2020/06/09 by Rohan Thavarajah, Xiang Zhai, Thavarajah, Rohan +5
Computer Science · Earth and Planetary Sciences · Physics and Astronomy · #Computational Physics (physics.comp-ph) #FOS: Computer and information sciences #FOS: Physical sciences #Generative Adversarial Networks and Image Synthesis #Machine Learning (cs.LG) #Model Reduction and Neural Networks #Seismic Imaging and Inversion Techniques
paper · pdf · doi:10.48550/arxiv.2006.05409
openalex publication_date 2020/06/09 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Is a deep learning model capable of understanding systems governed by certain\nfirst principle laws by only observing the system's output? Can deep learning\nlearn the underlying physics and honor the physics when making predictions? The\nanswers are both positive. In an effort to simulate two-dimensional subsurface\nfluid dynamics in porous media, we found that an accurate deep-learning-based\nproxy model can be taught efficiently by a computationally expensive\nfinite-volume-based simulator. We pose the problem as an image-to-image\nregression, running the simulator with different input parameters to furnish a\nsynthetic training dataset upon which we fit the deep learning models. Since\nthe data is spatiotemporal, we compare the performance of two alternative\ntreatments of time; a convolutional LSTM versus an autoencoder network that\ntreats time as a direct input. Adversarial methods are adopted to address the\nsharp spatial gradient in the fluid dynamic problems. Compared to traditional\nsimulation, the proposed deep learning approach enables much faster forward\ncomputation, which allows us to explore more scenarios with a much larger\nparameter space given the same time. It is shown that the improved forward\ncomputation efficiency is particularly valuable in solving inversion problems,\nwhere the physics model has unknown parameters to be determined by history\nmatching. By computing the pixel-level attention of the trained model, we\nquantify the sensitivity of the deep learning model to key physical parameters\nand hence demonstrate that the inversion problems can be solved with great\nacceleration. We assess the efficacy of the machine learning surrogate in terms\nof its training speed and accuracy. The network can be trained within minutes\nusing limited training data and achieve accuracy that scales desirably with the\namount of training data supplied.\n