Deep learning based quantitative property-consequence relationship (QPCR) models for toxic dispersion prediction

Abstract

It is crucial for emergency responders to makes a quick and accurate prediction of toxic chemical dispersions, which can lead to massive injuries and casualties. In this study, a toxic dispersion database is constructed by PHAST simulations, which consist of 30,022 toxic release scenarios of 19 chemicals. A quantitative consequence prediction model is then developed based on this database to efficiently and accurately predict dispersion downwind distances. Random forest, gradient boosting, and deep neural network algorithms are implemented and compared to find the best performing method for the model construction. The deep neural network is found to have the highest accuracy with the test set R2 higher than 0.994 and RMSE less than 0.1 for all key dispersion ranges. The developed toxic dispersion prediction models can be used to quickly generate instant toxic dispersion range estimations for any toxic chemicals at much lower computational costs.