Physics‐Guided Contrastive Learning for Facies Classification in Tight Gas Reservoirs Using Limited Core Data

ABSTRACT Facies classification in tight gas reservoirs faces a critical challenge: supervised machine learning requires extensive labelled core data that is economically prohibitive, while unsupervised clustering ignores geological physics and produces petrophysically implausible results. This study presents a physics‐guided contrastive learning (PGCL) framework that bridges this gap by integrating unsupervised representation learning with domain‐knowledge constraints. Applied to 3284 wireline log samples from five wells penetrating the coal‐bearing Shanxi Formation in the Eastern Ordos Basin, China, the framework employed SimCLR contrastive learning to construct 64‐dimensional embeddings from nine log features. Hierarchical refinement using binary density‐neutron cutoffs for coal separation and Ward's clustering on core permeability‐porosity space transformed initial unsupervised clusters into seven geologically coherent facies spanning three orders of magnitude in permeability. Random Forest classifiers trained on PGCL‐generated pseudo‐labels achieved 87.0% test accuracy, exceeding traditional supervised methods by 5–6 percentage points and unconstrained clustering by 15–16 percentage points despite requiring only 346 core measurements for physics‐guided refinement. Integration with reservoir quality metrics showed that 98% of predicted Sandstone intervals and 78% of Silty Sandstone intervals exceeded commercial production thresholds, enabling probabilistic well placement with quantified risk assessment. The results demonstrate that semi‐supervised learning enhanced with geological physics can reduce core dependence and provide reliable facies prediction for exploration‐stage reservoir evaluation.