Hybrid Transferable Deep Reinforcement Learning and Transformer Architecture for Enhanced Lithology Identification From Well-Logging Data

Abstract

This research presents an innovative framework for lithology detection that combines domain adaptation, an Actor–Critic reinforcement learning (RL) architecture and Transformer-based sequence modelling to enhance log interpretation reliability in complex depositional environments. The study first reviews conventional petrophysical characterization methods using wireline measurements, noting their limitations in dealing with varied lithofacies distributions and non-stationary formation properties. Subsequently, it emphasizes the superior capabilities of neural networks, particularly the Transformer architecture, in analysing temporal measurement sequences. The multi-head attention mechanism in Transformers effectively models contextual relationships within depth-dependent logging signals, which is vital for stratigraphic interpretation. The proposed framework incorporates the Actor–Critic reinforcement paradigm, where the policy network (Actor) generates lithofacies predictions, and the value network (Critic) evaluates prediction quality. This dual-network setup promotes iterative policy refinement through feedback, enhancing classification consistency and computational efficiency. Moreover, recognizing the potential for domain shifts in logging campaigns, the framework includes parameter transfer mechanisms to facilitate knowledge distillation from source to target domains. This ability to adapt across projects significantly boosts model robustness and deployment feasibility in diverse reservoirs. Experimental validation on multiple well-log datasets shows that the combined Transformer architecture, RL, and transfer strategies outperform traditional machine learning and standalone deep learning models. Quantitative results reveal improvements in prediction accuracy, cross-well generalizability and domain adaptation efficiency in novel geological environments.