Deep Reinforcement Learning Framework to Solve Oil and Gas Depth-Matching Well-Logging Problem

Today the oil and gas industry is in the midst of a digital revolution of reducing cost and gaining efficiency by automating many human-intensive processes. Well log depth matching from multiple wells has been traditionally labor-intensive and continues to pose a challenge to efficiently automate and remove human intervention. Multiple attempts and techniques such as cross-correlation and Dynamic Time Warping produced mixed successes. In complex reservoirs, human intervention is required from expert geologists for manual adjustment of some intervals thus negating the advantage of a fully automated depth-matching system. Recent solutions to the problems include supervised learning algorithms to automate and classify the well logs through a fully connected neural network. As part of my thesis I present to you a different approach from a new field of study known as deep reinforcement learning. The new framework is based on a Deep Q-Network used to optimize the reward of the agent within the environment defined as the well log inputs. The network is built on a Deep Neural Network filled with 1-D convolutional layers used to extract features from the input logs. After extensive training, the agent can then take the best actions until it finds a match.