Wildfires significantly impact ecosystems, human health, infrastructure, and the climate, making accurate prediction of fire behavior and its effects critical. Traditional physics-based models simulate fire-atmosphere interactions in detail but are computationally expensive and limited in real-time applications. In addition, uncertainties in input parameters and simplified combustion representations can reduce their reliability in forecasting wildfire-driven emissions and plume dynamics. On the other hand, empirical and statistical models are computationally efficient but often lack the ability to capture the nonlinear and coupled processes that drive wildfire spread. This study presents a deep learning approach using a convolutional neural network (CNN) to predict wildfire dynamics under varying environmental conditions of wind, fuel, and atmospheric stability. The model is trained on a high resolution Weather Research and Forecasting (WRF) model coupled with the SFIRE (WRF-SFIRE) simulation dataset and predicts the temporal evolution of wildfire spread, represented through ground-level heat flux (GHF) fields as an indicator for fire intensity and progression. Model performance is evaluated using root mean square error ( R M S E ) of 14.3 kW/m 2 , mean absolute error ( M A E ) of 6.6 kW/m 2 , correlation coefficient ( R 2 ) of 84%, and the Structure–Amplitude–Location (SAL) method for spatial verification. Results show that the CNN effectively reproduces the spatial and temporal evolution of wildfire dynamics, closely aligning with reference simulations across diverse conditions. By accurately capturing fire spread and intensity patterns at much lower computational cost, the proposed approach demonstrates the potential of deep learning to complement existing fire modeling frameworks and to support faster, scalable forecasting of wildfire behavior.
Moradpour et al. (Mon,) studied this question.