Robust Optimal Well Control using an Adaptive Multigrid Reinforcement Learning Framework

Abstract Reinforcement learning (RL) is a promising tool for solving robust optimal well control problems where the model parameters are highly uncertain and system partially observable in practice. However, RL of policies often relies on performing large number simulations. This could easily become computationally intractable cases with intensive To address this bottleneck, an adaptive multigrid framework introduced which inspired by principles geometric methods used iterative numerical algorithms. initially learned using efficient low-fidelity simulations coarse grid discretization underlying partial differential equations (PDEs). Subsequently, simulation fidelity increased manner towards highest that corresponds to finest domain. The proposed demonstrated state-of-the-art, model-free policy-based algorithm, namely proximal policy optimization algorithm. Results shown two case studies problems, from SPE-10 2 benchmark studies. Prominent gains computational efficiency observed framework, saving around 60-70% cost its single fine-grid counterpart.