Deep Learning achieves a 1000x boost compared to full-order Reservoir Simulations.

Reservoir simulation is dear to me since I was involved in scientific computing, building better linear/nonlinear solvers, and applying reinforcement learning for well control optimization at some stages of my life. The conventional simulation tools are costly to run for optimization, requiring thousands of simulation runs.

I have read the paper "Zhaoyang Larry Jin et al., 2020, Deep-learning-based surrogate model for reservoir simulation with time-varying well controls" that employs the embedded-to-control (E2C) procedure to speedup simulation by predicting key well parameters such as production and injection rates (and/or bottom-hole pressure (BHPs)) as well as global pressure and saturation distribution. Interestingly, they added an extra physics-based and data mismatch loss function to improve consistency with the governing flow equations.

They showed the result on a 2D two-phase setup with nine wells with varying controls. The Deep Learning-based Reduced Order Modeling outperformed the full-physics simulation by a factor of 1000's for GPU and 100's for CPU setup.

It would be interesting to see generalization on multiple geological distributions.
I'm curious about what else we can accomplish in the latent space.

GitHub repo: https://github.com/lonelysun1990/e2c_jpse
Dataset: https://drive.google.com/drive/folders/1P-R6uNkzw4lbVjgOIoe42okom08MtAN7
Paper: https://www.sciencedirect.com/science/article/abs/pii/S0920410520303533?via%3Dihub
Link for the LinkedIn post