Predicting micro-bubble dynamics with semi-physics-informed deep learning

Utilizing physical information to improve the performance of conventional neural networks is becoming a promising research direction in scientific computing recently. For multiphase flows, it would require significant computational resources for network training due large gradients near interface between two fluids. Based on idea physics-informed (PINNs), modified deep learning framework BubbleNet proposed overcome this difficulty present study. The (DNN) with separate sub-nets adopted predict physics fields, semi-physics-informed part encoding continuity equation and pressure Poisson [Formula: see text] supervision time discretized normalizer normalize field data per step before training. Two bubbly i.e., single bubble flow multiple microchannel, are considered test algorithm. fluid dynamics software applied obtain dataset. traditional DNN BubbleNet(s) utilized train fields flows. Results indicate frameworks able successfully inclusion significantly improves NNs. introduction also has slightly positive effects prediction results. results suggest that constructing semi-PINNs by flexibly considering into will be helpful complex problems.