Application of machine learning techniques to predict fire development in an ISO 9705 room

Abstract

Machine learning, a subset of artificial intelligence, shows potential for enhancing computational fire modelling compared to traditional methods such as computational fluid dynamics. This study explored using artificial neural networks to predict heptane fire development within a compartment, varying heat release rates from 100 to 3000 kW and ventilation areas from 0.16 to 4.8 m2. Artificial neural networks (ANNs) were trained using computational data from an ISO 9705 room. Network optimisation involved adjusting training-to-validation ratios and fine-tuning hidden layer neuron counts. Results indicate optimised ANNs achieved less than 7% error for heat release rate predictions and 1.5% for ventilation size predictions, with a notable computational cost reduction exceeding 104-fold. These findings suggest a promising future for integrating machine learning into fire engineering, significantly reducing analysis time therefore fostering safety improvements and innovation in the field.