Evolution of burning rates of tank fires under crosswind effect: experimental investigation and machine learning prediction

• A critical ullage height (0.8D) is identified for transition of combustion modes. • Dominant heat feedback shits from convection to radiation with increasing crosswind. • SEM model achieves superior prediction accuracy over SVR model. • SHAP analysis reveals coupled effects of crosswind and ullage height on burning rate. Storage tank fires pose critical risks to energy infrastructure safety and environmental sustainability, driven by complex nonlinear interactions among crosswind, heat transfer, and combustion dynamics. This study investigates the burning behavior in vertical storage tanks through integrated experimental and machine learning approaches, focusing on the nonlinear coupling mechanisms of crosswind velocity (0–2.5 m/s), dimensionless ullage height ( h/D = 0–1.2), and heat transfer mechanisms. Experimental analyses were conducted to assess the effects of crosswind velocity and ullage height on flame morphology, wall temperature distribution, and heat flux contributions (conductive, convective, and radiative). Key findings revealed two distinct flame regimes during stable combustion: bottom-contacting (flame contacts fuel surface) and non-bottom-contacting, with a critical transition occurring at an ullage height of 0.8D. The bottom-contacting mode exhibited higher burning rates due to intensified convective heat transfer, while non-bottom-contacting mode showed reduced wall temperature due to the presence of downwind airflow channels near the tank rim. Two machine learning (ML) models were used to predict the burning rate, namely a support vector regression (SVR) model and a stacked ensemble model (SEM), which resolve nonlinear scaling laws between pool diameter ( D ), crosswind velocity ( u ), and ullage height ( h ). It was found that SEM achieved superior predictive performance compared to the SVR model. Feature importance analysis identified D as the most influential parameter on the burning rate, followed by u and h . The findings can enhance predictive capability for thermal hazard assessment in energy storage systems and provide a basis for engineering solutions that improve thermal management, support emergency planning, and strengthen the resilience and sustainability of energy infrastructure.