Probabilistic real-time deep-water natural gas hydrate dispersion modeling by using a novel hybrid deep learning approach

Computational Fluid Dynamic (CFD) has been widely used for the gas release and dispersion modeling, which however could not support real-time emergency response planning due to its high computation overhead. Surrogate models offer a potential alternative to rigorous computational approaches, however, as the point-estimation alternatives, the existing neural network-based surrogate models are not able to quantify the uncertainty of the released gas spatial concentration. This study aims to fill a gap by proposing an advanced hybrid probabilistic Convolutional-Variational Autoencoder-Variational Bayesian neural network (Conv-VAE-VBnn). Experimental study based on a benchmark simulation dataset was conducted. The results demonstrated the additional uncertainty information estimated by the proposed model contributes to reducing the harmful effect of too 'confidence' of the point-estimation models. In addition, the proposed model exhibits competitive accuracy with R-2 = 0.94 compared and real-time capacity with inference time less than 1s. Latent size N-z = 2, noise sigma(z) = 0.1 and Monte Carlo sampling number m = 500 to ensure the model's real-time capacity, were also given. Overall, our proposed model could provide a reliable alternative for constructing a digital twin for emergency management during the exploration and exploitation of marine natural gas hydrate (NHG) in the near future. (C) 2020 Elsevier Ltd. All rights reserved.

Automated summary

This study introduces an advanced hybrid probabilistic Convolutional-Variational Autoencoder-Variational Bayesian neural network to address the inability of existing neural network surrogate models to quantify uncertainty. Experimental results demonstrate that the additional uncertainty information provided by the model helps to reduce the negative effects of overly confident point-estimation models. Additionally, the model shows competitive accuracy and real-time capacity.