Fire smoke, rich in toxic ultrafine particles and polycyclic aromatic hydrocarbons (PAHs), poses significant health risks to first responders and vulnerable populations. In this study, a reproducible combustion–smoke simulation platform was developed to mechanistically quantify fire behavior, particle emissions, and PAH toxicity under controlled heat flux and oxygen conditions. Consistent combustion and smoke emissions were achieved by measuring heat release rate, particle mass, particle number concentration, and PAH concentration, with an overall average coefficient of variation below 15%. Systematic experiments with representative biomass (pine, oak) and plastics (PVC, polystyrene) demonstrate that fuel composition, heat flux, and oxygen availability jointly govern particle formation and PAH partitioning. Regardless of the combustion factors, ultrafine particles dominated the particle number concentration (55.5–86.2%). Plastic combustion generated 7 to 59 times particle mass, up to 260 times higher PAH emissions, and up to 58,500 times greater PAH toxic equivalent quotient (PAH-TEQ) than wood. Oxygen-deficient and smoldering regimes shifted emissions toward fine and ultrafine particles enriched in high-molecular-weight PAHs, revealing a coupled physical–chemical hazard not captured by bulk PM metrics alone. These results establish a quantitative framework linking combustion regime, particle size, and PAH toxicity, providing critical insight for exposure assessment, PPE design, and mitigation strategies in ventilation-limited and mixed-fuel fire scenarios.

Optimizing thermal comfort for gloves requires precise predictions of localized thermal resistance ( R _ct ), yet traditional experimental methods are costly and limited in capturing complex interactions. This study developed a computational framework employing Bayesian-optimized machine learning models, including multiple linear regression (MLR), decision tree (DT), generalized additive model (GAM), support vector regression (SVR), and deep learning (DL), to predict localized R _ct of firefighting gloves. Utilizing 1680 measurements from firefighting gloves, predictive accuracy and stability were evaluated. DL demonstrated the highest predictive accuracy ( R ² = 0.924690, mean absolute error [MAE] = 0.009970), closely followed by SVR ( R ² = 0.913098, MAE = 0.011209) and GAM ( R ² = 0.909959, MAE = 0.011472), whereas DT and MLR exhibited lower accuracy. Response surface analysis revealed that GAM provides consistent and interpretable predictions in sparse data contexts; DL captures detailed, complex interactions but requires careful management of potential overfitting; SVR represents an intermediate choice, balancing nonlinear accuracy with smoothness but still vulnerable to occasional nonphysical predictions in sparse data scenarios. Key predictors were air layer thickness, glove thickness, wind speed, glove section, and surface area. Thermal resistance increased consistently with air layer thickness, whereas higher wind speeds reduced insulation. Predictions highlighted the little finger as the most critical region, with R _ct decreasing from 0.145 to 0.112 K m ² /W as the wind speed increased from 0 to 2 m/s. Future studies should incorporate experimental validation under realistic firefighting scenarios, expand dataset diversity, and apply physically constrained models to enhance prediction robustness.