Viability of Physics-Aware Deep-Learning based Proxy Reservoir Simulator for use in Well Control Optimization

Abstract

As oil production rises, and we are forced to undertake more challenging problems, the need for accurate and efficient reservoir simulations increases. Revitalizing old reservoirs and tackling more geologically complex reserves, require increasingly complex and computationally expensive reservoir models. It is necessary to develop new, more efficient alternatives to traditional reservoir simulation workflows to keep up with the demand in computational power. This study aims to explore the potential of neural network-based proxy reservoir simulators in the context of well control optimization, one of the most computationally demanding aspects of reservoir simulation.

The research objective is tested through an implementation of a particle swarm optimizer with an algorithm called E2CO – Embed to Control and Observe, a physics-aware neural network-based proxy reservoir simulator containing model input/output matching. The obtained results are analyzed and compared against the values obtained from a traditional reservoir simulator implemented on the same optimization framework.

The results show that implementation of E2CO with particle swarm optimization is a flexible and efficient alternative to traditional well control optimization workflows. Although not ideal, the algorithm provides acceptable accuracy, while delivering the results in a fraction of the time. Pairing this proxy model optimization framework with a traditional numerical simulator yields fast and sufficient results in a realistic workflow.

The testing methodology described in this research provides a vision for what AI can achieve when implemented in a reservoir simulation environment. Quantifying the potential improvements and highlighting areas that need additional exploration.