Key points are not available for this paper at this time.
Longwall mining faces persistent hazards from spontaneous combustion, methane explosion, and their coupled occurrence. Day-to-day hazard management practice relies on field monitoring, indicators, and empirical rules, while high-fidelity CFD is typically reserved for studies due to its high computational cost and specialist configuration. This study develops an operator-learning surrogate that maps three site parameters, ventilation flux, gas-emission rate, and drainage pressure, to continuous O₂ and CH₄ fields and to oxidation, explosive, and synergistic hazard masks. Trained on 2,000 CFD-generated cases and evaluated on 400-case held-out test sets, the surrogate attains an R² of 0.992 for O₂ and 0.998 for CH₄, and achieves precision of 98.6 and 86.5 percent for oxidation, 80.9 and 98.8 percent for explosive, and 83.8 and 99.0 percent for synergistic, reported for the positive and negative classes, respectively. The prediction of a single case completes within 0.015 seconds, which corresponds to an effective speedup of roughly 1.4×10⁶ relative to a six-hour CFD run. Increasing training-case cardinality yields rapid gains for the continuous fields from a few dozen to a few hundred cases, with diminishing returns beyond 200 cases, while positive-class precision for explosive and synergistic rises steadily and stabilises at 200 cases. Varying intra-case sampling density shows a marked deficit at 100–250 points per case and near-ceiling accuracy by 500–1,000. The work turns CFD from a specialist, time-consuming research workflow into a practical decision-support tool for rapid scenario evaluation. This enables instantaneous prediction of goaf concentration fields and hazard maps under prescribed operating conditions, supporting shift-time planning and rapid comparison of ventilation and drainage designs with immediate visualisation.
Hu et al. (Sat,) studied this question.