Predicting thermal pyrolysis of asphalt via comparative artificial intelligence methodologies with thermogravimetric experimental validation

Asphalt pyrolysis, a critical thermal process governing fire initiation and evolution in civil infrastructure such as tunnels, involves complex multi-stage and strongly nonlinear reactions. This study has established a substantial pyrolysis database (2488 publication and 6781 supplementary experimental data points) to enable data-driven fire consequence modeling. The thermogravimetry (TG) experiments first identify a sharp peak derivative thermogravimetry (DTG) of 10%/min at 420–480 °C, quantifying its high combustibility and associated fire intensity. A general artificial intelligence (AI) framework was subsequently developed, where four models (SVR, RF, XGBoost, ANN) were compared for predicting TG/DTG of asphalt. The RF model demonstrated superior accuracy (within ±20% error) attributable to its robust tree-based algorithm. It also rationally depicts the variation trends of DTG with heating strengthened, further confirming its robustness. Furthermore, both RF and XGBoost successfully predicted an additional DTG peak from aluminum hydroxide flame retardant, demonstrating their capability to reflect fire suppression mechanisms and multi-stage chemistry, with RF's peak value error at merely 15.2%. Herein, this data-driven AI framework demonstrates significant potential for application in modeling thermal decomposition behavior and fire consequences, thereby contributing to enhanced fire pre-warning capabilities and safety protection in thermal engineering systems associated with civil infrastructure.