Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/333981206

Machine-Learning in Oil and Gas Exploration: A New Approach to Geological


Risk Assessment

Conference Paper · June 2019


DOI: 10.3997/2214-4609.201900988

CITATIONS READS
0 802

9 authors, including:

Francisco Silva Susana Fernandes


Galp Energia Galp Energia
8 PUBLICATIONS   119 CITATIONS    9 PUBLICATIONS   29 CITATIONS   

SEE PROFILE SEE PROFILE

João Casacão Carlos Raoni De Alencar Mendes


Galp Energia IBM
22 PUBLICATIONS   1 CITATION    6 PUBLICATIONS   13 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Semiotic Engineering Tools for Human-Centered Software Development View project

Petroleum Systems at the Western Iberian Margin, NE Atlantic View project

All content following this page was uploaded by João Casacão on 24 June 2019.

The user has requested enhancement of the downloaded file.


Introduction

During the last decade, machine learning has been exponentially evolving to become a complementary
tool in geoscience research, frequently associated with the use of big data to analyse information, build
geological models and aid commercial decisions. In the oil and gas industry, the continuous need to
renew a company’s prospect portfolio is supported by a fair ranking (e.g. Surovtsev et al. 2019),
ultimately reducing drilling costs and ensuring the company’s successful exploration program by
optimizing long-term decisions.

As a key step in exploration management, prospect portfolio is dependent on geological risk assessment,
which, by principle, should be systematically performed to account for different team views (e.g. White,
1993), across different geographies. During this assessment, exploration geoscientists routinely estimate
the Probability of Success (PoS), vital to portfolio management (Smalley et al. 2008) and the final step
on prospectivity evaluation prior to economic assessment.

Oftentimes, geological PoS estimation for each individual risk factor is based on subjectivity related
with subconscious biases and heuristic judgments and fallacies (e.g. Milkov, 2015; Polson & Curtis,
2010). The biases we propose to address are controlled by an individual’s past experiences, such as
personal and technical knowledge, recent exploratory successes or failures, under or overconfidence,
instinctive ideas and external pressures. All of these preconceived judgements impact the accuracy of a
given assessment, ultimately resulting in poorer management decisions. The rationale behind our
research project is to constrain these biases, supported by cognitive advisors to foster consistency and
overcome the frequently associated subjectivity.

Methodology

In this research, we propose a new approach to geological risk assessment. This methodology is
supported by a Level of Knowledge (LoK) assessment, which is dependent on the existing data and
models for a given prospect evaluation. The LoK is assessed from a standard characterization and
evaluation of each geological factor, anchored to a powerful source of information, the Knowledge Base
(KB). The KB is a triplestore database that structures all the peer-reviewed inputs and other unstructured
information, allowing semantic queries on the modelled knowledge. Information is fed by both internal
assessments and external inputs (e.g. from papers, public data). The proposed approach is to collect
feedback from users and retrieve peer-reviewed information as an input to machine learning methods,
providing a consistent knowledge ranking that is continuously improved as the KB is fed. Methodology
evolution is supported in a controlled environment – a company, or a user – by adapting the geological
characterization and/or the algorithms behind it (Fig. 1).

The process starts with a consistent characterization of each geological factor (timing and migration,
reservoir, trap geometry, recovery effectiveness, and seal (Fig. 1A) – adapted from the classification
proposed by Otis & Schneidermann, 1997; CCOP, 2000; Rose, 2001; Milkov, 2015) and simultaneous
cognitive advisor suggestions based on relevant information from the KB (e.g. including public research
papers, internal data or reports, and other public data). This process supports the recognition of similar
characterizations for each risk factor, through a similarity algorithm, allowing end-users to compare
LoK metrics between the evaluated prospect and all previous characterizations stored in the KB (Fig.
1B). A rule-based method guarantees the consistency for the users’ LoK comparisons (Fig. 1C). Then,
they are used as training examples for a supervised machine-learning algorithm that models a LoK
ranking function and influences the similarity algorithm (Fig. 1D). This allows for a continuous
enhancement of the algorithms and the process itself. The LoK of the evaluated prospect is inferred
from the trained function. Finally, the estimated PoS values from similar prospects stored in the KB are
retrieved to advise the user on its current evaluation (Fig. 1E).

81st EAGE Conference & Exhibition 2019


3-6 June 2019, London, UK
Figure 1 The workflow of the Geological Risk Assessment methodology can be described in four
different stages, from A to D. These four stages can be defined or fed by the user (A, C), and by the
GRA advisor (A, B and D), leading to the final output addressed in stage E, the Probability of Success
(PoS). The Risk Assessment Advisor in B gives the initial Level of Knowledge (LoK) derived from the
Geological Characterization for each risk factor (A). The final LoK is obtained from the gathering of
all comparisons using ML algorithms. The estimated and constrained LoK offers a PoS suggestion for
each risk factor, which, if accepted by the user, leads to the final prospect PoS. These processes run
under a Knowledge Base environment system, advising the user with relevant data and incorporating
the final outcomes.

Case Study

Oftentimes, in prospect risk assessment, geoscientists tend to give different values of Probability of
Success when using the same dataset (e.g. Polson & Curtis, 2010; Milkov, 2015). With this new
methodology, where final PoS’s are constrained by the Level of Knowledge, we wanted to assess if the
same divergent results were identified using this LoK metric. In order to test this scenario, we asked the
Exploration team members at Galp to compare the Level of Knowledge for a given risk factor, between
two different prospects. In this exercise, 34 geoscientists with various experience in the oil and gas
industry were asked to compare the Level of Knowledge of the trap geometry risk factor for two
different prospects.

The assessment of trap geometry is commonly based on the reliability of the structure picking across
the seismic data, considering also its structural configuration. Therefore, geoscientists should assume
that the LoK for this particular risk factor relies mostly on seismic type, quality, and coverage, but also
the structural complexity of the trap, and the uncertainty of the velocity model. The trap characteristics
for both prospects are defined in Table 1.

81st EAGE Conference & Exhibition 2019


3-6 June 2019, London, UK
Table 1 Characterization of different trap geometry characteristics for Prospect A and Prospect B
Trap geometry characteristics Prospect A Prospect B
Seismic data type 3D 3D
Seismic data quality Good Good
Seismic tied to the well? No No
How do you evaluate the structure Easy/ reliable
Easy/ reliable correlation
interpretation? correlation
Structure complexity 4-way 3-way
Structure relief High-relief Low-relief
Uncertainty on depth conversion and/or
- Yes
rapidly changing lateral velocities?

Similarity algorithms evaluate these characterizations and provide a numerical value of the similarity
between prospects A (reference prospect) and B. A first LoK assessment is derived based only on the
characterization (as depicted on Fig. 1B). Then, users are asked to do comparisons between the
evaluated prospect (prospect A) and its similar (prospect B) that will be stored in the Knowledge Base.
The ultimate goal is to fairly rank the LoK of prospect A regarding prospect B and other similar
prospects. Overall, 67% of geoscientists considered that Prospect A has a higher LoK for trap geometry
than Prospect B. However, 13% claim that Prospect A has lower LoK than Prospect B, and the
remaining 20% of the geoscientists consider that the LoK between both prospects is similar. From the
achieved results, evaluation disagreements amongst people of the same company were clear.

To solve these inconsistencies, a given user can do its own LoK comparisons and store them in the
Knowledge Base. ML algorithms re-assesses the LoK of prospect A given the comparisons made by the
user with prospect B and other similar prospects present in the KB.

Figure 2 Prospect A LoK and PoS using initial (in red) and advisor (in green) methods for a given
risk factor. Using the LoK comparisons within a company’s portfolio, the level of knowledge for
this specific risk factor was increased, allowing peers to modify its PoS. Prospect A risk factor with
LoK advisor now fits within a cluster of similar prospects (in blue, prospects B, C, D and E) within
the prospectivity portfolio.

81st EAGE Conference & Exhibition 2019


3-6 June 2019, London, UK
In the end, the system gathers all comparisons stored in the KB made by every single user, using AI-
based algorithms to fairly rank the LoK for the evaluated prospect A, guiding the user to properly assess
the final PoS. In the case of Prospect A (Fig. 2), the LoK was altered from 0,5 to 0,7, allowing the user
to give a new PoS to the trap geometry factor.

Conclusions

This innovative methodology is the most consistent to incorporate a company’s prospect assessment,
avoiding economic or geologically biased evaluations, by comparing each prospect’s characteristics,
and building a continuously larger data and model-driven knowledge base. The assessment is based on
knowledge normalization, diminishing the associated subjectivity of the user’s analysis, which
frequently reflects the geoscientist’s experience and biased assumptions. A software prototype was
developed to implement this methodology, integrated within a multidisciplinary AI-based system,
which will allow geoscientists to gradually improve their daily work, with less time-consuming
assessments.

The results point to a successful technique to be systematically applied to geological risk assessment in
the oil and gas exploration workflow. Its efficiency depends on the continuous growth of the knowledge
base, which tends to be improved as more prospect evaluations are integrated, resulting in progressively
more accurate advices to end-users. At long-term, oil and gas companies that consistently apply this
machine learning-based method should benefit from a better prospect portfolio management, which
supports the exploratory strategies and reduces operational efforts. These are critical to consider,
especially given the ‘Big Data’ onset and the challenging expansion of oil and gas frontier exploration
into unknown areas.

Acknowledgements

The authors would like to thank Petrogal Brazil for funding this research project. We would like to
acknowledge Galp Exploration & Production and the Research & Development departments for
promoting the project and for their active participation on giving ideas and feedback to the team. Finally,
we thank IBM Research Brazil for the joint collaboration in this project, particularly for the
implementation of Machine Learning algorithms to geological risk assessment.

References

CCOP (Coordinating Committee for Offshore Prospecting in Asia), [2000] The CCOP guidelines for
risk assessment of petroleum prospects.
Costa, L., Zalán P. and Nobre, L., [2013] Estimativa da chance de Sucesso Exploratório: uma
abordagem em três passos consistente com a classificação de recursos petrolíferos. Petrobras
Geosciences. Bulletin, 21 (2), 313-324
Milkov, A., [2015] Risk tables for less biased and more consistent estimation of probability of
geological success (PoS) for segments with conventional oil and gas prospective resources. Earth-
Science Reviews 150, 453-476
Otis, R.M. and Schneidermann, N., [1997] A process for evaluating exploration prospects. AAPG
Bulletin, 81, 1087-1109
Polson, D. and Curtis, A., [2010] Dynamic of uncertainty in geological interpretation. Journal of the
Geological Society, 167, 5-10
Rose, P., [2001] Risk Analysis and management of Series, Petroleum Exploration Ventures. AAPG
Methods in Exploration, 12
Surovtsev, D, Levy, T and Sethi, M.U. [2019] Look at situations from all angles and you will become
more open: Advanced analytics approach to exploration portfolio allocation decisions. AAPG
Search and Discovery Article, 70378, DOI:10.1306/70378Surovtsev2019
White, D., [1993] Geologic Risking Guide for Prospects and Plays. AAPG Bulletin, 77 (12), 2048-2061

81st EAGE Conference & Exhibition 2019


3-6 June 2019, London, UK

View publication stats

You might also like