Applying Deep Reinforcement Learning for Strategic Drilling in Subsurface Exploration

Optimizing where to drill in a synthetic geological simulation to maximize hydrocarbon discovery, with no real seismic data to lean on. In this Expert Spotlight Series session, ThinkOnward Community Member, Roderick Perez, PhD, shows how he built a full pipeline to solve it: Markov-chain-driven geological sections to generate realistic subsurface structure, agent-based modeling to simulate hydrocarbon migration, Gassmann fluid substitution to produce synthetic seismic amplitudes, and a Deep Q-Network trained directly on the resulting seismic images to learn drilling strategy. The results showed that the agent's mindset matters as much as its inputs. Agents tuned for long-term thinking and a gradually narrowing search strategy consistently found stronger, more balanced drilling outcomes than agents that didn't. Roderick is candid about where the approach holds up and where it doesn't: it performs well on clear structural traps and struggles when the seismic signal gets noisy, a limitation worth understanding when considering how approaches like this might be evaluated in real-world settings. A technical, results-driven session for anyone curious where reinforcement learning is being explored in subsurface exploration today. The Expert Spotlight Series is ThinkOnward's recurring webinar program built around the expertise inside our community. Each session features a ThinkOnward Community Member presenting original work and hard-won knowledge across subsurface science, engineering, and AI.

Wednesday, June 24 at 11:00 AM CDT