Firebrand transport is a primary cause of fire spread in outdoor fires, but its prediction is challenging. This study develops a highly accurate and low-cost prediction model for firebrand transport using machine learning based on experimental data. Experimental dataset used is firebrand data from wood materials of single size and produced from a firebrand generator. First, five machine learning models were built and compared. The results showed that a Neural Network (NN) provided the best performance, accurately reproducing the landing distribution of firebrands. Next, SHAP analysis were used and it was found that physical indices, such as the Tachikawa number, were critical for accurate predictions. Based on this finding, a Physics-Constrained Neural Network (PCNN), which incorporates physical laws into the learning process, was developed. The PCNN model was tested against unseen data which is firebrand data generated from combustion experiments with roof assemblies. The PCNN, especially when constrained by the Tachikawa number, demonstrated more robust and accurate predictions on unseen, more complicated data compared to the standard NN. The ML technique developed in this study is able to predict firebrand transport with different shape. This work shows that machine learning, particularly PCNN, is a powerful tool for predicting complex firebrand transport, potentially contributing to future fire risk assessment. • NN proven to be superior for firebrand transport prediction to other four ML model • SHAP analysis identified Ta , Re , and velocity ratio as key physical features • PCNN showed improved generalization performance on unseen, complicated data • Ta as physical constraint was the most effective approach for improving PCNN
Kameyama et al. (Sun,) studied this question.