Deep Learning Approach to Study Fluid Object Interactions

The flow physics around bluff bodies at varying Reynolds numbers has always attracted researchers due to the complex behavior of vortex-shedding phenomena. For the case of a circular cylinder, the vortex shedding starts beyond a diameter Reynolds number of 48. At low Reynolds number flow is assumed to be laminar, however, at higher Reynolds numbers turbulence starts to dominate the vortex shedding phenomena. On the whole, quantitative and qualitative analysis helps in understanding flow physics. However, simulating fluid behavior across all Reynolds numbers is time-consuming and requires substantial computational resources. Moreover, conducting experiments at lower Reynolds numbers is quite challenging. This article presents two novel computer vision-based methodologies for generating accurate and rapid simulations for any Reynolds number and obstacle shape. The study has achieved precise results by leveraging popular computer vision models such as Autoencoders achieving an average Frechet Video Distance (FVD) of 89.49 among them. The authors propose a new approach to create embeddings for videos, enabling feature extraction across frames. Using the ResNet architecture, a pipeline has been developed to generate a one-dimensional vector embedding, which can be evaluated using standard metrics like Peak Signal to Noise Ratio (PSNR) and Cosine Similarity. The proposed evaluation metric has been validated by calculating similarities between three videos (two similar and one dissimilar) from the web using cosine similarity.