Instance-Level Lithology Classification Using Self-Distilled Vision Transformer Representations

Summary Automated analysis of drill cuttings provides a continuous, cost-effective, and objective approach to subsurface lithology characterization, overcoming the limitations of manual inspection, which is slow, subjective, and often inconsistent. Traditional image-level classification methods yield only global lithology estimates, potentially obscuring mixed lithotypes within a single sample. To address this, we propose an instance-level framework combining high-resolution multi-light imaging, instance segmentation, and Vision Transformer (ViT)-based feature extraction. Using a self-distillation strategy, the model learns robust, shape- and texture-aware embeddings for each cutting, which are then used to train supervised classifiers. The method was applied to over 48,000 segmented cuttings across six lithologies, achieving more than 7% higher accuracy than previous approaches, with further improvement when incorporating UV-light features. Predictions are complemented by a confidence-based quality control mechanism and active learning, ensuring reliability and adaptability to new lithological contexts. This approach enables detailed, instance-level lithology description, facilitating real-time, high-resolution geological interpretation, and supporting traceable, scalable, and AI-driven workflows for exploration and reservoir evaluation.