Professional Documents
Culture Documents
American Scientist, Vol. 111.4 (July-August 2023)
American Scientist, Vol. 111.4 (July-August 2023)
AMERICAN
Scientist
American Scientist
How
Volume 111 Number 4
Real Are
Models?
Simulations can address our biggest
questions—if we are honest about
their power and their limitations.
July–August 2023
FREE
the visually stunning stone with a
concave cut set in .925 sterling Spirit Lake Helenite Ring
silver loaded with brilliant white, Total value with FREE studs $478*
Helenite Studs Ultimate Diamond Alternative®,
Now, Offer Code Price Only $99 + S&P Save $379!
-a $129 value- DiamondAura ® . The classic
with purchase of
Spirit Lake Ring
pairing of colors in a vintage-
inspired setting makes for a
1-800-333-2045
statement ring that’s simply Your Offer Code: SLR745-02
You must use this offer code to get our special price.
impossible to ignore!
Beauty from the beast. Also
known as “America’s Emerald,”
Stauer ® 14101 Southcross Drive W., Ste 155,
Dept. SLR745-02, Burnsville, Minnesota 55337
www.stauer.com
helenite is not an emerald * Special price only for customers using the offer code
at all, but a brighter and versus the price on Stauer.com without your offer code.
Rating of A+
4 carat Helenite center stone • Ultimate Diamond Alternative®, DiamondAura® accents • .925 sterling silver setting • Whole ring sizes 5–10
Stauer… Afford the Extraordinary®
Scientist
Volume 111 • Number 4 • July–August 2023
213
research, but their application
requires intent and skill. 242 On the Hunt for
KEVIN HENG Another Earth
Dendritic cells 216 Building Better GrowthDendritic cells DC1 DC2 are
DC1 DC2 DC3 pDC Astronomers DC3making
pDC progress
198 Monocytes
Curves
Current standards for assessing
Mo1 Mo2 Mo3
in finding
Monocytes
growth in infants and children often
planets
resemble our own.
ABEL MÉNDEZ
that broadly
Mo1 Mo2 Mo3
Neutrophils raise
N1 unwarranted
N2 N3 N4 concerns.
N5 BetterNeutrophils N1 N2 N3 N4 N5 N6
models could improve care.
Mø4
Macrophages WILLIAM
Mø1 Mø2 E.
MøBENNETT,
3 Mø4 Mø5 JR Macrophages Mø1 Mø2 Mø3 Mø4
DC
July–August 2023
Scientific models can help
www.americanscientist.org
Real Are
Building knowledge Models? computer simulations
Simulations can address our biggest
questions—if we are honest about
M
ADMINISTRATION
odeling is a technique that in “Bias Optimizers” (pages 204–207), EDITORIAL CORRESPONDENCE
cuts across scientific fields Damien Patrick Williams of the Univer- American Scientist
and uses many different sity of North Carolina at Charlotte focus- P.O. Box 13975
technologies. A math- es more on the creation of AI and the in- Research Triangle Park, NC 27709
919-549-4691 • editors@amscionline.org
ematical model can be used to study herent biases that it can amplify, making
climate, or galaxy formation (pages it difficult to root out even unintentional CIRCULATION AND MARKETING
NPS Media Group • Laura Garrity, account director
222–225), or cancer. An animal model sources of problematic results.
can reveal new insights about a disease There are many cases in which data ADVERTISING SALES
advertising@amsci.org • 800-243-6534
(pages 213–215). Models sometimes selection can affect model performance.
are employed to interpret processes In “Building Better Growth Curves” SUBSCRIPTION CUSTOMER SERVICE
American Scientist
involving subatomic particles, such as (pages 216–221), William E. Bennett, Jr P.O. Box 193
neutrino detection (pages 226–231) or of Indiana University School of Medi- Congers, NY 10920
nuclear fusion (pages 211–212). Mod- cine discusses the history of infant 800-282-0444 • custservice@amsci.org
els can be used to examine the past, growth projections and why they can
PUBLISHER
such as the process of human evolution often lead to false positives when di- SIGMA XI, THE SCIENTIFIC RESEARCH
(pages 235–236), or they can be used agnosing children with what’s called HONOR SOCIETY
to try to predict the future, such as in failure to thrive. But he also explains that President Marija Strojnik
global demographics (pages 200–203). efforts to personalize growth curves Treasurer David Baker
President-Elect Kathy Lu
Although models are ubiquitous in can raise serious concerns about data Immediate Past President Nicholas A. Peppas
science, they are not a magic box, and privacy and implicit bias. Executive Director & Publisher Jamie L. Vernon
they are only as good as we make them. Anne Goujon of the International
EDITORIAL ADVISORY PANEL
Properly conceptualizing them is an art Institute for Applied Systems Analysis Richard Boudreault, University of Waterloo
as well as a science, and their limits of in Austria discusses how historical de- René Fuanta, East Stroudsburg University
applicability have to be carefully con- mographic data can help countries plan Simson Garfinkel, AI2050, Schmidt Futures
sidered. In addition, scientists have to for the future, in “Predicting an Aging Sonya T. Smith, Howard University
Caroline VanSickle, A. T. Still University of
be vigilant in creating models that don’t and Changing World” (pages 200–203).
Health Sciences
inadvertently introduce bias (pages Goujon describes models of Niger’s
204–207), especially when artificial intel- future population that are based on American Scientist gratefully acknowledges support
for “Engineering” through the Leroy Record Fund.
ligence is employed in attempts to cope whether or not funding is increased
with the mountains of data that a model now for education, and explains how Sigma Xi, The Scientific Research Honor Society is
can use and produce. these different models can help policy- a society of scientists and engineers, founded in 1886 to
In this special single-topic issue, we makers decide where to use their lim- recognize scientific achievement. A diverse organization
of members and chapters, the Society fosters interaction
take a step back and give the big picture ited resources to achieve the most pros- among science, technology, and society; encourages
of what models can and can’t do well. perous outcomes for their countries. appreciation and support of original work in science and
technology; and promotes ethics and excellence in
In “Approximating Reality” (pages Readers can find lots more in this scientific and engineering research.
198–199), Kevin Heng of Ludwig Maxi- issue, and additional online content
milian University in Germany kicks off is listed on page 196. Join us on social Printed in USA
the discussion with his reflections on media to share your own experiences
what it means to understand a concept with modeling.
in the natural world, and how that defi- —Fenella Saunders (@FenellaSaunders)
59
exchange for imparting all of this energy and experience, you
expect nothing in return.
The point? You deserve to treat yourself once in a while. You
do so much for the people you care about. Now it’s time to
focus on you for just a few minutes. That’s where the Men’s
Due Volta Watch comes in. This astonishing innovation
provides a digital readout that’s powered by a precise quartz
engine, combining both analog and digital timekeeping.
Outfitted with a stopwatch, alarm, a.m./p.m.
settings, and day, date and month complications,
this timepiece stands out from the crowd. With
its large face and handsome, masculine design,
this watch is perfect for the back nine and the
happy hour afterwards. Water-resistant up to
3 ATM, this timepiece won’t wimp out if you
have a run-in with a water hazard.
Other hybrid watches can sell for thousands
of dollars, but because we’ve built more
than 1 million watches, we know how to
create real value. That’s why we can offer
this timepiece for just $59! If you’re not
completely satisfied with this watch, return it
for a full refund.
Act quickly! This one-of-a-kind watch has been
one of our fastest sellers this year. Of our original
run of 2,872, we only have 389 left for this ad! In
the last couple of years there’s been a watch shortage,
but Stauer has got you covered. If you’re the kind of
man who gives everything and expects nothing in return,
it’s time to treat yourself.
Watch Specifications:
• Complications: stopwatch, alarm, a.m./p.m. settings,
and day, date and month
• Precision movement chronograph
• Stainless steel caseback and crown
• Water-resistant to 3 ATM
Men’s Due Volta Watch $399 $59* + S&P Save $320
*Special price only for customers using the offer code.
Join more than 1 million smart
1-800-333-2045 people who own stauer watches
Your Insider Offer Code: DVW205-03
Stauer, 14101 Southcross Drive W., Ste 155, Dept. DVW205-03, Burnsville, MN 55337 www.stauer.com
NASA/JPL/Caltech
archives that delve into the the companion podcasts on
many ways that scientists our website.
have used models.
The Internet of Animals
How Reliable Are Upgrades in global
Models? positioning system satellites, insights into animal
Mathematical descriptions the use of the internet to migration and behavior,
and simulations help link datasets and process and also how data from
scientists forecast events and information in real time, and animals can feed back into
recommend actions, but it the miniaturization of devices weather and environmental
Movebank/422 South
can be difficult to determine has allowed for more wild monitoring. A video of
whether those predictions are animal monitoring than ever the talk and social media
trustworthy. For a primer on before. Roland Kays of North highlights are available on
the topic, check out “In Models Carolina State University our website.
We Trust—But First, Validate” discusses how this new
Am
by Daniel Solow of Case Internet of Animals provides Sci Look for this icon on articles with a
ssociated podcasts online.
American Scientist (ISSN 0003-0996) is published bimonthly by Sigma Xi, The Scientific Research Honor Society, P.O. Box 13975, Research Triangle Park, NC 27709 (919-549-0097). Newsstand single
copy $5.95. Back issues $7.95 per copy for 1st class mailing. U.S. subscriptions: one year print or digital $30, print and digital $36. Canadian subscriptions: one year print $38, digital $30; other foreign
subscriptions: one year print $46, digital $30. Print institutional rate: $75; Canadian $83; other foreign $91. Digital site license $200, print and digital institutional rate $275. Copyright © 2023 by Sigma
Xi, The Scientific Research Honor Society, Inc. All rights reserved. No part of this publication may be reproduced by any mechanical, photographic, or electronic process, nor may it be stored in a
retrieval system, transmitted, or otherwise copied, except for onetime noncommercial, personal use, without written permission of the publisher. Periodicals postage paid at Durham, NC, and ad-
ditional mailing offices. Postmaster: Send change of address form 3579 to American Scientist, P.O. Box 193, Congers, NY 10920. Canadian publications mail agreement no. 40040263.
Discovering EGYPT
Including Abydos & Dendara
November 4-16, 2023 For info
about Covid
and travel,
please call
Or look at our web site for many more exciting adventures:
www.betchartexpeditions.com our office.
the World with Sigma Xi! 17050 Montebello Rd, Cupertino, CA 95014-5435
Phone: (800) 252-4910 Email: SigmaXiInfo@betchartexpeditions.com
sigmaxi.org/renew
Science Policy
Boot Camp
July 25-28, 2023
Join us for a four-day deep dive into state level policymaking
with practitioners and leading experts in science policy,
communications, and engagement.
sigmaxi.org/meetings-events/science-policy-bootcamp
Approximating Reality
W
hat does it mean to understand to mimic as much of an observed phenomenon
the natural world? To a classi- as possible. But we achieve understanding by
cally trained physicist, it means using idealized models to capture the essence
that one is able to construct of the phenomenon. By design, idealized mod-
a model that not only accounts for currently els necessarily employ assumptions.
measured phenomena but is also capable of
predicting future phenomena. The modern Models and Their Limitations
challenge posed by artificial intelligence is Models are approximations of reality, con-
that if one has a sufficiently large dataset of a structed by scientists to address specific
system, in principle one may train a machine questions of natural phenomena using the
learning model to predict future phenomena language of mathematics. Mathematical equa-
without having any understanding of the un- tions allow the practitioner to decide the level
derlying physical mechanisms. It reignites the of abstraction needed to tackle a given prob-
debate of what “understanding” really means. lem, thus transcending the “billiard ball” ap-
In a mechanistic view of the world, any phe- proach. One example concerns the modeling of
nomenon may be understood as a large ensem- disease transmission. Human beings are com-
ble of interacting billiard balls. The reductionist plex entities whose individual behavior cannot
asserts that as long as enough basic units (“bil- be easily modeled by mathematical equations.
liard balls”) are present in such a system, one However, the movement of ensembles of hu-
may model or simulate it exhaustively. For ex- man beings within a city or country may be
ample, if one wishes to understand how waves approximated by a set of fluid equations: the
behave in the ocean, one only has to simulate so-called compartmental models of disease
the interactions of each and every constituent transmission. Our incomplete understand-
molecule of water. ing of the biological properties of a pathogen,
This reductionistic approach has been ex- as well as how it mutates and is transmitted
tremely successful at describing nature and across hosts, is encoded within a single pa-
making useful predictions, but it does not rameter known as the reproduction number.
satisfy a more comprehensive meaning of In this manner, epidemiologists were able to
“understanding.” Even if one could run such study COVID-19 during the pandemic with-
a simulation in one’s lifetime, it would be dif- out having full knowledge of it, because the
ficult to identify the underlying mechanisms focus was on its macroscopic behavior across
responsible for producing different types of large spaces. (See “COVID-19 Models Demand
waves. Generally, examining the microscopic an Abundance of Caution,” April 23, 2020.) In
components of a complex system in order to such an example, the theorist is employing
understand how macroscopic behavior emerg- the principle of separation of scale in order to
es falls short—think biology and economics. isolate phenomena.
Implicitly, this brute-force approach suggests Ideally, one would like to construct a uni-
that one may study complex systems without versal model that is able to answer every ques-
having a scientific question in mind, and if tion that the scientist asks of it. In practice,
enough computing power is deployed then mathematical solutions to equations describing
insight naturally emerges. The billion-euro Hu- nonlinear systems that can be written down
man Brain Project is a spectacular example of on paper are exceedingly rare. (Nonlinear here
the limitations of such an approach. Essentially, means that changes between the inputs and
their simulations have failed to replace labora- outputs of the system do not obey a simple
tory experiments. relationship.) Instead, one must solve these
As has been noted by the climate scientist equations with a computer, which allows one
Isaac Held, there is a tension between simula- to study how different components of a physi-
tion and understanding. We simulate in order cal system interact and produce nonlinear out-
be decisively confronted by data. Is- causal mechanism responsible for the honorary professor at the University of Warwick
sues such as falsification and Occam’s observed trend. It is almost as if we in the United Kingdom. He is author of the 2017
razor become relevant—one wishes are missing part of the language of textbook, Exoplanetary Atmospheres: Theoreti-
to construct models that are complex models and simulations, a tool of in- cal Concepts and Foundations (Princeton Uni-
enough to include the relevant phys- termediate complexity that goes be- versity Press). Email: Kevin.Heng@physik.lmu.de
1 2 3 4 5?
stage high stationary early expanding late expanding low stationary declining?
40 birth rate
?
death rate
(per 1,000 people per year)
30
birth and death rates
natural
20 increase
10
?
total population
natural
decrease
?
0
100 100
90 90
women in the country.
80 80
Assuming a continuation of present
70 70 policies, by 2062, only 35 percent of 25-
60 60 to 39-year-olds in Niger will have a sec-
50 50 ondary education or higher. But by in-
40 40 creasing investment in education today,
30 30 and boosting enrollment in vocational,
20 20 primary, and secondary education, the
10 10 proportion of citizens in this age group
0 0 with a secondary education or higher
would increase to above 70 percent by
20 2
20 7
32
20 7
20 2
20 7
20 2
20 7
62
20 2
20 7
20 2
20 7
20 2
20 7
20 2
20 7
62
12
20 7
12
20 7
2
2
3
4
4
5
5
2
2
3
3
4
4
5
5
1
1
20
20
20
20
20
year year
2062. In this alternative scenario, Ni-
ger’s population size by 2062 would
upper/post-secondary complete primary no education/incomplete primary triple to 78 million, rather than quadru-
vocational/lower secondary Koranic school age 0 to 6 pling to 106 million as expected under
current policies. (See figures at left.)
The author’s projections (bottom graphs) suggest that improving funding for schools in Niger (top)
could dramatically increase the proportion of educated citizens by 2062 (evolving scenario), where-
as maintaining the status quo would result in more modest educational gains (trend scenario). The Future of Ukraine
Demography moves at a slow pace—
century, placing these countries, on av- tion. Niger’s rapidly increasing popula- except when shocks happen, such
erage, 80 years apart from more devel- tion presents several issues, including as a war. Such effects occurred on a
oped regions in the demographic transi- economic pressure, food insecurity, and huge scale during World War II. My
tion process. The last country that until malnutrition. Most rural regions are colleagues and I see them happening
the early 2010s had not started its transi- particularly prone to drought, deserti- in a different context now in Ukraine.
tion to decreasing fertility was Niger in fication, and environmental degrada- Before the 2022 Russian invasion, the
West Africa, which until recently had tion, while also having limited access to population of Ukraine was already
had an average fertility rate of greater education and health care. shrinking due to relatively low fertil-
than seven children per woman for sev- In my role as a demographer at the ity, high mortality, and especially high
eral decades. But a survey published in International Institute for Applied emigration rates. Now, the war has put
2021 shows a rapid decline to 6.2 chil- Systems Analysis (IIASA) in Laxen- Ukraine on a path of rapid popula-
dren per woman. burg, Austria, my team and I worked tion decline. Greater than eight million
In 2017, I worked with the United with local experts and stakeholders have fled—a loss of almost 20 percent
Nations Children’s Fund (UNICEF) to develop narratives about the pos- of the total population.
country office in Niger and with the sible future of Niger, with an emphasis Predicting Ukraine’s future popula-
Nigerian government to explore future on the potential impact of educational tion size can help the region and the
demographic prospects and their de- development. We know from past de- global community prepare for the
velopment implications. Niger has the mographic trends in other parts of the country’s postwar needs. At IIASA,
fastest population growth in the world world that giving girls access to educa- my colleagues and I worked with other
and is also among the least-developed tion is an effective way to curb popula- demographers at the European Com-
countries. A large majority of the popu- tion growth. Education also greatly in- mission’s Joint Research Centre to de-
lation does not have any formal educa- creases a nation’s economic prospects. velop four hypothetical scenarios for
R
ecently, I learned that men them. This means we must be purpose- or more milk. But you can’t adjust the
can sometimes be nurses and ful about how we build AI systems so proportion of the pie’s ingredients with-
secretaries, but women can that they amplify the values we want out considering the rest, or you’ll end up
never be doctors or presi- them to, rather than the ones acciden- with a crumbly or spongy mess; it won’t
dents. I also learned that Black people tally fed into them. We have to ask really be a good pie. You must adjust the
are more likely to owe money than to questions about the source material whole recipe, the whole algorithm.
have it owed to them. And I learned that trains them, including books, so- To the person using it, an algorithm
that if you need disability assistance, cial media posts, news and academic may look like a unitary thing that per-
you’ll get more of it if you live in a fa- articles, and even police reports and pa- forms one job: A Google search, for in-
cility than if you receive care at home. tient information. We must also examine stance, seems like a singular, powerful
At least, that is what I would believe the frameworks into which that data is operation that searches the web. In real-
if I accepted the sexist, racist, and mis- placed: What is the system doing with ity, platforms and search engines work
leading ableist pronouncements from to- that data? Are some patterns or relation- on dozens of algorithms that search,
day’s new artificial intelligence systems. ships between certain words or phrases sort, rank, weight, associate, suggest,
It has been less than a year since OpenAI given more value than others? Which amplify, and suppress words, concepts,
released ChatGPT, and mere months ones? Why? What are the assumptions and content. Those algorithms work
since its GPT-4 update and Google’s re- and values at play in the design of tools in concert, but when you take a ma-
lease of a competing AI chatbot, Bard. that transform human lived experiences trix of algorithms and automate it, it
The creators of these systems promise into data, and that data into algorithms looks as if your computer system is
they will make our lives easier, remov- that impact human lives? autonomous and self-directed. So it is
ing drudge work such as writing emails, It is much easier to see through the with the new AI chatbots: They seem to
filling out forms, and even writing code. mystique of ChatGPT and other AI deliver on “true artificial intelligence,”
But the bias programmed into these sys- applications once you understand ex- a seductive idea that goes back to the
tems threatens to spread more prejudice actly what they are and what they do. dawn of the computer age, but they
into the world. AI-facilitated biases can The truth about such algorithms is that are actually composed of series of al-
affect who gets hired for what jobs, who they’re literally just sets of instructions. gorithms even more complex than the
gets believed as an expert in their field, You have a set of standardized opera- systems that came before.
and who is more likely to be targeted tions within which particular weights
and prosecuted by police. and measures can be adjusted. In so do- A History of Bias
For some people, the word bias is syn- ing, you have to adjust every element of Since the 1940s, when mathemati-
onymous with prejudice, a bigoted and the whole to make sure the final prod- cians and cryptographers such as Joan
closed-minded way of thinking that pre- uct still turns out the right way. Clarke, Jane Hughes, Pamela Rose, the
cludes new understanding. But bias also Algorithms are often sold as magical, other 8,000 women of Bletchley Park,
implies a set of fundamental values and but they are neither unexplainable nor and Alan Turing used early computer
expectations. For an AI system, bias may even terribly unfamiliar. The recipe for technology to break complex codes
be a set of rules that allows a system or any food—just as for anything you have and help win World War II, people
agent to achieve a biased goal. to make—is an algorithm, too. My fa- have wondered about the possibility
Like all technologies, AI reflects hu- vorite algorithm is pumpkin pie. If you of intelligence in digital computers. In
man bias and values, but it also has an go to make a pumpkin pie, you might the 1950s, computer researchers began
unusually great capacity to amplify decide you’d like less butter, more sugar, to ask, “Can machines think?” And in
QUICK TAKE
It is not surprising that systems trained on Many groups are already integrating AI Generative pretrained transformers (GPTs)
biased source material would result in biased tools into their public-facing interfaces, billing such as Bard and ChatGPT cannot recontextu-
outputs, regardless of whether or not the them as assistants and interventions to help alize or independently seek out new informa-
original biases were intentional. people do their jobs more efficiently. tion that contradicts their built-in assumptions.
Creators pour their emotions and subjective reactions into the algorithms that guide our lives, cessing. Researchers working on natural
making these complex systems as idiosyncratic, volatile, and biased as their creators are. language processing sought to combine
linguistics, computer science, artificial
the 1960s, a rift formed between two gram developed by computer scientist neural networking, and AI to find ways
camps of AI researchers at Dartmouth. Joseph Weizenbaum at the Massachu- for computers to interpret, process, and
One group focused on computation setts Institute of Technology in 1964. At communicate in human-like, conversa-
and cybernetics, which are feedback first, ELIZA was meant to parody open- tional language. In the 2010s, the models
loops that mimic biological processes. ended psychotherapy; the program Global Vectors for Word Representation
The other group worked to replicate would do things such as rephrasing the (GloVe) and Word2Vec were two of the
human neural networks in electronic typed inputs from users as parroted foremost examples of natural language
form. Neither camp considered ma- questions rather than replying with any processing systems. They work by sta-
chine bodies, emotions, or socializa- new thoughts. Even knowing that they tistically mapping the relationships be-
tion, however. These researchers firm- were talking to a computer, users re- tween words, embedding layers of as-
ly believed that the key to AI was to peatedly formed emotional bonds with sociative meaning between them.
divorce any messy social factors from ELIZA, often in as little as one or two Early LMs could represent the se-
the purity of rationality and intellect. short conversations. Weizenbaum was mantic connections between words
As part of this work, scientists devel- astounded at what he called the “pow- such as “dog” and “dig” or “plane” and
oped language models (LMs), a meth- erful delusional thinking” such a brief “flight.” These early programs used so-
od of determining the probability of engagement could produce. called machine learning, a process of
words connecting to each other based ELIZA was one of the first main- encoding various elements of English
on context cues such as their starting stream LMs, but the work didn’t end language as data, and then training
letter and the preceding word. One of there. The dream of AI grew up along- the system to hit particular predictive
the earliest examples was E LIZA, a pro- side the dream of natural language pro- targets and to reinforce associations
Modeling a
Greener Future
W
hen scientists try to un- ers that combine electricity and water
derstand how Earth’s to produce hydrogen for use as a fuel
climate is changing and or as a feedstock for other fuels. We
how we might halt the reexamined the cost trends for these
rise of global temperatures, they turn four green energy technologies and
to models. Some models predict how forecasted that they are likely to be-
global temperatures will rise depend- come much cheaper than most other
ing on future rates of greenhouse gas energy models predict. We then made
emissions, and how those temperature new projections for the expected cost
increases will likely affect humans and of transitioning from fossil fuels to
the environment. Other models fore- clean energy sources. Using our up-
cast the costs and benefits of various dated estimates, we calculated that
strategies to mitigate climate change. transforming the global economy into
But all of these models have limita- a carbon neutral one by 2050 not only
tions and sometimes need to be ad- is economically achievable, but will
justed based on new data. Now is one
of those times.
The need to adjust models is espe- Over the past two
cially acute when it comes to human
and technological factors, which can
decades, the large
be hard to quantify and about which energy models
we often have poor intuition. Over the
past two decades, the half-dozen large used to inform
energy models used to inform the in- influential reports
fluential reports by the Intergovern-
mental Panel on Climate Change have have systematically were commercialized in the 1980s
systematically overestimated future
costs of key green energy technologies.
overestimated future and 1990s. Solar and wind are now
the cheapest forms of new electricity
For example, in 2010 a model used costs of key green generation to build, and it is cheaper
by the International Energy Agency to run an electric vehicle than one with
projected that solar energy would energy technologies. an internal combustion engine. Why,
cost $260 per megawatt-hour in 2020. then, do the major models keep over-
The actual price that year was $50 per likely produce trillions of dollars of estimating the price of green energy?
megawatt-hour, well below the aver- net economic savings compared with A group of assumptions embedded
age price for electricity generated by continuing a fossil fuel–based system. within the models seems to be the
coal or gas. By overestimating green If governments pursue smart poli- source of the problem.
technology costs, these models have cies, going green quickly is likely to be Economists build models of the
made the transition away from fossil cheaper than either a slower approach global energy system based on ener-
fuel–based energy sources to green or sticking with the status quo. gy demand, resource availability, and
technologies appear substantially It should not come as a surprise past technological trends to predict
more expensive than it is likely to be. that green energy technologies are how much different energy sources
Last year, my colleagues Doyne on track to become cheaper in the fu- will cost in the coming decades. These
Farmer, Matthew Ives, and Penny ture. Innovation and ingenuity have large energy models typically include
Mealy, and I published a model that been driving down the costs of these a few types of assumptions that con-
took a different approach to predict- technologies consistently for decades. strain the values of variables at play.
ing the costs of four key green energy The prices for solar energy, wind en- Floor cost constraints set hard limits be-
technologies: solar energy, wind en- ergy, and batteries have all dropped low which costs for different technolo-
ergy, battery storage, and electrolyz- by more than 90 percent since they gies are not allowed to fall, to avoid
Current energy models overestimate the cost of green energy technologies such as solar pow- Our model removes or greatly re-
er. A new model predicts that these costs will continue to drop, making the transition away laxes the assumptions built into the
from fossil fuel–based energy sources much cheaper than expected. leading energy models, based on ideas
that one of my colleagues, Doyne
Farmer, a complexity scientist at the
unrealistically low-cost projections. an influential model from one of the Institute for New Economic Thinking
Deployment rate constraints set upper top international modeling teams at the University of Oxford, developed
limits on how fast technologies can limited solar and wind to providing in the late 2010s. Farmer started his
be manufactured and rolled out, to a maximum of 20 percent of the pow- work after talking to energy modelers
avoid predicting unrealistically rapid er in the electricity grid. But in 2023, to understand how they came up with
deployment. And technology mix con- those sources routinely provide more technology cost forecasts that looked
straints limit the amount of solar and than 60 percent of the power in large decades into the future. He was dissat-
wind power allowed in the electricity economies such as the United King- isfied with what appeared to him to be
grid in the model, to reflect concerns dom and Germany. Similarly, before highly unscientific methods, including
about production gaps created when 2015 most major models included floor their conservative assumptions about
those green energy sources are not constraints for the costs to produce the future development of green tech-
generating electricity, such as during solar energy, the lowest of which was nologies. Farmer thought that the fore-
the night and on days without wind. $750 per kilowatt. Today, the lowest- casts seemed to rest on ideas that were
In the standard energy models, all cost systems produce electricity for less not grounded in history or based on
these constraints appear to have been than $600 per kilowatt. Setting pessi- observations about how technologies
set far too conservatively, systemati- mistic technology constraints has led change. Also, he noted that many of
cally underestimating technological to higher cost projections for renewable the ideas incorporated into the models
improvements. For example, in 2015 energy than what we actually observe. had not been tested.
From your perspective, what makes fu- all of this material down to extremely that implodes the capsule, and how
sion such a difficult problem? Why is it small volumes, extremely high pres- that happens symmetrically. And then
so hard to model, and then why is it so sures, and reaching very high temper- modeling the plasma physics condi-
hard to control the experiment? atures. But as the material implodes, tions in the center of the dense plasma
The thing that makes it so difficult it’s not just the end state that’s hard that we create, and the diagnostic sig-
to control and model is that we need to model. It’s the entire implosion natures that come out of that as well.
extreme conditions. We need extreme that’s hard to model as well. There are
temperatures. In ICF [inertial confine- many states along the equation of the What difficulties do you encounter
ment fusion] we also need extreme den- density-temperature-pressure relation- when you’re trying to deal with all of
sities. We’re reaching pressures that ship that we don’t understand about these different parts of the modeling?
are more than two times the center the materials we’re trying to implode. There’s quite a bit of simultaneous op-
of the Sun, and temperatures that are There’s no experimental data. timization that has to occur. When NIF
more than five times the center of the first started [in 2009], it went for the
Sun, in our experiments. We’re making You are dealing with temperatures, highest yielding potential design that
the most extreme plasma state that you pressures, and conditions that go be- we had. The highest potential gain de-
can make on Earth. As you can imag- yond known physics. How do you sign. And that design had issues. It
ine, since that’s not been done before, model something when you’re going was more susceptible to instabilities.
there’s quite a bit we don’t know about into such unfamiliar regions? It just didn’t work. There are a lot of
the materials science. We have some low-pressure data and situations where your model says one
In these experiments, laser beams single points of parameter space that thing, but when you actually go to do
enter and hit the inside of a hohlraum [a we use to benchmark our equation of the experiments you find that you’re
hollow cylinder made of heavy atoms state models. Then we have our the- missing physics, or you can’t calculate
such as gold or depleted uranium] and ory, our transport models, radiation that physics because you don’t have
create a very intense radiation bath. transport models, and hydrodynamics the resolution, or there’s something
We have to be able to model that con- models. We benchmark these models you don’t understand.
dition, and then we also have to model as we go to integrated experimental Over the last several years we’ve
the plasma conditions of the implosion data. We don’t have a lot of data to been working to rebalance, optimiz-
as it’s imploding. Inside this intense support the basic physics models, or ing what’s good for the implosion
radiation bath sits a spherical capsule, to test them, but we have generated a and what’s good for the hohlraum to-
and in our experiments it’s made out lot of integrated measurements, inte- gether. They’re usually not the same
of diamond. And inside of that spheri- grated datasets. You do an experiment thing. We want to increase the size of
cal capsule sits deuterium and tritium where all this complicated physics is the implosion, but when you do that,
fuel [two hydrogen isotopes]. happening, and then you do an after- it becomes a more massive target. If
When we make this intense radia- the-shot simulation to try to match all you don’t have more laser energy to
tion bath, it heats the outside of the the observables from that experiment. blow off the material and implode that
capsule that holds the fuel. That heat It’s like an integrated check on how extra mass, then the implosion goes
explodes the outside of the capsule, the modeling is doing. slower, and if you can’t get it going
which, in a rocket-like effect, sends the Integrated modeling is basically try- fast enough, you can’t squeeze the ma-
remaining capsule and the fuel inward ing to model the entire system, from the terial fast enough.
and squeezes it. We’re taking some- laser hitting the hohlraum to ignition We don’t have any more laser ener-
thing the size of a BB and squeezing it and what comes out of it. It’s modeling gy to drive the implosion, so we have
down to a size roughly half the diam- how the lasers interact with the gold to make the hohlraum that surrounds
eter of a human hair. We’re squeezing and depleted-uranium cylinder, how it more efficient. For a given amount
I
n a nondescript laboratory, a uni- addiction, heart disease, and much while deviating from humans in oth-
versity researcher using mouse more (see box on page 215), some still ers; the central question is the degree of
models to find new drugs to treat argue that too many preclinical targets congruence in the biological pathways
major depressive disorder finds have failed to be effective for human of focus and whether the differences ob-
an interesting potential drug target. use and that vast amounts of resources served are associated with the research
She is so sure this drug target will have therefore been wasted. Much of hypothesis. Any attempt to universally
work that she decides to go indepen- this disagreement stems from uncon- nullify or endorse animal model usage
dent and seek funding to start her own scious bias on the part of researchers. In will benefit neither scientific discovery
company. She is given start-up funds the aforementioned papers, the authors nor, ultimately, human health.
by a Big Pharma company and she who concluded that animal models
speeds up development of the drug as were ineffective were mostly physi- The World Needs Congruence
quickly as she can to make it ready for cians, whereas those who concluded In many biomedical fields, it’s common
clinical trials. The drug passes all its that the animal models were efficacious to see later studies contradict original
trials with flying colors except for one: findings and generate refutations. This
the final clinical trial in humans.
In this all-too-common scenario, tax-
In a variation of phenomenon is known as the Proteus
phenomenon, after the Greek god of
payers and sponsors might well argue George Box’s change. However, any data s cientist—
that the years of research and millions
of dollars in funding were wasted. In- assertion that “All statistician, computer scientist, or
bioinformatician—cannot accept con-
deed, according to the National Center models are wrong, tradictory conclusions drawn from a
for Biotechnology Innovation, it takes single dataset; it threatens the rigor and
15 years and an average of $900 million but some are credibility of the data science field.
to bring viable new drugs to market.
Mouse models have long been used
useful,” we’d argue Scientific research has more tools
than ever at its disposal, but the del-
in biomedical research, but in recent that “All biological uge of information can be overwhelm-
years, papers published in the Proceed- ing and difficult to properly interpret.
ings of the National Academy of Scienc- models are What we need to do first is to ask the
es of the U.S.A. (PNAS) claiming that
mouse models poorly mimic human
imperfect, but some right questions and then use the best
tools available to interpret the data. To
responses have fanned the flames of are applicable.” bridge the gaps between clinicians and
controversy in the biomedical commu- basic researchers, we’ve developed a
nity. It all began in 2013, when a paper were basic research scientists. These software program called Congruence
published in PNAS that ultimately was implicit biases have driven an unneces- Analysis of Model Organisms (CAMO).
covered in the New York Times main- sary wedge between the clinical and CAMO uses sophisticated statisti-
tained that mouse gene expression basic research communities. cal tools, aided by machine learning,
responses to inflammatory diseases In a variation of statistician George to compare organisms at the molecu-
such as those brought on by burns, Box’s assertion that “All models are lar level and pinpoint disease-driving
trauma, and endotoxemia (the buildup wrong, but some are useful,” we’d ar- genes. This mechanism allows re-
of dead bacteria, often from the gut, gue that “All biological models are im- searchers to take a more granular ap-
in the bloodstream) poorly mimicked perfect, but some are applicable.” Model
human responses and concluded that organisms often are highly similar ge-
research using mouse models was netically to human systems; mice and
therefore a waste of time and money. A humans share more than 90 percent of
year later, another paper reinterpreted the same genes. Animal models can
the dataset and presented the oppo- mimic human systems in some ways
site conclusion. It was both intriguing
and difficult to believe that the two
papers drew contradictory conclusions
after analyzing the same datasets from
mouse and human models.
Though animal models have clearly
helped in developing drugs that treat Panther Media GmbH/Alamy Stock Photo
DNA molecule
Dendritic cells DC1 DC2 DC3 pDC Dendritic cells DC1 DC2 DC3 pDC
Human cell types
Neutrophils N1 N2 N3 N4 N5 Neutrophils N1 N2 N3 N4 N5 N6
Mø4
Macrophages Mø1 Mø2 Mø3 Mø4 Mø5 Macrophages Mø1 Mø2 Mø3 Mø4
DC
Mice and humans share more than 90 percent of their DNA and have many cell types in com- Bibliography
mon. Researchers can use software tools such as Congruence Analysis of Model Organisms Ahn, A. C., M. Tewari, C. S. Poon, and R. S.
(CAMO) to find congruence between the two species, which can ultimately lead to more ef- Phillips. 2006. The limits of reductionism in
ficient discovery of potential drug targets. medicine: Could systems biology offer an
alternative? PLoS Medicine 3(6):e208.
Bryda, E. C. 2013. The mighty mouse: The
proach, using congruence between ani- her research. Perhaps it would have impact of rodents on advances in biomedi-
mal models and human systems to find helped her find better drug targets cal research. Missouri Medicine 110:207–211.
useful treatments for disease. For exam- faster or helped her to realize that Ericsson, A. C., M. J. Crim, and C. L. Franklin.
ple, when we reanalyzed the human- mouse models do not match humans 2013. A brief history of animal modeling.
mouse inflammatory disease data from closely enough in this instance to pro- Missouri Medicine 110:201–205.
the original papers that caused such vide a good drug target. CAMO could Seok, J., et al. 2013. Genomic responses in mouse
models poorly mimic human inflammatory
controversy, we found mice were like likely have saved her many hours and diseases. Proceedings of the National Academy of
humans in immune- and inflammation- millions of dollars and helped her Sciences of the U.S.A. 110:3507–3512.
related pathways and gene markers bring more successful targets to trial. Takao, K., and T. Miyakawa. 2015. Genomic re-
but different in the ways they translate Modern biomedical research offers a sponses in mouse models greatly mimic hu-
proteins. This revelation could help re- multitude of useful tools, but research- man inflammatory diseases. Proceedings of
the National Academy of Sciences of the U.S.A.
searchers focus their attention where ers also must learn to adjust their ex-
112:1167–1172.
it’s most needed and could also help pectations and ask the right questions
Zong, W., et al. 2023. Transcriptomic congru-
rule out ineffective potential drugs ear- with those tools. CAMO is one tool that ence analysis for evaluating model organ-
lier in the discovery process. may help by showing where mouse isms. Proceedings of the National Academy of
CAMO also provides enough statis- models and human systems are alike Sciences of the U.S.A. 120:e2202584120.
tical sensitivity to assess whether the enough to prove medically useful and
data will even provide enough infor- where they are different enough that
mation to draw congruence conclu- further research isn’t indicated. Effec- George Tseng is a professor and vice chair for research
sions in the first place. When sample tive integration of different types of ex- in the department of biostatistics at the University
size is small or biological variability perimental data at all levels helps us of Pittsburgh School of Public Health. Tianzhou
with a species is large, CAMO warns grasp the bigger picture when we eval- (Charles) Ma is an assistant professor of biostatistics
in the department of epidemiology and biostatistics at
researchers of insufficient information. uate the usefulness of animal models.
the University of Maryland School of Public Health.
Ultimately, it’s our hope that such tools Jian Zou received his PhD from the department of bio-
The Future Is Congruent will help usher in a new era of biomedi- statistics at the University of Pittsburgh and will join
Returning to the researcher with start- cal research, allowing us to bridge the the department of statistics and management in the
up funding, we wish we could have gap between asking the right questions School of Public Health at Chongqing Medical Uni-
provided her with CAMO early on in and finding the best answers. versity in China. Email for Tseng: ctseng@pitt.edu
H
umans have a long history
of using animal models to
learn about ourselves. The
first records of animal model us-
age date back to ancient Greece. In
the fifth century BCE, Alcmaeon
of Croton observed connections
between the brain and sensory or-
gans in dogs. A couple of centu-
ries later, Aristotle studied embryo
growth in chicks.
Since then, animal models have
played an indispensable role in
biomedical research, leading to
many of the biggest medical break-
throughs. William Harvey, founder
of modern physiology, discovered
blood circulation in the early 1600s
after studying the anatomy of sev-
eral species of animals. Surgeon
Frederick Banting and his student
Charles Best found that injections of
pancreatic cell extracts relieved dia- Science Source
betic symptoms in dogs, leading to Galen was the physician to Emperor Marcus Aurelius of Rome in the second century
the discovery of insulin in the 1920s. CE. He studied anatomy and physiology by dissecting pigs, focusing particularly on
The Salk and Sabin polio vaccine the functions of the kidneys and spinal cord. Known as the father of modern medicine,
he made many contributions in the fields of anatomy, physiology, pathology, pharma-
was developed based on more than
cology, and neurology.
40 years of experiments using mon-
key, rat, and mouse models, leading
to a successful double-blind trial on ied,” a concept known as Krogh’s only 15 percent of drug candidates
1.8 million children in 1954. principle. That principle no longer survive to Phase III trials in humans
Despite these obvious benefits, seems so certain, as many biomedi- after successful preclinical animal
the use of animals in biomedical cal investigations have shifted from studies, and then only half of those
research has been a subject of de- are eventually approved for clini-
bate and controversy for decades, In cancer, only cal use—a success rate of less than
as people have sought to balance 8 percent. For psychiatric disorders
medical advancements with animal 15 percent of such as depression, bipolar disor-
rights. Today, most scientists accept
that animals should be used for
drug candidates der, and schizophrenia, many drugs
tested in animal models failed to
research ethically and responsibly. survive to Phase produce results adequate for clinical
Research regulations and academ-
ic guidelines generally follow the III trials in humans use. Despite the advances of mod-
ern molecular biology, many of these
principles of the Three Rs: Replace- after successful diseases continue to stump clinicians
ment of animals where possible, due to the involvement of advanced
Refinement, and Reduction. preclinical neural networks and the potential
Many researchers have also be-
gun questioning long-held assump-
animal studies. for drug targets to produce toxic-
ity. The pharmaceutical industry has
tions about the validity and cost- anatomy and physiology to com- therefore reduced its research and
effectiveness of using animal mod- plex models of understanding, man- development of psychiatric drug
els. In 1929, Nobel Laureate August aging, and treating disease. targets over the past 50 years. Un-
Krogh wrote in the American Journal Modern styles of medical research derstanding the congruence of ani-
of Physiology that “for such a large require a much higher level of con- mal models to human systems may
number of problems there will be gruence between animal models increase the success rate in drug dis-
some animal of choice . . . on which and the specific human diseases covery and recover the investments
it can be most conveniently stud- they are used to study. In cancer, in drug development.
William E. Bennett, Jr
I
first met Michael years ago, when However, Michael’s parents did not spoken with hundreds of families dur-
he was just over a year old, the agree with the assessment of his doc- ing my career who have been told by
first and only child of a young tors. They saw a small, healthy boy who their pediatrician or other provider that
Amish couple. He was admitted didn’t eat much but didn’t seem to be their child has failure to thrive—only
to our children’s hospital to figure out thin or starving either. They saw a hap- to end up with a shrug of the shoul-
why he was growing so slowly. I was py, vibrant child who was full of energy ders from the same doctor after all the
the supervising gastroenterologist that and was thriving. The family had many tests come back normal. The variation
week, but he had already seen several questions and were hesitant to proceed in growth trajectory, body habitus, and
of my colleagues, including other spe- with any invasive tests or procedures. height seen in the healthy general pop-
cialists. His diagnosis was something Michael’s parents were correct, it turns ulation is enough on its own to throw
doctors call failure to thrive, a term out. After reviewing his records, I agreed growth curves for a loop.
plucked out of Victorian-era medicine with them: He was healthy and thriving. Even though context is so important
that has no clear consensus definition So where did this dissonance be- in assessing growth, we fail to teach
and is simply what we say when a tween family and medical team arise? trainees how these curves are made
child isn’t growing as we expect. The answer lies in a ubiquitous tool of and why this process can be mislead-
In Michael’s case, this diagnosis was pediatrics, the growth curve. It’s at the ing. We also fail to understand the
complicated by a rare genetic disorder front of every chart in every pediatric test characteristics—such as the false
caused by the deletion of a large portion clinic in the world. During every visit positive rate—in a tool we use every
of one of his chromosomes. Only a half to a pediatrician or family doctor, start- day. We fail to contextualize genetics
dozen cases of this specific deletion have ing soon after birth, a child’s height and medical history sufficiently when
been reported, so little was known about and weight are plotted on a standard making a binary decision, asking, “Is
the expected course of his life, other than curve. Medical students and residents this a health problem?” And perhaps
that he would have substantial delays in see hundreds of growth charts dur- most heartbreaking, we often fail to
development and might never learn to ing their training, and reviewing them look at the patient and talk to the fam-
speak or walk. Most clinicians will never is an essential part of both pediatric ily when arriving at a diagnosis. Un-
see this specific genetic problem once checkups and the assessment of many necessary testing, expense, time, pa-
during their entire careers. medical problems in children. rental anxiety, and even risks for the
Despite this complexity—and per- But like most health information, patient are the usual results.
haps because of it—Michael’s medical context is supremely important when Although I have had a clinical and
team felt something must be done. A assessing a growth curve, and Mi- research interest in these problems
battery of tests was planned, including chael’s case illustrates the many weak- for two decades, they never really hit
endoscopy and surgical placement of nesses of this ubiquitous medical tool. home until I experienced them first-
a gastrostomy tube, which would feed Even for kids without genetic disor- hand as a parent. Like Michael, my
him more calories. ders, the problem is pervasive. I’ve daughter has a rare chromosome
QUICK TAKE
Growth curves are a standard screening tool Failing to understand the patient’s context Improving growth curves is no easy task, re-
in pediatric clinics around the world. How- when they are not growing as expected can quiring the use of longitudinal data and more
ever, those using these curves end up flagging lead to unnecessary testing, expense, time, diverse datasets. Personalizing growth curves
healthy kids as having potential problems. parental anxiety, and even risks to the patient. is theoretically possible with machine learning.
From birth onward, children are weighed at every health checkup, and their weight is plotted overuse of tests and procedures. We can
along with their height on standard growth curves published by the U.S. Centers for Disease use sophisticated tools in data science to
Control and Prevention and the World Health Organization. But without understanding the enhance this mid-20th-century instru-
limitations of these growth curves, they can result in a high false positive rate for diagnosing ment. We can make the curves more
failure to thrive, a term indicating the child is not gaining weight or growing appropriately. inclusive and representative of human
diversity, more accurate at predicting
deletion— h ers is called Phelan– her body to grow at a different pace illness, and more personalized. We can
McDermid syndrome. Like Michael, than the growth curve suggests. build better growth curves.
most children with her condition have This problem is fixable. We can teach
profound delays in development, com- clinicians to use growth curves appro- Michael and Children like Him
munication, and intellectual ability. priately, and thus limit spurious diagno- Michael continues to thrive, years af-
And like Michael, her genes are telling ses of failure to thrive and its attendant ter I met him. The day I first spoke
140 25
growth curves metabolic machinery, so much higher
130 10 portions of our genes are involved,
120 healthy child 5
perhaps as much as 10 to 20 percent.
growth failure
110 One can see why a deletion of a sizable
Adapted from CDC; individual growth data from the author
repositioning
100 chunk of a chromosome will almost
90 always include something critical to
80
growth in children.
Based on these insights alone, we
70
can deduce that some individuals can
60 have variations that might not fit the
50 typical growth curve.
40
30 How We Use Growth Curves
The first part of the problem with
growth curves is how health provid-
ers use them, even in the best of cir-
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
cumstances. They are, in most cases,
age (years)
a screening test. That is, every child is
Three real-life growth curves are shown plotted with the standard curves in gray, ranging from monitored using growth curves, which
the 5th to the 95th percentile. The orange curve shows the growth of a healthy child whose rate are designed to detect growth prob-
of weight gain is far lower than normal. The green curve shows a child who is indeed under- lems with high sensitivity. But, just like
nourished and needs monitoring and health care guidance. The pink curve shows a healthy many other screening tests, doctors
child whose growth follows one percentile curve for a while and then drops to another. frequently encounter false positives.
False positives are a pernicious di-
with his family, I reviewed the growth sheer number of genes involved with lemma in medicine. We want to make
curve and what I had observed dur- brain development. As our primary sure nothing slips through detection,
ing my physical exam, and we elect- evolutionary asset, our big brains re- because we are taught that it’s much
ed to defer further workup. For the quire an enormous number of com- more serious to overlook a cancer or
next year, we just watched his weight plex interactions, so neurodevelop- an infection (false negative) than it is
closely. He has never caught up to the ment and neurological function utilize to be wrong in retrospect (false posi-
curve and likely never will. But he has 80 percent to 95 percent of our genes. tive). So where do we draw the line?
followed his own curve well below Missing genes or abnormal amounts Conventional medical wisdom tends
to separate tests into those for diagnosis
and those for screening. For instance,
the hemoglobin A1c blood test for dia-
Conventional medical wisdom tends to betes is more than 99 percent specific
if it finds elevated levels of glucose,
accept many false positives in our quest but only about 30 percent sensitive. It
therefore works great as a diagnostic
to keep people alive. test. But that’s not the same as a screen-
ing test. Diagnostic tests assume you’re
targeting a population with symp-
toms. Screening assumes you’re testing
the first percentile, and he has stuck to of activation of these genes can affect everyone, in a population with very
it. Perhaps more accurately stated, his development at any stage. low prevalence of the problem in ques-
genes have stuck to it. Growth has a similar, if not quite tion. Growth curves are screening tests.
A large change in genetic material, so dramatic, pattern. In a single-celled Most kids grow just fine, so they are
such as a chromosome deletion, invari- eukaryote such as the fungus Saccharo- more likely to have a measurement er-
ably affects multiple body systems— myces, 2,000 of the roughly 5,000 total ror or normal deviation from the curve
and almost always the brain. One of genes are involved with growth rate, than they are to have a serious underly-
the reasons that most children with and about half of those are critical ing disorder causing growth failure.
major chromosome abnormalities have to growth. Of course, with complex Pediatricians use growth curves
developmental delays is because of the multicellular organisms like humans, just like they do any other screening
Stephanie Freese
even among that population, most pa-
tients I see referred for failure to thrive
are completely healthy. In the mean-
time, though, many of these perfectly
healthy children have undergone a
variety of potentially unnecessary or Environmental context and genetics both influence a child’s growth rate in complex ways.
harmful interventions—feeding tubes, The acquisition of body mass involves much of the human genome’s metabolic machinery,
excessive calories and force-feeding, and the ways those genes interact with one another, the environment, and epigenetics are not
radiographs that involve unnecessary well understood. Smaller parents tend to have smaller children, but diet, energy expenditure,
radiation, potentially harmful surgical pollution, stress, and many other environmental variables can play a role. Children with chro-
procedures, and disruptions in breast- mosome abnormalities also often have abnormal growth.
feeding, among many others. Further-
more, the effects on family dynamics, centile or the fifth percentile, or cross- visualized in a small enough window, it
parental anxiety, and parents’ percep- ing two or more percentile lines, or no can seem as if normal growth suddenly
tion of their child’s overall health can growth for six months, or no growth arrests, but if you zoom out, the line
be drastic and hard to reverse. for two months, or a growth rate below smooths. This pattern can be even more
20 grams per day, or any weight loss pronounced once a child is diagnosed
Why Curves Fail during childhood. You get the picture. with failure to thrive, because addition-
Growth curves do not effectively iden- Another common problem is called al monitoring takes more data points
tify problems, and they are not good repositioning. Many children will grow closer together in time. These stairsteps
at identifying who is “thriving” and for months or years along one line on can be misinterpreted as the earliest
doing well, either. These problems the growth curve, then inexplicably signs of a curve flattening, which could
stem from the lack of a clear defini- move to another percentile line and create more concern and more testing,
tion for the diagnosis for which growth grow along it instead. This shift hap- even when unnecessary.
curves screen: failure to thrive. The first pens commonly in the first two years These systematic deficiencies in
mention of the term in the medical of life and is widely regarded as a nor- growth curves are probably more
literature in 1906 is by Meinhard von mal phenomenon. Despite this preva- pervasive than is usually taught. My
Pfaundler, an Austrian pediatrician lence, a significant portion of referrals research group at Indiana University
and director of the children’s hospital I receive for failure to thrive turn out School of Medicine collected data from
in Munich. He used the term in an ar- clearly to be repositioning when I as- a large cohort of 9,369 children in an
ticle about what was then called “ma- sess the shape of the growth curve. Indiana primary care network and
ternal deprivation syndrome”—when This kind of shift doesn’t even take tracked the expected weight percentile
neglected, abandoned, or orphaned into consideration children such as Mi- at 12 months compared with where it
infants showed growth rates and de- chael or my daughter, who may be on was at 1 month. In the paper we pub-
velopment far below other children. their own line far below the curve and lished in JAMA Pediatrics, we found
The term evolved over time, and in are perfectly healthy, nonetheless. that nearly half of children fell by one
the 1930s it began to be used to indi- In a similar vein, many children or two lines on the growth curve be-
cate undernutrition or lack of adequate grow in a discontinuous fashion. Rather tween the ages of 1 and 12 months.
growth from any cause. Today, the than a smooth line, if enough measure- Far from “following the curve” as we
term failure to thrive is a ubiquitous ments are taken, one often sees a stair- might expect, most kids drifted down-
fixture of medical education. Various step pattern indicating periods of calor- ward. Something significant must be
sources have since defined failure to ic intake without weight gain followed happening between those time points,
thrive as falling below the third per- by growth spurts. If these stairsteps are right? Usually not.
Cosmos Ex Machina
T
he universe exists on scales Just a few decades ago, it was not at within 10 percent. Most recently, the
of time and distance that lie all clear that we would be able to reach powerful James Webb Space Telescope
entirely outside the range of this point. As recently as the 1990s, cos- (JWST)—which launched in December
human experience. It is domi- mologists had not yet discovered dark 2021—has shown us that the era of
nated by two substances—dark matter energy, an omnipresent and elusive galaxies stretches back at least to the
and dark energy—that we cannot yet form of energy that is now recognized time when the universe was a mere
create, capture, or even measure in a as the driver of the accelerating expan- 320 million years old and was about
lab. In the face of our ignorance, rel- sion of the universe. Researchers dis- 1/14th its current size.
egated as we are to a fleeting moment agreed about the expansion rate of the JWST is also keeping us humble. Its
on one small planet, it may seem an universe by a factor of two, a discrep- early observations indicate that galax-
absurd ambition for us to make sense ancy that provoked bitter arguments at ies formed earlier and faster than we
of it all. Yet that is exactly what we scientific conferences. And astrophysi- had expected, in ways that our models
cosmologists attempt to do. cists could only speculate about when did not predict. These findings are forc-
We work to combine observations, the first galaxies began to form. ing us to reexamine our ideas about the
mathematical models, and computer Then in the 2010s, the European earliest generation of stars and galaxies.
simulations to retrace the path from Space Agency’s Planck satellite pinned But out of this current confusion could
the chaos of the Big Bang to the mod- down the recipe for the modern uni- emerge the next revolution in our un-
ern universe. Astonishing as it may verse: 68 percent dark energy, 27 per- derstanding of the universe.
seem, we are succeeding. Propelled by cent dark matter, and just 5 percent
a succession of ever-more-powerful atomic matter. Planck’s high-precision Universe in a Box
telescopes, along with modern super- measurements of cosmic radiation also In some ways, nature has made it re-
computers that can perform millions indicated that the universe is 13.8 bil- markably easy to simulate the uni-
of calculations in a trillionth of a sec- lion years old, with an error of just verse. If we can know the energy con-
ond, we can now provide a detailed 23 million years. Analyses of distant tent of the universe at any one time
account of the growth and develop- stars and supernovas have determined (with matter defined in terms of its
ment of galaxies over cosmic time. the expansion rate of the universe to equivalent energy via E = mc2), we can
simply plug those numbers into the dark matter; their atomic matter is too simulations to model the universe. The
equations of general relativity and un- sparsely distributed to form stars. The ensemble technique attempts to simulate
derstand how the energy density of inexorable pull of gravity subsequently many galaxies from a representative
the universe has evolved at all oth- brings together many smaller dark ha- section of the universe, encompassing a
er times. Nature has also provided a los to form larger ones. It takes time for volume that is hundreds of millions to
blueprint of what the early cosmos them to grow massive enough to trig- billions of light-years across. The zoom-
looked like. We can observe the cos- ger the formation of galaxies. in technique places a computational
mic microwave background—relic radia- Ordinary matter may make up only magnifying glass on individual galactic
tion that has traveled to us unimpeded 5 percent of the universe, but it produc- systems and explores these in great de-
from a time 370,000 years after the Big es most of its complexity. Protons, neu- tail. Each approach has its benefits and
Bang—to study the highly homoge- trons, and electrons interact in varied its limitations. The ensemble technique
neous (but not perfectly so) primordial and complicated ways. The laws that allows us to make predictions about
distribution of matter and energy that govern atomic matter and its interac- collections of galaxies but can resolve
seeded the galaxies we see today. tions with radiation (and, through grav- individual systems only fairly coarse-
With this knowledge, we can predict ity, with dark matter) are well known, ly; the zoom-in technique can provide
how dark and atomic matter assembled but the outcomes are hard to predict. spectacular detail but only on a galaxy-
into collections called halos, the forma- Gas cools, condenses, and forms stars; by-galaxy basis.
tion sites of galaxies. Using the biggest the stars, in turn, inject energy and mo- Although my background lies in
supercomputers and cranking through mentum into ambient gas during their conducting big-picture, ensemble stud-
the equations for gravity, my colleagues lives and their often-spectacular deaths. ies, I have been captured by the allure
and I can evolve simulated mini- It is essentially impossible to analyze of zoom-in simulations, because they
universes and see how they change these competing processes by hand. allow us to limit our assumptions and
over time. Our models, combined with Instead, we feed the equations of gas pose more specific questions. Can we
observations of the real universe, tell dynamics into our supercomputer sim- form individual galaxies that have
us that structure grows hierarchically. ulations and explore the consequences. thin disks and spiral structures like the
Small halos form first, dominated by We use two complementary classes of Milky Way? Can we also form huge,
Today
DARK
undetectable observed observed
AGES
era of light and first cosmic earliest modern
cosmic matter are atomic microwave known local
inflation coupled nuclei background galaxies universe
This cosmic time line highlights how much of the universe we can observe. It is fundamentally from the European Space Agency’s
impossible to observe events from before the time of the cosmic microwave background. After Euclid satellite, to NASA’s upcoming
that, there was a long gap—the dark ages—before galaxies formed and became visible to us. New Roman Space Telescope, to the Dark
telescopes are exploring ever-deeper into the dark ages in search of the earliest stars and galaxies. Energy Spectroscopic Instrument in
Arizona, to the Vera C. Rubin Observa-
Those galaxies are so distant that their most speculatively, maybe these obser- tory taking shape in Chile—are set to
radiation has taken billions of years to vations point to a fundamental short- explore the nature of dark matter, dark
reach us; over that long journey, the ex- coming in our cosmological model. energy, and galaxy formation.
pansion of the universe has stretched One hypothesis is that there could There will almost certainly be ad-
energetic light from young stars in be a distinct form of dark energy that ditional surprises, and we will have
those galaxies into longer-wavelength operated very early in cosmic history to update our models for how galax-
infrared rays that cannot be detected (just 50,000 years after the Big Bang), ies form and evolve, perhaps in major
by the Hubble telescope or by ground- catalyzing the growth of galaxies. As ways. Personally, I am waiting with
based observatories. wild as this ”early dark energy” scenar- bated breath for the Rubin Observatory
JWST was designed to search for the io may sound, it is not ruled out by ob- to make sensitive observations of the
infrared glow of those early stars, and servations; in fact, it could help explain numerous dwarf satellite galaxies that
the results are already exceeding ex- a small but notable discrepancy seen surround the Milky Way. The num-
pectations. In its first months of opera- between two different ways of measur- ber and distribution of those satellites
tion, the telescope uncovered a startling ing the expansion rate of the universe, are strongly influenced by the prop-
abundance of well-developed galaxies a problem known as the Hubble tension. erties of the galactic halos in which
that appear to have taken shape with- Many diverse groups of scientists are they formed, so they should provide
in the first billion years after the Big investigating these ideas, coming up insights into the nature of dark matter.
Bang. If these observations check out— with models, arguing with one another The dwarfs are also some of the old-
astronomers are now racing to confirm about both observations and predic- est and least evolved galaxies in the
them—then stars must have formed tions, and disagreeing about the un- nearby universe; they may be living
much faster, and in much greater abun- derlying causes. The process may look fossils from the early era of the cosmos
dance, in the early universe than previ- messy or confusing from the outside, that JWST is starting to directly reveal.
ous studies and models indicated. but each of these possibilities is being By combining new observations,
This news has created an anxious carefully vetted. Disagreement between theoretical models, and computer
buzz among those of us who study our expectations and observations is simulations, we are getting closer than
galaxy formation and evolution. Are what drives improved understanding, ever to our goal of understanding how
we misinterpreting the JWST findings? and sometimes overturns established the universe works. What will our pic-
Are we missing something fundamen- models. This situation is the cauldron ture of galaxy formation and cosmol-
tal in our simulations of early galaxy of progress, caught in mid-boil. ogy look like in 20 years? I cannot say,
formation? Or do we need to modify We are likely to learn more soon: and I cannot wait to see.
the underlying cosmological model on These surprises come from a relatively
which those simulations are based? small amount of data taken from just
Perhaps JWST has, by chance, ob- the initial JWST observations. Much Mike Boylan-Kolchin is a theoretical astrophysicist
at the University of Texas at Austin who works
served a highly unusual portion of the more data have already been collected
on galaxy formation theory and its interface with
sky. Perhaps star formation proceeded and are being analyzed. JWST was so
cosmology. Recently, his research has focused on
differently in the very early universe, efficient in getting to its observation lo- near-field cosmology, using detailed studies of near-
when gas clouds did not yet contain the cation (an orbit that keeps it 1 million by galaxies to address a wide variety of questions
heavy elements that were created by miles from Earth) that its nominal five- related to dark matter and galaxy formation physics
later generations of stars. Perhaps we year mission may be extended to 20 across cosmic time. His work combines numerical
are actually seeing emissions from black years or more. And over the coming de- simulations, analytic models, and observations.
holes and confusing it with starlight. Or, cade, many other new o bservatories— Email: mbk@astro.as.utexas.edu
QUICK TAKE
Neutrinos are lightweight particles that IceCube is a telescope at the South Pole Studying how neutrinos change their “fla-
rarely interact with other matter, allowing that can detect these ghostly neutrinos, en- vor” during their journey across the cosmos
them to travel for millions of light-years across abling scientists to trace their origins in dis- is helping to test new theories on quantum
the universe. tant galaxies. gravity and the origin of neutrinos’ mass.
trinos directly, so they have to look for we could not track these neutrinos back
buried in the ice, which can identify neutri-
nos from distant cosmic sources. These detec- indirect signs. When a high-energy neu- to a specific point of origin. Although
tors not only help scientists understand those trino occasionally (very occasionally) in- IceCube’s detectors can roughly work
neutrino sources—such as the roiling clouds teracts with the ice, it produces charged out which part of the sky a neutrino has
of matter around black holes—but also pro- particles that move faster than the speed come from, the observatory does not
vide a way to test new theories at the cutting of light in that medium. The result is have enough resolution to identify an in-
edge of fundamental physics. generated light known as Cherenkov ra- dividual source. It’s hardly surprising—
large-scale evolution of the universe. this phenomenon could make all of the about 100 billionths of a second, the tau
Unfortunately, these theories cannot neutrinos from a source arrive on Earth disintegrates into other particles, releas-
both be completely true. In some cir- as a single flavor, rather than the three- ing more light. So the signature of a tau-
cumstances, they give wildly inconsis- way mix we would otherwise expect. neutrino is a distinctive double flash:
tent or nonsensical answers. To test these far-reaching ideas, we one signaling the production of the tau
Quantum gravity aims to overcome first needed to prove that IceCube can particle, followed by a second one that
these problems by forming a bridge be- detect all flavors of neutrinos—even shows its disintegration. For very high-
tween the two theories. The effects of the super-obscure tau-neutrinos. That energy neutrinos, these two emissions
quantum gravity will be incredibly sub-
tle in most situations, and only become
significant at extremely small distances
or at extremely high energies. One po-
Researchers working with IceCube are
tential signature of quantum gravity in-
volves the breakdown of a fundamental now using the flavor of detected neutrinos
symmetry in the universe. This sym-
metry rests on the idea that there is no to test and improve theories about the
preferred direction in the universe—in
other words, the cosmos has no label fundamental principles of the cosmos.
showing an arrow and the word “up.”
If this symmetry is broken, it means
that there are discernible elements in effort pushed our observatory to its might be only about 50 meters apart in
space-time—glitches in the very fabric limits. Last year, I was part of the team the ice surrounding IceCube, the aver-
of the universe, you might say—that that announced the first detection of age distance a tau with a million times
do indeed point in a particular direc- a tau-neutrino from a cosmic source. the energy of a proton can travel during
tion. And that’s where neutrinos come Because the neutrino almost certainly its fleeting existence.
in. As they travel from their faraway did not start out with its tau flavor, our To look for these signatures, I
sources to Earth, neutrinos can inter- detection confirmed that it must have worked with an international team to
act with these space-time elements in undergone neutrino oscillation during meticulously reanalyze IceCube data
ways that imprint distinct signatures on its vast journey. collected between 2010 and 2017. This
the neutrino flavor states. For example, We were able to identify this tau- tau-neutrino hunt was led by Juliana
some quantum gravity theories sug- neutrino by its unique light signal Stachurska, currently a postdoctoral
gest that the interactions can switch off picked up by our IceCube detectors. researcher at the Massachusetts Insti-
neutrino oscillations, forcing the neu- When a tau-neutrino hits the ice, it pro- tute of Technology.
trino to maintain the same flavor from duces a very short-lived tau particle While our team was at an IceCube
source to detection. In some scenarios, and an accompanying flash of light. In Collaboration meeting in Atlanta, we
Building Knowledge
Henry Petroski
S
imulation and modeling are est thing since sliced bread, but it may will interact with the system into which
virtually synonymous with en- have to forever remain at best a half- it will be built. In the case of a structure
gineering, for an engineer can baked, private thought if it cannot be such as a building or bridge, engineers
hardly proceed with the ana- communicated to someone who can must translate the originator’s idea into
lytical design of a machine, system, engineer it into something concrete. a form that others can grasp unambig-
or structure without having at least The typical first step in doing so is to uously and thus test analytically the
an image of it in mind. That idea may simulate or model the idea through im- structure’s strength, stability, and suf-
conceivably promise to be the great- ages that give it boundaries at which it ficiency for fixing the problem that the
idea aims to solve.
Before Galileo Galilei analyzed the
cantilever beam with the same scien-
tific rigor he had the heavens, design-
ing a physical structure was largely a
matter of craft tradition rooted in trial
and error. As late as the 17th century,
what worked was imitated; what did
not was abandoned. Galileo distin-
guished himself by asking and pursu-
ing in a rational manner the question
of why some structures failed. At the
time, it was well known that ships built
in a geometrically similar way to ones
that were successfully sailing the seas
would themselves be successful—up
to a point. It was natural to build larger
and larger ships based on previously
successful designs, but when the largest
ones failed upon being launched, there
appeared to be a limit to size. But why?
Galileo was able to explain why by
changing the common conception of a
ship. Rather than considering the en-
tire structure, Galileo modeled the ves-
sel at port as a simple beam supported
only at its ends. When the ship leaving
Your work diverges from the idea that where I could see genetic splits be would get stuck on the idea that came
modern humans emerged from a sin- tween populations. I’m also doing first or that was from the most presti
gle population in East Africa. At what analysis that can trace the evolution of gious person. Partial connectivity allows
point did you start to think that there certain words and o bjects—almost do populations to explore and build on a set
were details that didn’t match the con- ing a gene tree, but with objects. of solutions to build a greater repertoire.
ventional models? Even if these populations now For example, you might have a
There is a long tradition of paleo speak languages that are unrelated, population that has optimized fish
anthropology in East Africa. We know it does seem that these people had a ing methods and another one that has
that some of the oldest human fossils very deep past in which they were great antelope hunting techniques.
were found there, so we must have connected to one another. They were You can recombine and even trade
come from there and left. But in the exchanging genetic material, and I’m these ideas. And with genes it happens
past 10 to 15 years, very old, human finding that they exchanged culturally in a similar way. You’re able to access
like things were found in Southern with one another as well. And if the a greater gene pool, even if you’re
Africa and in Morocco. We saw that whole stretch of Central Africa was ex adapted to a local environment.
some of the oldest genetic lineages changing genes and culture, they must We conducted a study in which I
were from populations that now live have been also exchanging genes and compared the bits of DNA that I knew
in other parts of Africa. Some research culture with people a bit further north, had been exchanged between Central
ers started developing a theoretical a bit further south, and to either side. African hunter-gatherers with the struc
model in which humans came from tural diversity of their musical instru
more than one place in Africa. The idea of connectivity networks is ments and of their foraging tools. These
As I was reading this literature, I very important in your research. Do two sets of objects have completely
wondered whether the Central African genetic connectivity and cultural con- different purposes; the musical instru
hunter-gatherers I worked with fit this nectivity go hand in hand? ments might serve as identity markers,
model. If there was one place that could Not necessarily. In order to have a di whereas the tools might be specific to
connect Morocco with Southern Africa verse set of solutions to complex prob a foraging niche. When I did a relation
and East Africa, it was Central Africa, lems, two things are important: inno test between the structure of genetic
and I also knew that hunter-gatherers vation and collaboration. When I talk diversity and the structure of musical
had networks created for mobility. about connectivity being the key to ge instrument diversity, the two matched
I started looking at the genetic netic and cultural diversity and an im almost perfectly—populations that
evidence and environmental models portant element of human evolution, it’s were exchanging genetically were also
and asking, Where do contemporary the idea that different populations living exchanging musically. But for the for
Central African hunter-gatherers live? in different environments that come into aging tools, there was no relationship
Where are the fossils, the few that contact every once in a while are able at all. It just may not make sense to use
we’ve found in Central Africa? Un to optimize these two processes. The the tools that someone in a different
der which environmental conditions? separate populations are each innovat environment is using.
Then I created a model using data on ing, and when they come together and
contemporary hunter-gatherers from exchange those ideas, they’re able to get You are describing deeper and wider
Central Africa to try to predict where the best of what the other one has. levels of connectivity than what’s com-
we’ve found fossils and genetic diver monly represented in the scientific liter-
gences in the past. What is the role of partial connectivity ature. Are there barriers that have kept
I found that the moments in which in this process? people from taking this broader view?
my environmental model predicted If you have a population that’s fully con Traditionally ethnographers who
fragmentation coincided with points nected from the beginning, everybody worked with hunter-gatherers would
experienceIFoRE.org
Sigma Xi’s Annual Meeting and Student Research Conference is now IFoRE
www.americanscientist.org Special Issue: Scientific Modeling 2023 July–August 237
2023-07Q&APadila-Iglesias.indd
IFoRE JULY-AUG AD.indd 1 237 6/7/2023 1:50:13
5/18/2023 11:55:20PM
AM
Farming and the Risk
of Flash Droughts
In the coming decades, every major food-growing region faces
increasing incidences of sudden water shortage.
Jeffrey Basara and Jordan Christian
F
lash droughts develop fast,
and when they hit at the
wrong time, they can devas-
tate a region’s agriculture.
They’re also becoming increasingly
common as the planet warms.
In a study we published this year in
the journal Communications Earth and En-
vironment, we found that the risk of flash
droughts, which can develop in the span
of a few weeks, is on pace to rise in ev-
ery major agriculture region around the
world in the coming decades.
In North America and Europe,
cropland that had a 32 percent annual
chance of a flash drought a few years
Left: NASA Earth Observatory/Lauren Dauphin/Christopher Hain; right: Pasquale Mingarelli/Alamy Stock Photo
ago could have as much as a 53 per-
cent annual chance of a flash drought
by the final decades of this century.
That result would put food produc-
tion, energy, and water supplies un-
der increasing pressure. The cost of
damage will also rise. A flash drought
in the Dakotas and Montana in 2017
caused $2.6 billion in agricultural dam-
age in the United States alone.
QUICK TAKE
The warming climate has accelerated the al- When certain conditions combine, flash Better prediction of flash droughts could
ready increasing occurrence of flash droughts, droughts can arise over the span of just a help farmers and ranchers plan for a chang-
which can cause crops to fail as well as put few weeks, quickly pulling moisture from the ing future, which modeling shows will only
water and energy supplies at risk. ground surface and plants. see further increases in flash droughts.
Can Agriculture Adapt? Flash droughts affect agriculture directly, and they can also interrupt the transport of crops and
One way to help agriculture adapt to even people. During Europe’s 2022 drought, which began as a flash drought and then intensified,
house boats in the Netherlands were left stranded in a dry riverbed.
the rising risk is to improve forecasts for
rainfall and temperature, which can help tists. For example, the U.S. Drought Nothing is getting easier for farm-
farmers as they make crucial decisions, Monitor at the University of Nebraska- ers and ranchers around the world as
such as whether they’ll plant or not. Lincoln has developed an experimen- global temperatures continue to rise.
When we talk with farmers and tal short-term map that can display de- Understanding the risk they could
ranchers, they want to know what veloping flash droughts. As scientists face from flash droughts will help
the weather will look like over the learn more about the conditions that them—and anyone concerned with
next one to six months. Meteorology cause flash droughts and about their water resources—manage yet another
is pretty adept at short-term forecasts frequency and intensity, forecasts and challenge of the future.
that look out a couple of weeks, and monitoring tools will improve.
Bibliography
Christian, J. I., et al. 2023. Global projections
of flash drought show increased risk in a
Cropland in much of North America and warming climate. Communications Earth and
Environment 4:165.
Europe would have a 49 and 53 percent an- Christian, J. I., et al. 2021. Global distribution,
trends, and drivers of flash drought occur-
rence. Nature Communications 12:6330.
nual chance of flash droughts, respectively, Otkin, J. A., et al. 2018. Flash droughts: A
review and assessment of the challenges
by the final decades of this century. imposed by rapid-onset droughts in the
United States. Bulletin of the American Meteo-
rological Society 99:911–919.
at longer-term climate forecasts using Increasing awareness can also help. Jeffrey Basara is an associate professor at the Uni-
versity of Oklahoma with a joint appointment be-
computer models. But flash droughts If short-term forecasts show that an
tween the School of Meteorology and the School of
evolve in a midrange window of time area is not likely to get its usual pre- Civil Engineering and Environmental Science. Jor-
that is difficult to forecast. cipitation, that should immediately set dan Christian is a postdoctoral researcher in meteo-
We’re tackling the challenge of off alarm bells. If forecasters are also rology at the University of Oklahoma. This article
monitoring and improving the lead seeing the potential for increased tem- is adapted and expanded from one that appeared
time and accuracy of forecasts for peratures, that heightens the risk for a in The Conversation, theconversation.com.
flash droughts, as are other scien- flash drought’s developing. Email for Basara: jbasara@ou.edu
Abel Méndez
E
xtraterrestrial life is a genu- rather a fuzzy comparison between We are still very early in the pro-
ine scientific possibility. Earth a selected set of planetary properties cess of identifying exoplanets that are
proves that at least one hab- with Earth—primarily size, mass, not just broadly Earth-like but that
itable planet, even a planet and insolation, the amount of radia- seem plausibly habitable. The first
with technological intelligence, can tion received from the planet’s star— claimed potentially habitable exoplan-
exist. Astronomers have discovered because those details are often all that ets, discovered around the nearby red
more than 5,000 exoplanets—worlds we are able to measure for exoplanets. dwarf star Gliese 581, turned out to
circling other stars—and that number Worlds with high ESI values are not be illusions. Then in 2012, a team of
increases weekly. Now the search for necessarily more habitable, because so astronomers led by Guillem Anglada-
habitable worlds is on, pushing our many other factors influence a plan- Escudé of the University of Göttingen
data and models to the limits. etary environment. Venus (ESI = 0.78) in Germany announced the discovery
Most known exoplanets have been and Mars (ESI = 0.65) are similar in size of Gliese 667C c, the first confirmed
detected indirectly, primarily using and insolation to Earth, but they have potentially habitable world. Today, we
two techniques. The transit method starkly different surface conditions. know of about 60 potentially habitable
measures the slight dimming of a star Most known exoplanets have ESI planets that have the right size and
that occurs when a planet passes in values well below that of our Moon orbit to hold surface water. However,
front of it. The radial velocity method (ESI = 0.60). There are a few exoplanets the world “potentially” carries a lot of
measures tiny shifts in a star’s light with ESI values close to one. Maybe weight here. We do not know anything
spectrum caused by the back-and- their surfaces are genuinely similar to about these planets’ atmospheres or
forth gravitational pull of orbiting Earth’s, with oceans and continents. surfaces yet. (See “Decoding Light from
planets. The amount of light blocked Then again, they could be unsuitable Distant Worlds,” May–June 2020.) They
during a transit indicates the size of for any life because they lack other life could be barren, Moon-like worlds, or
a planet. The intensity of a radial ve- requirements, such as water, or have something else unexpected.
locity change indicates a planet’s ap- harmful or toxic conditions. We don’t Even today’s most powerful tele-
proximate mass. When combining the know how likely it is that superficially scopes, such as the James Webb Space
two methods is possible, the results Earth-like planets truly resemble our Telescope (JWST), are able to analyze
reveal the planet’s density. And that own; that is a major goal for the next few the atmospheric composition of only a
is where our knowledge of most exo- decades of astronomical observations. few, particularly well-situated exoplan-
planets ends. The ESI is also relevant only for surface ets. We will need far better telescopes,
To complement the search for Earth- life, not subsurface life that could exist such as NASA’s proposed Habitable
like planets, I have created an Earth independent of the star’s energy, much Worlds Observatory, to discover more
Similarity Index (ESI) that indicates like the life around geothermal vents in Earth-like planets, observe some of
how much planets resemble Earth in Earth’s deep oceans. For example, astro- them directly, and provide more infor-
their gross physical properties. The ESI biologists consider it possible that there mation about their atmospheres, sur-
is a number between zero (no similar- could be life in the oceans under the icy faces, and suitability for life. Detailed
ity) and one (identical to Earth). It is surface of Jupiter’s moon Europa, even theoretical studies about habitability
not a direct measure of habitability, but though it has an ESI of 0.22. and planetary modeling will then be es-
QUICK TAKE
Potentially habitable planets require rocky No true Earth analog has yet been found, Gases associated with life on Earth, such as
surfaces with liquid water and a substantial but the author has developed an Earth Simi- oxygen alongside methane, could indicate bio-
but not overwhelming atmosphere—at least, larity Index to identify planets that broadly logical activity on another planet. The search
those are requirements for life as we know it. resemble our own and deserve closer scrutiny. for chemical biosignatures is starting now.
starlight + starlight +
planet dayside planet dayside
starlight
only
NASA, ESA, CSA, J. Olmsted. Data: Thomas Greene, Taylor Bell, Elsa Ducrot, Pierre-Olivier Lagage
Planets around red dwarf stars are relatively easy to study, making them natural starting Exoplanet surveys indicate that there
points in the search for habitable conditions. Recent observations of planet TRAPPIST-1 b are, on average, one to two planets
(illustrated here) detected no atmosphere when it passed in front of its star (inset), but models per star—and yes, that includes all the
had predicted it would be airless. Future studies will target TRAPPIST-1’s six other Earth- stars we see at night. What fraction of
sized planets, three of which orbit in the habitable zone around their stars.
those stars have a habitable world is
an open question. The rarer such plan-
sential for interpreting the results. The is universally applicable to all extrater- ets are, the less likely we are to have
confident detection of life won’t be easy. restrial life, or if it is a result of the evo- one close enough to study using cur-
Using current technology, we would lution of our particular planet. Never rent telescopes. The term Eta-Earth is
have to get extremely lucky to get a clear theless, liquid water is essential to life used to refer to the average number of
indication of life, or even a clear signal on Earth, so astrobiologists start from Earth-sized planets, in orbits that allow
of habitability. Then again, if we do get the assumption that habitable exoplan- Earth-like temperatures, around each
lucky, we might start to get an answer as ets must have a surface temperature star in our galaxy. NASA’s Kepler space
soon as this decade. Even if it takes us a compatible with liquid water. telescope, which operated from 2009
long time to get there, we are on the path Life has many requirements beyond to 2018, was the first mission to pro-
to discovery—which is truly exciting. water, however. It requires an environ- vide empirical estimates of Eta-Earth.
ment where the three phases of mat- Upper estimates of Eta-Earth put the
What Life Needs ter (gas, liquid, and solid) plus energy value around 0.5, which means that less
In astrobiology, habitability is defined as coexist for a long time. The classical than half of the stars we see at night
the overall suitability of an environment elements (air, water, earth, and fire) are might have planets with the fundamen-
for biological processes. A habitabil- helpful analogies for understanding tal properties for habitability: not just a
ity analysis cannot determine whether this requirement. suitable orbit, but also a suitable size.
a world is able to sustain life; only life The three phases of matter are nec- The right size is another key habit-
detection experiments can do that. Even essary because the elements used to ability factor, because life as we know it
with that restriction, researchers can- construct the molecule of life, DNA (hy- needs a rocky surface and a substantial
not study all the factors affecting habit- drogen, carbon, nitrogen, oxygen, and but not excessive atmosphere. Planets
ability, so we focus on the ones that are phosphorus), are not present in a single in the habitable zone must be between
most important and that are relatively phase. For example, nitrogen is usu- 0.5 (Mars-sized) and 1.5 times the di-
straightforward to measure. ally found as a gas in the atmosphere, ameter of Earth to meet those require-
The right temperature is a key con- whereas inorganic sulfur and phospho- ments. Worlds that are too small will
sideration for habitability. Most animal rus are typically found in rocks as sul- have a rocky surface but won’t be able
life on Earth requires temperatures be- fates and phosphates. Furthermore, the to hold a significant atmosphere, just
tween 0 degrees to 50 degrees Celsius, solvent, diffusion, and flow properties like Mars. Worlds that are too large will
and the optimum is around 25 degrees, of air and water enable these elements have too much atmosphere and no ac-
not far from Earth’s average global to mix with rock and become readily cessible rocky surface; an atmosphere
temperature of 15 degrees. Microbial available in any habitat. On a planetary of just 1 percent of the planet’s mass
life has an extended thermal range if scale, the classical elements become the would create pressures so extreme that
the water is still liquid. We don’t know atmosphere (air), hydrosphere (water), surface water would remain solid no
whether this principle of biochemistry lithosphere (earth), and insolation (fire). matter what the temperature.
Abel Méndez/adapted by Barbara Aulicino; photos: Wikimedia Commons (3), Travel Stock/Shutterstock, Smileus Images/Alamy Stock Photo
The right orbit of a planet around and luminosity of the star, but also on oceans to be converted into steam. Stel-
its star is related to both the size and the planet’s properties, such as its size lar wind and radiation will then erode
temperature requirements for habitabil- and its atmospheric composition. For this atmospheric water vapor, decom-
ity. The stellar habitable zone is defined example, the boundaries for a star like posing it into hydrogen (which easily
as the region around a star where or- our Sun extend from about 0.95 to 1.37 escapes into space) and oxygen (which
biting planets potentially receive the astronomical units (the distance be- quickly reacts with other atmospheric or
surface compounds). Venus may have
had early oceans that were lost in this
way. The outer limit of the habitable
Surveys indicate that there are, on average, zone is constrained by the greenhouse
capacity of carbon dioxide to keep tem-
one to two planets per star—and yes, that peratures above freezing. Mars, with its
relatively low gravity, lost its water and
includes all the stars we see at night. most of its initial atmosphere, leaving it
cold and inhospitable.
The right type of star is necessary to
have a long-lived, potentially Earth-
right amount of warming radiation to tween Earth and the Sun), but the zone like planet. Small stars generally have
support liquid water on their surfac- could be smaller or larger depending small planets, whereas big stars have
es. This region is sometimes called the on the specific characteristics of a plan- more giant planets. About 75 percent
“Goldilocks zone” because planets in et. Planets with thick, heat-trapping at- of the stars in our galaxy are red dwarf
that zone are neither too hot nor cold, mospheres, like Venus, may be able to stars (known to astronomers as type M
but are just right for life as we know support liquid water at greater distanc- stars), which are more prone to having
it. If a planet is too close to its star, its es from their stars than planets with Earth-sized planets. Only 19 percent of
surface temperature will be too high thin atmospheres, like Mars. stars are Sunlike (referred to as types F,
for liquid water to exist, whereas if it The inner edge of the habitable zone G, or K); they are more likely to hold
is too far away, the temperatures will is constrained by warming induced a mix of Earth-sized planets and giant
be too low. (Those considerations don’t by the evaporation of water, a strong worlds like Jupiter and Saturn. Our
exclude subsurface life, however.) greenhouse gas. Once the average tem- Sun is a typical G star, but we don’t
The boundaries of the habitable zone peratures exceed about 60 degrees, a know if G stars are uniquely or even
depend on the effective temperature runaway feedback effect causes all the especially well-suited to life.
Venus Mars
Kepler-452 b Earth
stellar temperature (K)
5,000
Kepler-62 f
Kepler-442 b
4,000
K2-18 b
GJ 667 C c
Proxima Cen b
Teegarden's Star b
3,000
TRAPPIST-1 h
TRAPPIST-1 g
TRAPPIST-1 d TRAPPIST-1 e TRAPPIST-1 f
S c i e n t i s t s’
Nightstand
individual fixes. For instance, the Na-
The Scientists’ Nightstand, Glitching Out tional Football League (NFL) famously
used race in its algorithms to deter-
American Scientist’s books mine payouts for players with brain
section, offers reviews, review injuries, with Black players receiving
Ashley Shew lower amounts by thousands of dol-
essays, brief excerpts, and more.
For additional books coverage, More than a Glitch: Confronting lars. The NFL’s individual fixes in-
please see our Science Culture Race, Gender, and Ability Bias in Tech. cluded eliminating data such as race,
blog channel online, which Meredith Broussard. 248 pp. The MIT and paying penalties, although only
explores how science intersects Press, 2023. $26.95. after they were ordered to by a federal
judge. But algorithmic biases are more
T
with other areas of knowledge,
he biases caused by and perpet- than a temporary blip or something
entertainment, and society. uated through algorithms are no we can deal with through individual
glitch: They are consistent with patches; simply fixing blip by blip is
ALSO IN THIS ISSUE our world as it exists now. In More than not a structural solution and won’t let
a Glitch: Confronting Race, Gender, and us move quickly enough. Our biases
MODELS AND MATHEMATICS: Ability Bias in Tech, Meredith Brous- are systemic and structural, and they
Q&A WITH ERICA THOMPSON sard offers a thorough exploration of are reflections of the world.
page 251 how algorithmic systems encode and By the end of the book, Broussard
perpetuate injustice—and what we can offers clear ways to recognize and
ONLINE do about it. manage biases and work toward a
Broussard’s 2018 Artificial Unintelli- more just world through public inter-
On our Science Culture blog: gence: How Computers Misunderstand the est technology and the use of algorith-
www.americanscientist.org/blogs/ World demonstrated how we’ve been mic audits. She highlights the work of
science-culture misled by tech enthusiasts and devel- groups already doing tech monitor-
opers in their quest for artificial general ing and reporting, along with many
Rethinking Menstrual Norms intelligence (AGI). More than a Glitch recent cases in which tech companies
Digital features editor Katie is no less important to understanding have failed us, and she suggests how
Burke reviews Kate Clancy's Peri- where we are now with tech. Explain- we can add to these efforts through
od: The Real Story of Menstruation. ing the current state of algorithmic in- different types of algorithmic audit-
justice, she offers strong calls to action, ing and an emphasis on public interest
along with very concrete steps, to help technology. She calls for the public to
us manage, mitigate, and recognize the “hold algorithms and their creators ac-
biases built into algorithms of all types. countable” to the public, and demon-
A thread running through the book is strates ways we can do just that.
the idea of technochauvinism, a “bias that The book is divided into 11 chapters,
considers computational solutions to be each of which could be read on its own
superior to all other kinds of solutions.” or in concert with other readings. Each
This bias insists that computers are chapter is grounded in many contem-
neutral and that algorithms are fair. But porary cases of biases built into algo-
this bias is its own type of ignorance, rithms that then perpetuate or cause
because what we build into algorithms discriminatory harm. Her journalis-
captures and perpetuates existing social tic approach makes this work come
biases, all while being touted as “fairer” alive through accounts of not only the
because of the assumed neutrality of way algorithms work, but also of how
algorithms. the use of algorithms has caused real
Broussard defines glitch as “a harm, including arrests, educational
temporary blip,” which is how tech setbacks, administrative nightmares,
boosters often want to explain away and even death.
publicized instances of bias in tech as One compelling example from
one-offs that can be dealt with through Chapter 3 is from the area of facial rec-
Models are primarily useful as What do you think is the most impor-
tant thing we should keep in mind for
working with models in the future, and
metaphors and aids to thinking, rather where do you see your work going?
The most important thing to keep in
than prediction engines. mind when working with models is
that any model can only give us one
perspective. It can’t tell the whole
story. A photograph is great to show
could become unreliable on some time prediction engines, and so you might you what someone looks like, but it
scale. The hawkmoth effect is a similar think of the “flatten the curve” concept doesn’t tell you their political opinions
idea, but relating to the accuracy of the that was so important in the spring of or what they want to have for lunch.
model rather than the data. If the mod- 2020, or more qualitative concepts like When we make and use models, we
el is slightly “wrong,” even by just a the idea that the national budget either need to keep in mind what they are
tiny amount, then the forecast it makes is or is not like a household budget. good at and what they aren’t good at.
could be significantly wrong if you are In that light, I hope it becomes clear- My own work has two strands, a
predicting far enough into the future. er how models also change the way mathematical strand about the statis-
Of course, the problem is that just that we think: If you have a particular tics of model calibration and evalua-
as all data are subject to uncertainties model (mathematical or conceptual) tion, and a more sociopolitical strand
(measurement errors), also “all models for something in the real world, then about the value judgements that are
are wrong” in the sense that we can you use that model as a tool for un- embedded in models—both illustrated
never really know that we have fully derstanding, and as a tool for commu- through case studies of different kinds
represented absolutely everything that nicating with other people. It can be of models, from public health to finance
is important about the system. In some illuminating by helping us to think in and climate change. I’m working on
cases, like the weather forecast, we new ways, but the model and the lim- bringing these two strands closer to-
have lots of past evidence from our its of the metaphor also constrain the gether, to learn from each other and
successful predictions, which help us ways that we are able to think about hopefully to improve the usefulness of
have confidence; we know that tomor- the system. the modeling methods that are so im-
row’s weather forecast is pretty good, To take an example, if you construct portant for today’s decision-making.
next week’s is indicative but not all a simple epidemiological model for
that reliable, and we wouldn’t even a disease outbreak, you are focusing
look at the forecast six months in ad- (probably rightly) on the infection Jaime Herndon is American Scientist's acting
vance because we know, based on past and its consequences, and you might book review editor.
Sigma Xi Today A NEWSLETTER OF SIGMA XI, THE SCIENTIFIC RESEARCH HONOR SOCIETY
On March 30, 2023, Sigma Xi installed Wang. The founding members will The ceremony was also attended by
its newest chapter at the Elmezzi Grad guide the chapter through its initial Sigma Xi’s executive director and CEO,
uate School of Molecular Medicine at years of development and activity, Jamie Vernon, and Sigma Xi’s director
Northwell Health. The ceremony was including recurring meetings, events, of membership and chapters, Richard
held in person in Manhasset, New membership growth, and participation Watkins. The CEO of Northwell Health,
York, and presided over by Sigma Xi in Sigma Xi’s annual conference, the Michael Dowling, delivered the key
then President-Elect Marija Strojnik. It International Forum on Research Excel note address.
was a celebration of the new chapter’s lence (IFoRE). The Elmezzi Graduate School of
officers, members, and commitment Molecular Medicine at Northwell
to growth and advancement of the Health, located in Manhasset, New
school’s research enterprise. York, is an individually tailored and
Annette Lee, dean of the Elmezzi accelerated three-year doctoral pro
Graduate School, delivered the inau gram awarding a PhD degree to
gural presidential address. She will individuals who hold an MD or equiv
lead a group of founding members alent. The graduate school directly
that include Yousef Al-Abed, Lance addresses the crucial need for phy
Becker, Betty Diamond, Daniel Grande, sician-scientists who are trained to
Christine Metz, Barbara Sherry, Bet conduct translational and clinically
tie Steinberg, Kevin Tracey, and Ping relevant research.
of GIAR
April Stabbins
Grant: $1,000 in Fall 2020
Education level at time of the grant: PhD student
Project Description: The project sought to investigate how certain coral species interact with
methane seeps in the deep sea. By receiving this grant, I was able to collaborate with researchers
from other laboratories and complete more detailed analysis with advanced microanalytical
techniques that otherwise would not have been possible.
How did the grant process or the project itself influence you as a scientist/researcher? My
first application was not accepted, but this outcome caused me to go back and review my pro
posal in detail to make adjustments. Being able to take a step back and review how I describe
my work to others, especially those not in my field, has influenced how I have applied for
other grants and how I write manuscripts.
Where are you now? I am a fifth year PhD candidate at Temple University in Philadelphia. I
plan to graduate in the next year and am now looking for a postdoctoral position.
Chhandak Basu
Grant: $1,000 in Fall 2002
Education level at time of the grant: PhD student
Project Description: Thanks to the GIAR funding, I was able to visit and work in the labora
tory of Brian Friskensky, Department of Plant Science, University of Manitoba, Canada. As
part of my PhD thesis, I evaluated several gene promoters that are suitable for the genetic
transformation of plants. I was fascinated to examine gene expression in plant cells at a tran
scriptome level. However, this was not a direct topic of my research. My major professor was
very supportive of my idea to visit another laboratory to learn techniques for studying stress-
responsive gene expression in plants.
As a result, I visited Dr. Friskensky’s lab to learn how transcription analysis with cDNA-
macroarray can be used to study gene expression patterns after pathogen attack in canola
cells. My research involved preparing pathogen-induced cDNA from canola cells and
mastering macroarray techniques. Various genes were nonradioactively labeled and hybrid
ized onto membranes containing cDNA from canola cells. In addition to understanding
pathogen-induced gene expression in plants, the genotyping data could also be used to
develop disease-resistant crops.
How did the grant process or the project itself influence you as a scientist/researcher? Working in a different lab abroad,
outside my familiar environment, was an experience I will never forget. Having this research experience motivated me to
become an independent researcher. Within five years of receiving the GIAR fellowship, I attended a week-long workshop at
the University of Arizona, Tucson, to learn microarray. As of now, my lab focuses on transcriptomic-level gene expression stud
ies in plants, and therefore I am indebted to Sigma Xi for granting me the GIAR award.
Where are you now? I am currently a professor of biology at California State University, Northridge. As part of my research,
I study how plant cells respond to environmental stress at the molecular level. Additionally, I have a research affiliate appoint
ment at NASA’s Jet Propulsion Laboratory. I conduct research at NASA on how microbes survive in hostile environments.