AI-Based Material Parameter Prediction from Cone Calorimeter Measurements

Punnoli Valayanad, Swathikrishna; Arnold, Lukas

doi:10.1088/1742-6596/2885/1/012013

Journal Article/Contribution to a conference proceedings

FZJ-2024-06723

AI-Based Material Parameter Prediction from Cone Calorimeter Measurements

Punnoli Valayanad, S. ; Arnold, L.FZJ*

2024
IOP Publ. Bristol

4th European Symposium on Fire Safety Science, Barcelona, Spain, 9 Oct 2024 - 11 Oct 2024 Journal of physics / Conference Series 2885(1), 012013 - (2024) [10.1088/1742-6596/2885/1/012013]

This record in other databases:

Please use a persistent id in citations: doi:10.1088/1742-6596/2885/1/012013 doi:10.34734/FZJ-2024-06723

Abstract: In pyrolysis simulation, methods such as inverse modelling are used to determine the material parameters from small-scale experiments. This often requires significant computational resources due to its iterative nature. This study investigates an artificial intelligence (AI)based alternative approach, that can give instantaneous predictions for material parameters once trained. A dataset based on Fire Dynamics Simulator (FDS) simulation of a cone calorimeter experiment is used for training these AI models. Di!erent AI models are trained to predict polymethyl methacrylate’s thermo-physical parameters using heat release rate (HRR) curves as input. AI models including Random Forest, 1D-convolution, and Recurrent Neural Networks showed the ability to predict the material parameters accurately with low mean squared error on the test dataset. These models were also able to recreate the HRR curves in FDS using their predictions, following the trend of the experimental HRR curve closely. Expanding the dataset to include materials with di!erent behaviours and modelling di!erent experiments could give these AI models broader applicability. However, FDS version dependence is a limitation for the AI models explored here because they were trained on a simulation-based dataset. Looking at the results, AI models in general can be used to predict material parameters required for pyrolysis modelling, potentially saving time and e!ort or, at the very least, used to complement the existing inverse modelling approaches.

Classification: