Testing for knowledge: maximising information obtained from fire tests by using machine learning techniques

A machine learning (ML) algorithm was applied to predict the onset of flashover in 1:5 scale Room Corner Test experiments with sandwich panels. Towards this end, a penalized logistic regression model was chosen to detect the relevant variables and consequently provided a tool that can be used to make predictions of unseen samples. The method indicates that a deeper understanding of the contributing factors leading to flashover can be achieved. Furthermore, it allows a more nuanced ranking than currently offered by the commonly used classification methods for reaction to fire tests. The proposed methodology shows a substantial value in terms of guidance for future large and intermediate scale testing. In particular, it is foreseen that the method will be extremely useful for assessing and understanding the behaviour of innovative materials and design solutions.