Coupling a neural network technique with CFD simulations for predicting 2-D atmospheric dispersion analyzing wind and composition effects

The Computational Fluid Dynamics (CFD) tool has a remarkable applicability for the prediction of gas dispersion flows by numerically solving the proper governing equations in realistic scenarios. Depending on the problem complexity, undesirably high computational costs can be incurred, which has encouraged the combined use of Machine Learning (ML) seeking to attenuate the CFD simulations requirement for multiple scenario studies. The present work aims at demonstrating the employment the coupling between CFD and the Artificial Neural Network (ANN) algorithm for representative problems in atmospheric dispersion in a preliminary assessment. A limited set of CFD simulation results was used for training neural networks, whose output is given by flow field interpolators, the potential uses of which include digital twin designing and optimization procedures. One possible strategy is the local approach, which treats the network as a transition rule in the scope of Cellular Automata (CA) modeling, allowing it to learn the dynamic behavior of the addressed physics locally. This method gives rise to simpler neural network architectures with closer computing relatively to the CFD calculation. Assessments have been done by predicting, initially, a scalar field time evolution governed by a 1-D advection-diffusion transport equation to verify the method implementation. Subsequently, species concentration distributions were sought in atmospheric dispersion cases from CFD simulations datasets, comprising four case studies followed in the performed analysis, all considering a bidimensional flow domain and a scenario involving methane leaks. The first one indicated an accurate reproduction of subsequent time steps concentration field referring to the displacement of a methane cloud. The second and third cases concerned a plume formation, in transient and steady-state regimes, respectively; their main outcome was the evidence of the CA-ANN methodology's flexibility to address time-dependent and permanent flow simulations interpolation. The last CFD-based case study comprised an additional complexity feature of gas dispersion problems: the wind influence. By redesigning the investigated data-driven approach in terms of ANN's features and labels choice, promising results followed from the analysis with respect to the simultaneous capturing of two global simulation parameters (wind and leakage speeds boundary conditions) in the species concentration field interpolation.

Automated summary

The present work demonstrates the application of Computational Fluid Dynamics (CFD) and Artificial Neural Network (ANN) algorithm in atmospheric dispersion problems. The combination of CFD and ANN allows for digital twin design and optimization. The local approach reduces the computational cost by treating the network as a transition rule. Several case studies were conducted to validate the effectiveness of the method, including methane cloud displacement, plume formation, and wind influence.