Physics-Constrained Neural Hawkes GNN for Reactor Runaway Precursor Recognition

Source abstract

Thermal runaway in stirred and semi-batch reactors is often preceded by weakly coupled alarm bursts, heat-removal imbalance, and delayed pressure response, yet these signatures are difficult to distinguish from benign operating transients when measurements are irregular. This paper proposes a physics-constrained Neural Hawkes graph neural network in which a process-topology encoder conditions marked event intensities and a normalized heat-balance residual jointly regularizes training and gates online risk. The principal novelty is this coupled mechanism: event clustering can increase risk only when graph-propagated process evidence and thermal disequilibrium support the escalation. Evaluation uses a controlled synthetic and historian-like benchmark containing 12,480 operating windows, 38 measured variables, and 412 incipient-runaway cases; it is intended for method comparison rather than plant certification. The model achieved an F1 score of 0.918, an AUROC of 0.973, and a median warning lead time of 18.6 min. Relative to LSTM, temporal convolution, Hawkes-only, and graph-only baselines, it reduced missed precursors by 31.4% and nuisance alarms by 22.7% at a fixed recall of 0.92. Ablation tests indicate that event timing supports aggregate recognition, graph propagation captures cross- unit coupling, and the physical residual is particularly important under cooling degradation. The framework therefore provides an interpretable supervisory warning layer for subsequent pilot- and plant-scale validation.