Sorry, you need to enable JavaScript to visit this website.
Skip to content Skip to navigation

Machine Learning with Small Data for Basin Modeling

Project Leads: 
Tanvi Chedda

Optimizing decision making under uncertainty using Bayesian Networks

“.. the scientific cast of mind examines the world critically as if many alternative worlds might exist, as if other things might be here which are not. Then we are forced to ask why what we see is present and not something else.”-  Carl Sagan

Basin and Petroleum System Modeling (BPSM) involves model building (using understanding of basin stratigraphy, geochemistry, timing of events, and boundary conditions) and forward modeling (pressure, compaction, fluid flow, heat flow, kinetics calculations, etc.) to quantitatively predict petroleum generation, migration and accumulation, among other things. Because of the large number of input parameters, insufficient data, and imperfect data, among other things, there exists inherent uncertainty in modeling results. Although basin modeling results are typically communicated in the form of a single solution, they are better communicated in a probabilistic manner. In this work, we aim to understand the sensitivity of data and prediction variables to the variation in model parameters. We also propose a structured workflow for calibration of models to observed data.

In practice, the procedure of varying model parameters tends not to be systematic. The first model run is a best guess value of the parameters and the input is changed in subsequent simulations according to informed intuition to match the calibration data. Calibration is often done using visual inspection, which lacks quantitative comparison, and increases the possibility of subjectivity or approximation in interpretation. These ad-hoc methods are especially cumbersome when new data becomes available from additional testing, and the modeler must rerun the simulations to update the understanding of model parameters with new best calibration.

Hence, we elucidate the application of Bayesian networks in basin modeling to quantify and reduce uncertainty with measured data. Once the wide range of possible geologic scenarios and possible outcomes is explored, the question of drilling the exploration well or prospect arises. Associated with it is also the question of whether it is worth the cost of performing additional tests to obtain more data to reduce uncertainty. More information or data may reduce uncertainty but is not necessarily a true requisite for optimal decision making and we show how exploration costs can be minimized with more quantified decision analysis.

We test newer workflow for uncertainty reduction and decision analysis in the Jeanne d’Arc basin in the Grand Banks region offshore Newfoundland, Canada. Primary objectives of the study include characterizing uncertainties in the source rock quality and thermal history of the basin and developing a workflow to reduce uncertainty in models created at individual wells, extending the study to a 3D basin model with spatial uncertainties like lateral facies variation, migration pathways, and trapping mechanisms, and creating a structured decision-making process that is informed by a quantitative evaluation of risks and returns from exploration decisions.

This workflow, although shown with the example of a conventional offshore basin in the east coast of Canada, is fundamentally applicable to various basin types, locations, and decisions scenarios. The graphical formulation is an excellent communication tool that can incorporate expert knowledge and decision maker preferences. It also allows for modularity: risk components like source, reservoir, and trap can be evaluated separately as inputs from different teams as well as recombined for accumulation and level analyses at the management level in a large organization setting.