Machine learning accelerated radiative transfer modeling in CFD fire simulations

Radiative heat transfer is the dominant heat transfer mode in many fire scenarios; yet, its solution procedure is often a major performance bottleneck in Computational Fluid Dynamics (CFD) simulations. Conventional solvers for the governing radiative transfer equation (RTE) are computationally expensive and can suffer from numerical artifacts, such as “ray effect”. This paper presents a machine learning-accelerated RTE solver based on Deep Operator Networks (DeepONet) to address these limitations. The DeepONet is trained as a neural surrogate to learn the underlying solution operator of the RTE, directly mapping input functions, such as absorption coefficient and temperature, to the output radiative intensity solution. This pre-trained DeepONet model is then integrated into the open-source CFD code FireFOAM, replacing the baseline finite-volume discrete-ordinate method (fvDOM) solver. The performance of this hybrid ML-CFD framework is evaluated in transient simulations of canonical fire scenarios, including two-dimensional (2D) and three-dimensional (3D) pool fires and a 3D wall fire. Results demonstrate that the DeepONet solver reproduces the radiation solutions with significant computational speed-up. Moreover, when using the DeepONet solver at high resolution, the ray effect is mitigated.