Development of Flammable Dispersion Quantitative Property–Consequence Relationship Models Using Extreme Gradient Boosting

Abstract

Uncontrolled release of flammable gases and liquids can lead to the formation of flammable vapor clouds. When their concentrations are above the lower flammable limit (LFL), or 1/2 LFL for conservative evaluation, fires and explosions can happen in the presence of an ignition source. The objective of this work is to develop highly efficient consequence models to precisely predict the downwind maximum distance, minimum distance, and maximum vapor cloud width within the flammable limit. In this work, the novel methodology named quantitative property–consequence relationship (QPCR) is proposed and constructed to precisely predict flammable dispersion consequences in a machine learning and data-driven manner. A flammable dispersion database consisting of 450 leak scenarios of 41 flammable chemicals was constructed using PHAST simulations. A state-of-art machine learning regression method, the extreme gradient boosting algorithm, was implemented to develop models. The coefficient of determination (R 2) and root-mean-square error (RMSE) were calculated for statistical assessment, and the developed QPCR models achieved satisfactory predictive capabilities. All developed models had high precision, with the overall RMSE of three models being 0.0811, 0.0741, and 0.0964, respectively. The developed QPCR models can be used to obtain instant flammable dispersion estimations for other flammable chemicals and mixtures at much lower computational costs.

Machine Learning and Deep Learning in Chemical Health and Safety: A Systematic Review of Techniques and Applications

“…Example of the gradient boosting tree algorithm. Reproduced with permission from ref . Copyright 2020 American Chemical Society.…”

“…Sun et al and Jiao et al. used PHAST simulation to construct a consequence database for fire radiation distance and flammable dispersion and used them to train the model to develop a quantitative property–consequence relationship (QPCR) model which can efficiently predict the corresponding consequence results. , …”

“…Sun et al and Jiao et al used PHAST simulation to construct a consequence database for fire radiation distance and flammable dispersion and used them to train the model to develop a quantitative property−consequence relationship (QPCR) model which can efficiently predict the corresponding consequence results. 132,133 The biggest limitation of ML-based consequence analysis is the lack of experimental data since it is extremely expensive and hazardous to conduct field dispersion experiments. The current studies try to tackle this problem by using widely validation dispersion simulation tools like CFD models of COMSOL, FLACS, and ANSYS or an integrated model like PHAST, and hence, the simulation will generate data for model training and validation.…”

Computational fluid dynamics model based artificial neural network prediction of flammable vapor clouds formed by liquid hydrogen releases

“…14 Jiao et al developed a quantitative property consequence relationship model and adopted different machine learning (random forest, gradient boosting) and deep learning (deep neural networks) algorithms combined with a datadriven manner to quickly generate the estimations of key toxic dispersion parameters, which achieved satisfactory predictive capabilities. 15,16 However, the model lacked consideration for the effects of wind speed and atmospheric turbulence on gas dispersion. Zhao et al developed a helium leak localization system using a gas sensor network and machine learning.…”

“…The artificial neural network (ANN) is extensively used in the analysis of data and prediction of results, which has significant speed advantages in result prediction 14 . Jiao et al developed a quantitative property consequence relationship model and adopted different machine learning (random forest, gradient boosting) and deep learning (deep neural networks) algorithms combined with a data‐driven manner to quickly generate the estimations of key toxic dispersion parameters, which achieved satisfactory predictive capabilities 15,16 . However, the model lacked consideration for the effects of wind speed and atmospheric turbulence on gas dispersion.…”