The Value of Integrating Machine Learning for Enhanced Understanding of Petroleum Systems
The evaluation of petroleum systems via basin modeling, mapping, geochemical analyses, etc., is now a “mature” field thanks primarily to 1) the codification of the petroleum system concept by Magoon and Dow in AAPG Memoir 60 and 2) the rise in sophisticated basin modeling software. Even so, the evaluation of petroleum systems, especially with basin modeling, remains an iterative process resulting in non‐unique solutions. In other words, there are always multiple models that adequately fit the data, leaving the user to decide which suite of models best reduces parameter uncertainty and ultimately, exploration risk. There is no substitute for traditional workflows that involve defining the four essential elements of the petroleum system—source rock, reservoir rock, seal rock, and overburden rock—and the two processes that link them (generation‐migration‐accumulation and trap formation). However, with the rise of big data and computational power, there is a resurgence in using machine learning techniques in the petroleum industry. These techniques are used and/or misused to supplement traditional geological workflows, automate them, or completely replace them. In this talk, we focus on the use of machine learning techniques for enhanced geologic and/or geophysical understanding of petroleum systems, one that moves beyond buzzwords and proprietary algorithms. We shape this talk in terms of the conference themes of “past, present, and future” as follows: Past: Over the past several years, machine learning has been deployed in exploration workflows in a variety of ways. Some of these are: identification of missed pay on petrophysical logs; geologic facies classification from thin sections, core, or seismic data; fault identification from seismic data; geobodies segmentation (e.g. salt bodies); data mining and handling of unstructured data (e.g., getting data from .pdf reports); and geologic cross‐section construction using automated well log correlation. The main drawback that we see in these techniques is several fold, chief among them being: 1) blindly applying algorithms to large data sets without understanding of the underlying techniques, which can lead to obtaining results with no or little geological meaning, and 2) the absence of integrating results with the petroleum system as a whole. In particular, machine learning has not yet been not integrated with basin modeling in any formal sense. Present: Today, machine learning workflows are being deployed as pre‐processors and post‐ processors to basin modeling. For example, in our research group, we utilize machine learning for lithology classification of data. The results are used to estimate prior probability distributions of corresponding parameters. Such operations aid in bounding the uncertainties in model building. We also decrease parameter space by applying machine learning techniques after model building. For example, we are utilizing modern computational power to run many scenarios of the basin model, incorporating the range of uncertainty in input parameters. We then use automated workflows to decrease human time spent on testing these scenarios: data science methods allow us to compare modeling outputs to measured data, thereby eliminating subjective visual calibration. Our methods also facilitate understanding the value of acquiring additional data, given the cost of acquisition, and the uncertainty in our predictions of volume and quality of petroleum accumulations or other prediction quantities. Future: We see opportunity for the application of machine learning techniques within basin modeling itself, especially with regard to the physics of petroleum migration. It is common to test models with a variety of migration physics, whether it be flowpath, Darcy, invasion percolation, or a combination of these. When combined with real‐world geology, machine learning combined with basin modeling may allow us to predict with higher confidence preferred flow pathways for petroleum migration in the subsurface. Another important research challenge that we aim to address is the prediction of various geo‐history parameters such as present‐day and paleo horizons and pore‐pressure models directly from seismic data by learning the direct relationship between data and target variables. Machine learning techniques can be critical in boosting the efficiency and accuracy of our current workflows by combining the disjointed workflows of geologic modeling and seismic inversion into an integrated step. In conclusion, the applicability of machine learning to the evaluation of petroleum systems is still not completely understood. In this talk, we use both synthetic analyses and real world case studies to elucidate how these sometimes “cryptic” techniques can reduce exploration uncertainty. In these examples, we share the reasoning for workflow design and the challenges encountered rather than just showing the final results. Further, we discuss the applicability of the workflows to other scenarios. Our main goal is to demystify machine learning by showing how it can be effectively used in an exploration context, if domain experts work together to integrate, and thereby improve, results.
AAPG Datapages/Search and Discovery Article #90349 © 2019 AAPG Hedberg Conference, The Evolution of Petroleum Systems Analysis: Changing of the Guard from Late Mature Experts to Peak Generating Staff, Houston, Texas, March 4-6, 2019