H048-04
Development of a software platform for machine learning-accelerated decision support in reservoir engineering

Tuesday, 8 December 2020: 17:42
Virtual
Catherine Yonkofski, Pacific Northwest National Laboratory, Seattle, United States, Alex Hanna, Clemson University, Clemson, SC, United States and Casie L Davidson, Pacific Northwest National Lab, Richland, WA, United States
Recorded Presentation
Abstract:
One goal of the Science-Informed Machine Learning for Accelerating Real Time Decisions in Subsurface Applications (SMART) Initiative is to develop a software platform which leverages recent advancements in machine learning and cloud computing infrastructure to produce a real-time forecasting and decision support tool.

The platform consists of an application programming interface (API) managing the efficient simulation, post-processing, storage, interpretation and visualization of reservoir models. An object-oriented, modular software development approach is employed allowing extendibility and forward-compatibility with new instrumentation and optimization algorithms that become available. The API facilitates handshaking with various container and code repositories, as well as supercomputing resources such as national lab grid computers, Open Science Grid, and Amazon Web Services.

This allows for quick deployment of reservoir simulators to whatever mix of computing resources are appropriate to the urgency of the wellfield decision at hand as well as the budget and skillset of the user.

A GPU cluster is then used to train a suite of machine-learning approximations, using the full physics reservoir simulations as a training dataset. A stochastic optimization process is then conducted to fit these trained ML algorithms to field data and build an ensemble of models representing the

range of possible subsurface realizations indicated by the data.

This ensemble is then used to drive a reinforcement learning or agent-based approach, constructing a Bayesian network that represents the range of possible random events and operator decisions that could occur in the field. This facilitates improved decision-making and optimized wellfield operations.

The scientific workflow and data infrastructure driving this ML platform are presented, as well as an example use-case and a preliminary mockup for the user interface.

<< Previous Abstract | Next Abstract >>