Development and Applications of Hazardous Gas Dispersion Surrogate Model Using Machine Learning

Abstract

The method of predicting hazardous gas dispersion using Computational Fluid Dynamics (CFD) is advantageous in that it can be utilized in various ways due to its high accuracy. However, it has structural complexity and requires a lot of computing resources, which hinders its utilization.<br/> First, TOXIM, which is a toxic gas dispersion simulator that can be easily used by anyone, has been developed solving the structural complexity of computational fluid dynamics. The stability of the simulator is confirmed by carrying out case study under various conditions using TOXIM. In addition, TOXIM's methodology of simulating gas leak accident that occurred in Gumi prove similarity between actual and simulation results.<br/> Next, a surrogate model is developed to reduce computation time of computational fluid dynamics and utilize it in real time. Compressed computational fluid dynamics calculation results are obtained by using various deep learning methods, and a surrogate model is created using neural network based functions. Among them, VAEDC-DNN shows the highest accuracy. It is confirmed that the computation time is reduced to several seconds in the surrogate model compared to several hours in the actual computational fluid dynamics. This demonstrates that computational fluid dynamics can be used in real time in a gas leak accident.<br/> Finally, the CFD model is applied to the gas leak detector placement problem. A surrogate model is used to generate various gas leak accident scenarios and the position of the detector is optimized by applying the Mixed Integer Linear Programming using the generated samples. Optimization results showed better performance when using surrogate model.<br/>