Session: 08-09-01 Neural Network for Waves & Cylinders, Symposium Summary

Paper Number: 105343

105343 - On the Performance of a Backward Compatible Physics Informed Neural Network (Bc-Pinn) for Prediction of a Flow Past a Cylinder 

Physics informed neural networks (PINNs) are a class of Physics informed machine learning (PIML) techniques where the underlying physics of the problem is enforced as a constraint in loss function of the model. PINNs are used to solve Partial differential equations (PDEs). The traditional PDE solving approaches like Finite Element Method(FEM), Finite Difference Method(FDM), Finite Volume Method(FVM) require precise knowledge of the underlying physics of the problem, boundary conditions and initial conditions. Generating mesh for these traditional approaches is still time consuming and can be considered as an art. PINNs solve the PDE by converting the problem to an optimization of a loss function, thereby bypassing the mesh generation workflow. PINNs have been used to solve Navier Stokes equations in the literature, however the usage of traditional training scheme has resulted in poor accuracy of predictions across the temporal domain. A training scheme, backward compatible PINN (BC-PINN), was proposed in literature, where the temporal domain is divided into smaller time segments and the model is trained sequentially, whilst satisfying the already obtained solutions for all previous time segments. The benchmark problem of 2D Unsteady laminar flow past cylinder, owing to its rich dynamics, was selected to test the robustness of the BC-PINN model. In this context, we propose a sparse data driven methodology where the model is additionally trained on low fidelity data from the CFD simulations. Extensive testing has shown that utilizing sparse data driven approach can significantly improve the accuracy as opposed to existing methodologies. Furthermore, transfer learning of parameters of a pre-trained model in different Reynolds numbers resulted in reduction of computation time by 3 hours. Additionally, it is shown that the proposed methodology can solve an ill-posed problem of missing boundary conditions with 6 times less error than the existing methodologies.

 

Presenting Author: Suresh Rajendran Indian Institute of Technology Madras

Presenting Author Biography: Suresh Rajendran is a faculty in the Dept. of Ocean Engineering , Indian Institute of Technology Madras, India. He works in the field of ship hydrodynamics and guidance, navigation, and control of marine vehicles.

Authors:

Vamsi Sai Krishna Malineni Indian Institute of Technology Madras
Suresh Rajendran Indian Institute of Technology Madras