Deep learning estimation of scalar source distance for different turbulent and molecular diffusion environments

Abstract

In order to adopt convolutional neural networks (CNN) for practical use in estimating the source of scalar dispersion in turbulent flows, such as gas leaks in industrial plants, the inference accuracy was verified using scalar concentration distributions in various turbulent environmental conditions. Training and test data were obtained through quasi direct numerical simulations on a flow system with a scalar-source-attached cylinder downstream of a turbulent grid, at two Schmidt numbers. An inference accuracy of at least 90% was confirmed under trained flow conditions, provided that the observation window was large enough to capture the integral scale in scalar fluctuations in turbulence. When the flow conditions (in terms of the turbulence intensity and the Schmidt number) differed between the training and testing phases, the accuracy was significantly reduced. However, the learner supervised with images under all different conditions showed high generalization performance. Singular value decomposition was used to discuss the features essential for training and prediction. In our CNN inference, we concluded that the source estimation was achieved by extracting from the integral-scale scalar patch distribution the features related to each of the turbulent and molecular diffusion progressions and their correlations. The results provide insights into the potential solutions for accurately predicting the sources of material dispersion in turbulent environments.