Exploring the application of machine learning for facies classification in oil and gas exploration

Trustworthy machine learning for facies classification depends critically on leakage-free evaluation. This study proposes the Hybrid Explainable Transformer Framework (HETF), combining convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and Transformer branches, with SHapley Additive exPlanations (SHAP) and Monte Carlo Dropout uncertainty estimates. Using 830 samples from two wells in the SEG 2016 dataset, we correct leakage caused by random splitting of overlapping windows and apply depth-contiguous, blocked five-fold cross-validation. Across four repeated cross-validation runs, HETF achieved 58% mean accuracy (run range 52.5--60.8%) and 31% mean macro F1-score (run range 25.5--33.6%). Training accuracy of 85--98% substantially exceeded validation accuracy, indicating overfitting attributable to class scarcity. Under an identical leakage-free protocol, the Transformer-only model (76.4% accuracy; 49.2% macro F1), Random Forest (72.5%; 40.8%), and XGBoost (75.2%; 46.8%) outperformed full HETF (61.4%; 38.0%) in this single ablation comparison. Thus, branch fusion does not justify its complexity at this dataset scale. Preliminary KernelSHAP analysis ranked deep resistivity (ILD_log10), neutron porosity (PHIND), and gamma ray (GR) as the most influential features. The reproducible workflow, with released code, emphasizes methodological validity over inflated accuracy estimates.