Open-set combustion-fingerprint classification with out-of-distribution alerting for environmental fire-hazard...

Fire effluent gases - carbon monoxide (CO), hydrogen cyanide (HCN), hydrogen chloride (HCl), and nitrogen oxides (NO_x) - are hazardous environmental contaminants released during structural and emerging fire incidents; rapid identification is critical for environmental hazard assessment and respiratory protection. The present framework does not measure these species directly; it classifies fire-material class and flags novel scenarios from bulk combustion-fingerprint features (CO, CO₂, smoke extinction, and heat release rate), and is therefore a combustion-fingerprint screening tool rather than a comprehensive fire-effluent speciation method. Existing machine-learning (ML) fire-gas classifiers operate under a closed-set assumption that fails on novel combustion sources such as wildland fires, electric-vehicle battery thermal runaway, and atypical industrial materials; these emerging scenarios motivate the open-set formulation, although the real-data validation here uses held-out wildland combustion only and no electric-vehicle data are employed. We present a framework integrating closed-set fire material classification with plug-in out-of-distribution (OOD) detection using multi-parameter combustion fingerprints from the publicly available NIST Fire Calorimetry Database (FCD). Seven calorimetry features (peak CO/CO₂, time-integrated CO/CO₂, CO/CO₂ ratio, peak smoke extinction K_smoke, peak heat release rate HRR_max) classified four structural fire material classes (Wood, Polymer, Textile, Composite; n=160, balanced n=40 per class). Random Forest achieved macro F₁=0.679±0.067 (95% CI [0.658, 0.700]; Cohen's d=6.76 vs. stratified baseline); LightGBM and XGBoost reached F₁=0.647 and 0.640. Under leave-one-class-out (LOCO) cross-validation the three plug-in detectors performed only marginally above chance (mean AUROC 0.560-0.569), reflecting chemical-space overlap among the structural classes; in a separate real-data test on n=20 held-out wildland scans - a chemically distinct novelty - LOF reached AUROC =0.812 and Isolation Forest 0.720. OOD detectability is thus strongly scenario-dependent. SHAP identified HRR_max, K_smoke,max, and ∫CO₂ as dominant features. The AEGL-2 (CO 150 ppm) and IDLH (CO 1200 ppm) breach rates under ISO 13571 are CO-based and yield a four-tier acute-exposure screening order (Wood 28% > Composite 22% > Polymer 20% > Textile 10%) rather than a comprehensive toxicity ranking. Beyond raw accuracy, the contribution is a reproducible, open-set formulation of fire-effluent hazard screening that maps what bulk combustion fingerprints can and cannot resolve, providing an interpretable, extensible screening layer and an honest capability baseline for environmental fire-hazard monitoring. Environmental Implication. Fire effluent gases - including carbon monoxide, hydrogen cyanide, hydrogen chloride, and nitrogen oxides - are hazardous environmental contaminants generated during structural and emerging fire incidents, contributing to over 350,000 fatalities annually through inhalation exposure. Reliable, streaming identification of combustion source class and detection of novel hazard scenarios (electric-vehicle thermal runaway, wildland-urban interface fires, atypical industrial materials) is essential for environmental hazard assessment, exposure mitigation, and respiratory protection decisions. This study presents a publicly available, lightweight, interpretable framework that translates standardized calorimetry data into screening-level environmental hazard signals, advancing data-driven monitoring of acute environmental contaminant exposure during fire incidents.