Neural network for real-time estimation of solid phase pyrolysis parameters

Abstract

The rate at which solid fuel is pyrolyzed is of great importance to determine combustion reaction rates at fires. As this transient process is highly sensitive on temperatures, the mass loss rate (MLR) tends to be derived by Arrhenius expression. Over the last decade years, novel optimization methodologies have emerged to obtain estimations of the thermokinetic parameters to simulate fire scenarios with the support of sophisticated models. While these works are based on different artificial intelligence technics, they share a common approach: when the user need to study a specific material, they launch thousands of simulations to find an optimized set of parameters which fix the best with a dependent parameter, like the mass loss rate curve. This study aims to define an inverse approach, training a neural network (NN), with thermokinetic parameters and thousands of mass loss rate simulations to create a single database, that once generated, can be applied to all subsequent analyses, reducing computational costs. Therefore, when one user needs to implement thermokinetic parameters in their fire engineering calculations, they must only supply the AI model with a MLR curve from bench-scale experiments, like gasification apparatus or Fire Propagation Apparatus (FPA). By processing this input, the corresponding reaction scheme and thermokinetic parameters are directly supplied. To validate the proposed AI model, it was directly applied to ten gasification apparatus experimental MLR curves obtained from the Fire Dynamics Simulator (FDS) Validation Guide. The results showcased an improvement in the uncertainty parameters compared to the Guide, underscoring the potential of utilizing neural networks to enhance the accuracy of FDS simulations for solid-phase pyrolysis. The proposed AI model can characterize the materials in few seconds, making it a tool of great interest for obtaining the input parameters of the FDS model from experimental tests.