I started today (Dec 5) work on the evaluation and Proof of Concept to port the Matlab Reservoir Simulator MRST to PyTorch for GPU acceleration.

Part of the work requires the Eclipse datasets from SPE papers for testing. MRST. I have uploaded most important datasets to its own repository. See references below.

Since PyTorch has built-in functionality to work with Graphics Processing Units or GPUs, we expect demonstrating that PyTorch GPU-based tensors could significantly reduce compute time in reservoir simulation. That is the idea in a nutshell.

A handful of scientists have already ported some of their computational physics work to such ML frameworks but do not address specifically reservoir simulation.

1. Test the Partial Differential Equations (PDE) that constitute the core of the MRST solvers. Test the running times of the solvers using Matlab and Octave. Some code for performance testing available in latest book An Introduction to Reservoir Simulation Using MATLAB, Octave by Knut-Andreas Lie (see Appendix).
2. Replicate the functionality in Python using PyTorch for GPUs. Convert the Matlab code to PyTorch; measure the compute time of MRST solvers.

If GPU compute times are 10 to 100 faster in PyTorch, then proceed with converting more Matlab code to PyTorch tensor based calculations. Thanks Lukas Mosser for the brainstorming.

Already have some scripts running the basics of the Partial Differential Equations (PDE) required for proving that an experimental reservoir simulation tool can run using machine learning libraries such as PyTorch and TensorFlow

Scripts in Matlab/Octave, Python.

Started by testing Jacobians with TensorFlow. See references for code.

https://www.linkedin.com/posts/alfonsorreyes_f0nziereservoirdatasets-activity-6608176122876264448-gwST

https://www.linkedin.com/posts/alfonsorreyes_mrst-pde-reservoirsimulation-activity-6609474020737941504-8GaW

https://www.linkedin.com/posts/alfonsorreyes_mrst-matlab-pytorch-activity-6613066657558556672-wQYA

• Code and notebooks: https://github.com/f0nzie/mrst-pytorch
• Eclipse datasets: https://github.com/f0nzie/reservoir_datasets
• Book [An Introduction to Reservoir Simulation Using MATLAB, Octave] by Knut-Andreas Lie

#petroleumengineering #spe #oilandgas #reservoirengineering #computationalphysics