Inverse modelling of pyrolization kinetics with ensemble learning methods

Abstract

To simulate fire spread, especially the pyrolysis process – the thermal decomposition of a solid material – must be predicted. Yet, needed material dependent reaction kinetic parameters cannot be directly measured. Common methods infer them from small scale tests with inverse modelling approaches, which are computationally costly. Here, a novel machine learning based approach utilising extremely randomized trees (ERT) is presented and evaluated. It aims to derive these parameters almost instantly due to an inverse pre-trained surrogate model. The approach consists of an ERT classifier, a non-linear least squares optimiser and an ERT regressor. A thorough hyperparameter study was conducted. The model is evaluated with a synthetic thermogravimetric analysis (TGA) dataset. Calculated from an Arrhenius model, it contains data for more than 18⋅106 synthetic materials consisting of up to three components. The method is also applied on real experimental data, here polymethyl methacrylate (PMMA), gained from the Measurement and Computation of Fire Phenomena (MacFP) working group. Evaluation of the model demonstrated that it can instantly predict reaction kinetic parameters from TGA experiments for synthetic and real materials. Systematic analysis showed an overall R2 score of 0.77 for the complete model predictions. The code and datasets are published as open access.