Thermal Runaway of Chemical Reactors: An Experimental, Modeling, and Machine-Learning Investigation

Thermal runaway of chemical reactors remains a central challenge in chemical process safety because small imbalances between heat generation and removal can trigger rapid temperature escalation. In this contribution, we combine mechanistic modeling, targeted experiments, and data-driven tools to predict both the onset and intensity of chemical reactor runaway, demonstrated here for batch and fixed-bed systems and formulated to be extensible to other reactor types and reaction networks. Experimentally, we investigate selective methanol oxidation over Pt–Bi/activated carbon in a fixed-bed reactor and measure axial temperature profiles under varied feed temperature and oxygen concentration. A five-reaction pseudohomogeneous plug-flow model, coupled with sensitivity analysis of Tmax, identifies critical operating thresholds and clarifies how Bi promotion suppresses highly exothermic CO2 pathways, widening the safe operating window. To generalize prediction beyond single-reaction systems, we formulate data sets of dimensionless groups (γ, B, Da, St, ψ) for batch and flow reactors and introduce a kinetic weighting factor σn to map multireaction mechanisms onto an effective parameter space. Logistic Regression, Random Forest, and Support Vector models are trained and validated on literature data (acetic anhydride hydrolysis, batch) and our fixed-bed experiments (methanol oxidation), with Random Forest providing the highest accuracy and the lowest miss rate (false negatives) for classifying runaway onset. We propose a continuous, ML-informed criticality metric, η, which aggregates feature importance with proximity to classical critical values to quantify “how close” a system is to runaway. This unified framework links mechanistic insight with fast, reliable prediction, enabling safer operating envelopes, catalyst-informed mitigation strategies, and practical early warning diagnostics for industrial reactors.