CASE STUDY

ADSORPTION ENTHALPIES IN ZEOLITES .

HIGH-ACCURACY, MACHINE LEARNING

A team of researchers led by 􀁉 rofessor Dario Rocca at the

Universite de Lorraine computed adsorption entalpies in
zeolites with high accuracy. Calculations were based on the
random phase approximation (RPA), an accurate but highly
expensive quantum mechanical approach.

The studies were made feasible by coupling molecular

dynamics simulations with machine learning. High-memory
nodes were instrumental to models zeolites with over 500
valence electrons per cell. At the peak, over 400 zeolite
configurations were studied simultaneously employing over
12,800 cores.

RAPID EXPLORATION OF NEW
CATALYTIC PATHWAYS
The possibility of achieving highly accurate
results using machine learning and random-
phase approximation was demonstrated for
Zeolite-based porous compounds.
Outcome
STATE-OF-THE-ART COMPUTING
CAPABILITIES
Scaling to 12,800 computing cores and utilizing
ultra-high memory nodes with over 600GB
memory was demonstrated in order to achieve
high accuracy for the dataset obtained.
VERSATILE PLATFORM-LEVEL
AI/ML APPROACH
The workflows, materials structures,
and properties calculated are organized
in an online platform and used to build
a predictive AI/ML approach.

Challenge
Predicting the properties of porous catalytic compounds
from ab-initio modeling and simulations is often considered
beyond today’s computing capabilities due to the processes
that occur at a significantly large time- and length scales.
Efficiently modeling the effects of small-molecule (e.g. CO2)
adsorption is especially challenging.
Solution
Scalable on-demand supercomputing resources paired with
accessible and data-centric digital tools allow to rapidly and
intuitively deploy high-fidelity simulations to obtain data on
a diverse landscape of structural variations. The data
produced is automatically organized and made available to
AI/ML models allowing for high accuracy predictions.

“The platform allowed us to leverage the latest advancements in cloud-based

high-performance computing and complete the calculations that otherwise would
be impossible at other computational facilities.“

Universit’e de Lorraine, Prof. Dario Rocca

1
CASE STUDY: ADSOPRTION ENERGY
Method
Zeolites are microporous materials intensively investigated for
their ability to selectively adsorb target compounds, including
carbon dioxide, methane and other gases. This property is used in
numerous applications such as depollu-tion, chemical catalysis
and/or separation of chemicals. Quantitatively reliable predictions
of adsorption energies can presently be obtained only by using
sophisticated (and usually prohibitively computationally
expensive) quantum mechanical models that accurately describe
the interactions in the systems of interest. Including temperature
effects through molecular dynamics simulations is also required.
finite-temperature RPA energy is computed by properly
In this study we used a state-of-the-art methodology based on
machine learning (ML) and molecular dynamics (MD) to compute
adsorption enthalpies at the random phase approximation (RPA)
level of theory, which is considered as one of the most accurate
but highly demanding quantum mechanical materials modeling
technique.
Figure 1. Visual sequence illustrating the essence of the work done in this paper. Adsorption of molecules in zeolites was simulated at the atomic scale using highly sophisticated quantum
mechanical approximations. The figure shows the specific example of chabazite, that has been studied with a periodic model containing about 40 atoms in the primitive cell.
Introduction
Thermodynamic perturbation theory (TPT) is a numerical
approach used to compute free energy differences in classi-cal
molecular dynamics simulations [1]. Here TPT is used in the
context of ab initio MD to obtain finite temperature properties
using high-accuracy/high-cost correlated approximations for the
electronic correlation [2]. This can be achieved with through: (1)
ab-initio MD simulation performed using a numerically robust
approximation (i.e. PBE); (2) RPA RPA calculations for each one of
the configu-rations generated by the MD in step 1; (3) finally, the
reweighting the contribution of the configurations sampled in the
first step. By computing internal energies of clean zeolite,
standalone molecule, and adsorbed system the enthalpy of
adsorption is subsequently obtained.
From a practical point of view the step above (2) is highly
demanding if feasible at all. For the largest system consid-ered in
this project this computation would require hundreds of millions
of core hours. This issue can be overcome by using a machine
learning approach. As shown in Fig. 2, by selecting a small number
of training examples (red dots) a ML regres-sion model can be
used to predict a certain property (i.e. the RPA total energy) for
many other structures. In this way the number of costly RPA
calculations is reduced to a few dozens (a maximum of 200
configurations was considered for a single system). Because of
the high requirements in terms of the computational time and
runtime memory, the RPA ener-gies have been computed on the
exabyte.io platform using high-memory nodes available from the
Google Cloud.
2
CASE STUDY: METALLIC ALLOYS

RESULTS
In Table 1. the enthalpies of adsorption (in kcal/mol) are presented for CH4
and CO2 adsorbed in protonated chaba-zite (denoted as HCHAB) and
siliceous chabazite (CHAB). The four corresponding adsorbed systems are
indicated as HCHAB+CH4, HCHAB+CO2, CHAB+CH4, and CHAB+CO2.
Additionally, results are presented for a much larger system, DON+CH4,
containing about 100 atoms in the primitive cell. RPA enthalpies of
adsorption are generally within chemical accuracy from experiment.

Table 1. Enthalpies of adsorption (in kcal/mol) of small molecules in zeolites computed at the RPA level Figure 2. Sketch illustrating a machine learning regression
of theory; experimental results are provided as reference. model in materials science. Starting from a few training
examples (red dots), the dashed curve is used to approxi-mate
the computationally demanding quantum mechanical results
represented by the continuous curve. (п¬Ѓgure from [4])

COMPUTATIONAL DETAILS
In order to highlight the computational challenges involved in this project it is
important to mention that the models of zeolites contained up to a maximum of
520 valence elec-trons in the primitive cell (DON+CH4) and that the RPA has
unfavorable scaling with the 4th power of the system size.

Calculations were performed on the exabyte.io platform using the RPA

implementation in VASP [5]. Because of the memory requirements, the use of
Google Cloud high-memo-ry nodes was instrumental for the success of the
project. At the peak computations of up to 400 zeolite configurations were
performed at the same time on 12,800 cores for a period of 12 hours.

Dario Rocca is an Associate Professor of physics at the Universit'e de Lorraine in

Nancy, France. He obtained a PhD from Sissa, Trieste (Italy) and was a postdoctoral
researcher at the University of California, Davis (USA). His research concerns the
development and application of advanced methodologies to describe the properties
of materials.

References: [1] A. Pohorille, C. Jarzynski, and C. Chipot, The Journal of Physical [3] S. Grimme, J. of computational chemistry 27, 1787 (2006).

Chemistry B 114, 10235 (2010).

[4] M. Rupp, International Journal of Quantum Chemistry 115, 1058 (2015).

[2] D. Rocca, A. Dixit, M. Badawi, S. Leb`egue, T. Gould, and T. Bucko, [5] J. Harl and G. Kresse, Phys. Rev. B 77, 045136 (2008).
Physical Review Materials 3, 040801 (2019).