Application of Machine Learning Techniques to Formulate Simulation Inputs for Fire Models

Inspired by the advanment of repid development in “machine learning” techniques, we now proposed using genetic algorithm theory to optimise the pyrolysis kinetics input data of any polymer composites for our fire field models (i.e. computational fluid dynamics based). Typically, pyrolysis kinetics is extracted via thermal gravemetry (TGA), where the solid decomposition is studied in the format of Arrhenius chemical reactions. Herein, serveral key parameters including pre-exponential factor, activation energy, fractional variable of the reactions, exponential factor are required to be optimised for quality input data. Our pervious work has demonstrated with a simple hand-calculation, then followed by curve-fitting procedures, we can establish data with reasonable accuray prediction of small-scale flame for enigineered wood products. Now, we have improved the kinetics search with a fully automated iterative procedure utilising genetic search algorithms. This also greatly improve the accuracy of the prediction accuracy of our fire models.

A fire assessment model has been developed to provide a better understanding of the flame propagation, toxic gases and smoke generations of polymer composites. In this study, the effectiveness of the Chitosan/Graphene Oxide layer-by-layer fire retardant coating on flexible polyurethane foam was investigated experimentally and numerically via Cone Calorimetry. To generate quality pyrolysis kinetics to enhance the accuracy of the model, a systematic framework to extract TGA data is proposed involving the Kissinger–Akahira–Sunose method followed by Genetic Algorithm, with less than 5% of RMS error against experimental data. The proposed fire model is capable of predicting and visualising fire development and emitting gas volatiles.