A Bayesian-optimized machine learning framework for predicting localized thermal resistance of firefighting gloves

This thesis develops and evaluates a Bayesian-optimized machine learning framework for predicting localized thermal resistance of firefighting gloves under varying material and environmental conditions. A dataset of 1,680 measurements obtained from 20 firefighting gloves was used to train and compare multiple linear regression, decision tree, generalized additive model, support vector regression, and deep learning models. The study examines predictive accuracy, stability, and physical consistency, and analyzes the influence of glove thickness, air-layer thickness, wind speed, glove section, glove type, and surface area on localized thermal resistance. Deep learning achieved the best predictive accuracy on held-out test data (R² = 0.924690, MAE = 0.009970), while support vector regression and generalized additive modeling showed comparable performance with different tradeoffs in flexibility and interpretability. Response-surface analysis showed that generalized additive models produced more physically consistent predictions in sparse-data settings, whereas deep learning captured more complex nonlinear interactions but was more sensitive to nonphysical behavior in underrepresented regions. The most influential predictors were air-layer thickness, glove thickness, wind speed, glove section, and surface area. Localized analysis identified the little finger as the most thermally vulnerable region, with thermal resistance decreasing substantially as wind speed increased. These results show that machine learning can support virtual prototyping and targeted thermal-performance optimization of firefighting gloves while reducing dependence on extensive physical testing.

Committee: Ali Jannesari (major professor), Guowen Song (major professor) and Regis Kopper