A Bayesian Regularized Artificial Neural Network for the estimation of the Ignition Probability in accidents in Oil & Gas plants

Within the Quantitative Risk Assessment (QRA) of Oil & Gas (O&G) plants, the estimation of the Ignition Probability (IP) following the release of flammable material in an accident (e.g., a Loss of Primary Containment (LOPC)) is commonly conducted by timeconsuming and computationally demanding Computational Fluid Dynamics (CFD) simulations, for only a limited number of operational configurations and accident scenarios. In this work, we propose an Artificial Neural Network (ANN) to overcome these limitations. Specifically, a Bayesian Regularized ANN (BRANN) is developed from a limited set of operational configurations and LOPC characteristics relative to a representative onshore O&G plant, then benchmarked and shown to outperform a traditional polynomial regression approach often adopted in O&G industry.