Rockburst Intensity Prediction Using TabPFN with Ensemble Consistency-Based Data Quality Screening

Abstract

Rockburst intensity prediction from heterogeneous engineering databases is constrained by small sample sizes and inconsistent grading practices across data sources. This study develops a framework that couples the pre-trained tabular foundation model TabPFN with ensemble consistency-based data quality screening and evaluates the screening effect within a nested design. Repeated predictions on the final 356-case database identified 69 cases (19.4%) with persistently high misclassification rates as potentially inconsistent. Sensitivity analysis showed that single-pass TabPFN accuracy varied by only 1.19 percentage points across 20 screening settings. Single-pass and nested screening produced different performance estimates, while nested screening did not consistently improve generalisation over the unscreened baseline. Under 30 repeated nested 60:40 splits, TabPFN achieved the highest mean accuracy of 63.03% ± 3.60% without task-specific hyperparameter tuning. Its advantage over LightGBM was significant (+6.95 percentage points, p = 0.031), while differences from RF and SVM were not. Random Forest proxy SHAP and direct TabPFN permutation importance both ranked Wet first, with σθ and rock type also ranking among the three most important feature groups in both analyses. Predictions for two Erlihe field cases were consistent with their engineering records. The framework therefore provides a practical approach for auditing heterogeneous rockburst databases and evaluating screening effects in small-sample rockburst classification.