Fast Reservoir Simulation and Multi?Well Placement Optimization with Low-Rank FNO

Abstract We develop a low rank Fourier Neural Operator (LR-FNO) surrogate to address the dual challenge of accelerating large-scale 3D reservoir simulation while enabling efficient multi well placement optimization. By applying CP-decomposed spectral convolution kernels, LR-FNO reduces parameter count by several orders of magnitude and alleviates overfitting, while maintaining high predictive fidelity. Compared with the original FNO, LR-FNO lowers GPU memory usage by roughly two-thirds, allowing the use of larger spectral kernels and improving model expressiveness. Trained on thousands of OPM simulations with randomly sampled well locations, the surrogate predicts full 3D pressure and saturation fields across all timesteps in under one second. LR-FNO also exhibits improved generalization in limited-data regimes and accurately reproduces cumulative oil production. When coupled with a Bayesian optimizer, the surrogate completes 1,500 iterations of the well placement optimization aimed at maximizing oil production within a few hours, compared to approximately ten days using OPM, providing a practical, efficient, and reliable decision support workflow for complex reservoir systems.