Interpretable Short-Horizon Prediction of High-Energy Microseismic Events for Rockburst Risk Screening Using TabPFN-SHAP

Routine microseismic monitoring records event time, location and energy, yet converting these data into interpretable guidance for mine inspection remains difficult. We combine Tabular Prior-Data Fitted Networks (TabPFN) with permutation SHapley Additive exPlanations (SHAP) to estimate whether a high-energy microseismic event will occur in the same grid within 24 h, thereby supporting rockburst risk screening without directly predicting rockburst occurrence or damage. We aggregated 9116 events into 3298 grid-time samples described by 28 energy, temporal-activity and spatial-distribution variables. Samples were divided chronologically using 48 h buffers, and thresholds were selected only from validation data. On the held-out test period, TabPFN achieved the highest observed areas under the receiver operating characteristic and precision-recall curves (ROC-AUC 0.8064; PR-AUC 0.5330) and the lowest Brier score (0.1190) among six classifiers. Four controlled semi-synthetic domain-shift tests produced TabPFN ROC-AUC values of 0.6992–0.7752. SHAP identified preceding-window cumulative energy and event count as the leading contributors, with stable rankings under expanded sampling. The framework therefore links grid-level screening to monitoring quantities that engineers can inspect. These semi-synthetic results are stress tests rather than independent-mine validation, and no transferable warning threshold is claimed.