Coupling Neural Netowrok technique with CFD simulations for predicting 2-D atmospheric dispersion analysing wind and composition effects

Computational Fluid Dynamics (CFD) defines a noteworthy methodology to carry out the numerical modeling of generic flow field problems, encompassing correlated phenomena like mass and heat transfer. Especially in gas dispersion context, such a tool presents a remarkable applicability, allowing one to perform comprehensive analysis in areas as gas leakage and explosion risky and consequence. An expressively widespread branch of Artificial Intelligence paradigm consists of Machine Learning, a techinque of which is given by the Neural Network framework, whose combination with CFD experiments a considerable growing recently in the literature and industry. The present work seeks to demonstrate the employment of such a coupled method in atmospheric dispersion representative problems. To do so, a limited set of CFD simulation results are used for developing neural networks by supervised trainings. The outcome are predictors for flow field interpolation purposes, the potential uses of which include digital twin designing, gas leakage detection optimization procedures etc. Particularly, two ways have been structured regarding the use of neural networks in replacement of CFD simulations: a global and a local approaches. The first one receives directly one or more system conditions boundary information (e.g. entrance temperature, flow rate, composition or other) and is expected to return the entire property distribution over the computatinal domain. Although leading to denser archtetures and less physical-based, this strategy involves single applications of the obtained model for each interpolated condition. On the other hand, the local approach treats the network as a transition rule in the scope of Cellular Automata (CA), making it able to locally learn the dynamic behaviour of the addressed physics. In this via, each nodal point is looked at in different time-steps; property values in the previous instant at the particular node and at its neighbours are taken, from which the respectives fluxes are computed and serve as input to the network providing then the current node value. This method gives rise to simpler neural network archtetures (fewer input and output) with closer computing relatively to the CFD calculation, but, as a downside, requires several runnnings of the model for each scenario in a transient fashion. Evaluations have been done by predicting species concentration distribution in atmospheric dispersion cases varying wind condition (velocities direction and magnitude), entrance composition and obstacle shape and placement.