Interpretable SHAP Model Combining Meta-learning and Vision Transformer for Lithology Classification Using Limited and Unbalanced Drilling Data in Well Logging

Abstract

Recent advances in geological exploration and oil and gas development have highlighted the critical need for accurate classification and prediction of subterranean lithologies. To address this, we introduce the Meta-Vision Transformer (Meta-ViT) method, a novel approach. This technique synergistically combines the adaptability of meta-learning with the analytical prowess of ViT. Meta-learning excels in identifying nuanced similarities across tasks, significantly enhancing learning efficiency. Simultaneously, the ViT leverages these meta-learning insights to navigate the complex landscape of geological exploration, improving lithology identification accuracy. The Meta-ViT model employs a support set to extract crucial insights through meta-learning, and a query set to test the generalizability of these insights. This dual-framework setup enables the ViT to detect various underground rock types with unprecedented precision. Additionally, by simulating diverse tasks and data scenarios, meta-learning broadens the model's applicational scope. Integrating the SHAP (SHapley Additive exPlanations) model, rooted in Shapley values from cooperative game theory, greatly enhances the interpretability of rock type classifications. We also utilized the ADASYN (Adaptive Synthetic Sampling) technique to optimize sample representation, generating new samples based on existing densities to ensure uniform distribution. Our extensive testing across various training and testing set ratios showed that the Meta-ViT model outperforms dramatically traditional machine learning models. This approach not only refines learning processes but it also adeptly addresses the inherent challenges of geological data analysis.