Investigation of Reinforced machine Learning for History Matching

Summary History matching is critical for predicting reservoir behavior by aligning simulation data with historical production records. Traditional methods, like manual adjustments and gradient-based optimizations, are either slow or prone to local minima issues. Advanced techniques, including Ensemble Kalman Filter (EnKF) and geostatistical inversion, offer continuous updates but are resource-intensive or data-dependent. Recent developments integrate Mimetic Finite Difference (MFD) and Operator-Based Linearization (OBL) for improved scalability and precision. This study proposes reframing history matching as a Markov Decision Process (MDP) using reinforcement learning (RL) with Proximal Policy Optimization (PPO). In this setup, a deep neural network agent iteratively adjusts reservoir parameters across parallel simulations, significantly improving training speed and solution exploration. The agent’s reward function encourages convergence by focusing on reducing errors between simulated and actual data, with multiple parallel instances enhancing model accuracy and reducing computational time. Results indicate RL’s potential to achieve efficient and precise history matching, offering insights for optimizing reservoir management and field development.