Abstract
Risk is a combination of the frequency with which consequences occur, and their cost. We calculate it by estimating the range and frequencies for different initial conditions and calculating predictions that cover these to achieve frequency distributions of consequences, which we combine with loss and fatality models to account for cost. This requires calculation of enough consequence predictions to cover all realistic possibilities, which is often computationally impractical using computational fluid dynamics (CFD) simulations. We investigated using a surrogate model to cheaply calculate reliable consequence predictions for tens of thousands of scenarios. The surrogate is built using the inputs and outputs from CFD simulations. We investigated different methods of selecting scenarios for these simulations, and different methods of building a surrogate model. An approach based on a Gaussian process model and a recurrent neural network resulted in predictions of maximum equivalent cloud volume and a predicted probability density function for Q9 volume that are similar to the CFD predictions. The surrogate-predicted distribution is wide enough to capture less frequent events (that are often higher-cost) and can be combined with loss and fatality models to calculate a risk assessment. We used the FLACS-CFD model to demonstrate the method for an example case, focused on a dispersion scenario with a single leak and predictions for equivalent cloud volume. The method could be implemented for other scenario types and other predictions, and for other CFD models. The results are promising, suggesting that this type of surrogate may be useful for risk studies, where it is helpful to have many more predictions than it is computationally practical to simulate with CFD.
