Well Placement Optimization for Avoiding Caves Using GANs and POMDPs

Abstract

Abstract Exploration for subsurface resources, such as groundwater and hydrocarbons, involves a high degree of uncertainty. This is because the spatial distribution of targeted resources and the geology of their overburden is largely unknown. In some areas, eogenetic caves known as banana holes overlay targeted resources. Because of pressure loss, drilling into these caves causes operational hazards that lead to project delays, well abandonment, and even loss of life. The goal of this paper is to formulate well placement as a sequential decision-making problem and solve for an artificially intelligent agent that avoids drilling into these types of caves. We formulate the decision problem as a partially observable Markov decision process (POMDP). To model the spatial uncertainty of caves, a Generative Adversarial Network is used. The generative model is trained based on a Light Detection and Ranging survey from San Salvador Island, Bahamas. To solve for a policy that recommends drilling actions, the Fast Informed Bound algorithm is explored. The proposed sequential approach for well placement is shown to recommend a sequence of actions that avoids drilling into caves and demonstrates the potential of POMDPs in this problem.