Applying Deep Reinforcement Learning to Seismic Amplitude Data for Strategic Drilling in Geological Exploration
Strategic drilling decisions depend on the ability to interpret complex subsurface signals and act under uncertainty. In this highly technical presentation, Roderick Perez explores how deep reinforcement learning can be applied to seismic amplitude data to support well placement and improve hydrocarbon discovery in geological exploration.
The presentation introduces an integrated framework that combines geological modelling, agent-based hydrocarbon migration, synthetic seismic data and Deep Q-Network reinforcement learning. By training an AI agent on 100×100 RGB seismic frames, the study examines how machine learning can learn drilling strategies from seismic patterns, geological structures and reward-based outcomes.
A key part of the work is the simulation environment. Markov-chain-driven geological sections are used to generate different subsurface scenarios, while agent-based modelling simulates hydrocarbon movement through facies such as sandstone, silt and shale. These outputs are then translated into synthetic seismic amplitude data, giving the reinforcement learning agent a realistic environment in which to make drilling decisions.
The results show how the right training strategy can influence exploration performance. Agents using high discount factors and annealing exploration policies achieved stronger long-term rewards, better economic balance and more strategic drilling depth, particularly in structural traps where seismic signals are clearer.
This presentation offers a detailed look at the intersection of geoscience, seismic interpretation and AI-driven decision-making, showing how deep reinforcement learning could help advance strategic drilling workflows and future exploration planning.
