Uncertainty quantification of flammable gas dispersion numerical models driven by hybrid variational inference deep learning

Accurate modeling of flammable gas dispersion is essential for fire and explosion risk assessment. However, CFD models that rely on fixed hyperparameters preclude uncertainty quantification, leading to overconfidence prediction. This work proposed a hybrid deep learning framework with variational Bayesian inference to inversely solve distributions of numerical model parameters. The gas dispersion database under different Froude numbers Fr is developed, which contains repetitive experimental data and corresponding numerical simulation values. CNN-AM architecture is developed to capture nonlinear relationship between model parameters and concentration outputs. Using experimental data, ADVI is employed to derive posterior distributions of the optimal model parameters. The results indicate that the parameter-optimized model obviously improves prediction accuracy for 80 % scenarios, with overall error below 5 %. Furthermore, spatial distribution characteristics of plumes are characterized probabilistically. Near leakage nozzles, local concentration fluctuations peak when gravity and initial momentum jointly dominate plume dynamics at Fr = 74.38. In terms of plume morphology, variability in horizontal extent increases monotonically with Fr, while uncertainty in vertical drop attains a maximum at 0.060 when Fr = 55.79. These findings demonstrate the robustness of the proposed method for uncertainty quantification in gas distribution modelling, thereby enhancing risk evaluation in industries.