Point Source Estimation via Deep Learning for Passive Scalar Turbulent Diffusion

For the practical application of convolutional neural networks (CNNs) in predicting the source of substance diffusion in turbulent environments, such as gas leaks in chemical plants, we tested the accuracy of the predictions using image data of substance concentration distributions affected by turbulent motions. We found that if the image size is sufficient for capturing the characteristic scale of turbulence, an inference accuracy of at least 90% can be expected under the target flow conditions. However, if the training and testing images have different flow conditions, the accuracy is considerably reduced, highlighting the need for identical flow conditions during the training. This study also demonstrated that the generalization performance was improved by using a learner trained using images from all different flow conditions. Overall, the present results provide insights into the challenges and potential solutions for accurately predicting the source of substance diffusion in turbulent environments.