Interpretable Machine Learning for Thermal Conductivity Prediction of Silica Aerogel-Incorporated Cementitious...

Silica aerogel-incorporated cementitious composites possess low density and thermal conductivity. Their thermal conductivity is influenced by the interplay of mix composition, pore structure, mineral admixtures, and environmental testing conditions. In this study, a literature-derived database containing 208 data records was established for thermal conductivity prediction. Eight variables were utilized as inputs: aerogel content, water-to-cement ratio (W/C), sand content, foam content, silica fume content, fly ash content, testing temperature, and testing relative humidity. Thermal conductivity was designated as the output. The models developed for this study included XGBoost, random forest (RF), support vector regression (SVR), and their counterparts optimized using particle swarm optimization (PSO), which were subsequently compared. Among the optimized models, PSO-SVR showed the most balanced predictive performance, with test-set R², RMSE, and MAE values of 0.9306, 0.1005, and 0.0658, respectively. SHAP analysis identified sand content as the most important variable, followed by W/C and aerogel content. Mechanistically, silica aerogel reduces effective thermal conductivity by introducing low-conductivity phases, weakening solid heat-transfer networks, increasing heat-flow tortuosity, and accumulating interfacial thermal resistance. This study provides a data-driven and interpretable approach for thermal conductivity prediction and low-conductivity mix design of silica aerogel-incorporated cementitious composites.