Application of Deep Learning in Reservoir Simulation

Abstract

Summary Reservoir simulation plays a vital role as oil and gas companies rely on them in the development of new fields. Therefore, a reliable and fast reservoir simulation is a crucial instrument to explore more scenarios and optimize the production. In each simulation, the reservoir is divided into millions of cells, and rock and fluid attributes are assigned to these cells. Then, based on these attributes, flow equations are solved with time-consuming numerical methods. Given the recent progress in machine learning, the possibility of using deep learning in reservoir simulation has been investigated in this paper. In the new approach, fluid flow equations are solved using a deep learning-based simulator instead of time-consuming mathematical approaches. In this paper, we studied 1D Oil Reservoir and 2D Gas Reservoir. Data sets generated using the numerical models were used to create the developed simulators. We used two metrics to evaluate our models: Mean Absolute Percentage Error (MAPE) and correlation coefficient (R2). Given the low value of these matrics (MAPE 0.84 for 1D and MAPE < 0.84%, R2 ≈ 1 for 2D), the results confirmed that the deep learning approach is reasonably accurate and trustworthy when compared with mathematically derived models.