Machine learning framework to predict instantaneous heat release rate of polymer nanocomposites in cone calorimetry

Abstract

A comprehensive understanding of fire behavior in a specific combustion scene is essential concerning the fire safety of materials. Cone Calorimetry (CCT), a worldwide benchmark lab-testing method in this field, reveals the heat release rate (HRR) over time, indicating the size and scale of the fire. On account of the Process-Structure-Property-Performance (PSPP) relationships, machine learning models have provided a shortcut to predict the HRR of a given composition of polymer composites before experimental testing. Here, we developed a machine learning framework using chained Random Forest Classifiers (RFC) to predict the HRR over time of polymer composites, which are loaded with Metal Hydroxide (MH) as flame retardant. Nine models were built with high R2 over 0.83 and low mean absolute errors (MAE). Comparison between predicted and real HRR curves of additional validation samples show good overlapping in 4 samples. Typical characteristics like time to ignition, peak HRR and total heat release rate obtained from the predicted and measured curves were basically identical.