NVIDIA GTC: Deep Reinforcement Learning for Field Planning

 

 

 

 

Recently, GPU-accelerated reinforcement learning (RL) using deep convolution neural networks has gained significant traction. This presentation describes RL applied to the sequential decision-making process of well-placement optimization in the upstream energy industry.

 

High-fidelity multiphase porous media flow simulations of geological subsurface reservoirs are coupled with the placement of sources/sinks corresponding to wells; the goal of this process is to maximize net present dollar value (NPV) obtained through cumulative production.

 

Due to high-resolution reservoir grids and a combinatorial explosion in the number of possible spatial and temporal steps for placement of wells, the optimization routine is intractable and requires high-performance computing (HPC) strategies to solve realistic systems.

 

In this presentation, the advantages of RL over other conventional methods will be described, together with benchmark examples and results obtained on several reservoir models. The use of A100 GPUs to accelerate the learning and inference aspects of reinforcement learning will be shown to yield significant improvements in both performance and the best NPV solutions obtained by the optimization routine. Finally, the future of deep RL in the context of HPC and scientific simulation in energy will be outlined.