Machine learning assisted convective wall heat transfer models for wall fire modeling

A significant source of error arising from wall fire modeling is the convective wall heat transfer model, especially when a thermal boundary layer along a wall is highly under-resolved, i.e., the size of the first grid cell from the wall is much greater than the viscous layer thickness. Traditional convective wall heat transfer models developed for forced or natural convection in a turbulent boundary layer tend to fail when used in a wall fire problem due to different challenges like the effect of blowing pyrolysis gas from a wall. In this work, machine learning with the random forest model is employed to construct a data-driven convective wall heat transfer model to provide an assessment of the feasibility and potential of using machine learning for enhancing wall fire predictions. To improve the predictions of convective heat flux to a wall in wall-modeled large-eddy simulations (LES) of wall fire, an amplification factor /i is introduced for the improvement of the calculation of the heat flux that accounts for the sub-filter scale effect. The factor /i compensates for the under-resolution of temperature gradient on the wall normal to its surface when the thermal boundary layer is under-resolved. A fine-resolution LES of afire case along a vertical wall is conducted to provide the training data for machine learning. The fine-resolution LES uses a grid size a few times larger than the viscous layer thickness but reasonably captures the inner thermal boundary layer. Different choices of input parameters for the machine learning models are examined. Several models for /i are constructed by using machine learning. Three different strategies for machine learning training are compared. Botha priori testing and a posteriori testing are performed in the vertical wall fire case to examine the performance of the developed machine learning models. The machine learning model is also tested in an intermediate-scale parallel-wall fire spreading case (0.6 m wide and 2.4 m high) which has not been seen by the trained model. Overall, the results show the promise of using machine learning approaches to enhance wall fire predictions.