Prediction of BLEVE blast loading using CFD and artificial neural network

Abstract

Boiling Liquid Expanding Vapour Explosions (BLEVEs) are extreme explosions driven by nonlinear physical processes associated with explosively expanded vapour and flashed liquid. Blast loading generated from BLEVEs may severely harm structures and people. Prediction of such strong explosions is not currently feasible using simple tools. Physics-based Computational Fluid Dynamics (CFD) methods are commonly utilized to predict the blast loading of BLEVE by going through many empirical formulas that map input variables to the target progressively. The calculation is often time-consuming, and it is therefore impractical to apply these methods to predict explosion loads from BLEVE in normal design analysis. Thinking of the composition of empirical relations in CFD models as a complex and nonlinear function, it is necessary to find an approximation of this function that can be efficiently calculated. The Artificial Neural Network (ANN) is a data-driven computational model that is capable of approximating any functions by learning from training data. Once properly trained, ANN can produce accurate predictions even for unseen inputs. This article presents the development of an ANN model to predict blast loading of BLEVEs in an open environment. A rigorous validation process is presented for the design of ANN structure, and the selected ANN is trained using validated simulation data from CFD models. Extensive evaluation of the network predictive performance is conducted, and it shows that the developed ANN can reproduce the result of CFD models effectively and efficiently, not only on simulation data but also on real experimental data. The prediction of ANN has a percentage error around 6 % and R2 value over 0.99 with the result of CFD simulated data. It speeds up the processing time from hours to seconds and only increases the error from 26.3 %–27.6 %, compared to the CFD simulations of real experimental data. Therefore, the developed ANN model can be potentially applied in the process engineering to generate a large number of reliable data for safety and risk assessment of BLEVEs in a more efficient way.