Hazardous Chemical Dispersion Analysis and Surrogate Model Optimization using Artificial Neural Network

Abstract

CFD simulations can be used to estimate the accident consequences by hazardous chemical releases. Both strength and weakness of this method lies on the complexity of the CFD simulation. In other words, CFD can predict chemical dispersion in the complex geometry like urban area, but it takes much computational resources. Comparison between actual dispersion and CFD simulation is done to show usefulness of the CFD simulation. Anhydrous hydrogen fluoride dispersion field test is simulated and it is within the dispersion model criteria. Actual hydrogen fluoride accidental release in 2012 is also simulated and its human and environmental damage prediction is similar to actual accident consequences. Case study of the toxic gas dispersion in the urban area varying source and meteorological condition is conducted, and numerical method is tested for the worst case scenario. Time variant CFL stepping method and multi-threading reduce the calculation time. Also ANN is able to classify the safe and danger based on the input condition. CFD simulation can be applied in the detector allocation problem in the process unit. Simulation inputs are determined by computer experiment procedure and other sets of data detection time is estimated with ANN. Then based on these simulated and estimated data MILP is solved and showed better result than using simulation data alone.