Abstract
In this study, we propose a new data-free framework, Feature EnforcingPhysics Informed Neural Network (FE-PINN), to overcome the challenge of animbalanced loss function in vanilla PINNs. The imbalance is caused by thepresence of two terms in the loss function: the partial differential loss andthe boundary condition mean squared error. A standard solution is to use lossweighting, but it requires hyperparameter tuning. To address this challenge, weintroduce a process called smart initialization to force the neural network tolearn only the boundary conditions before the final training in a designedprocess. In this method, clustered domain points are used to train a neuralnetwork with designed weights, resulting in the creation of a neural networkcalled Foundation network. This results in a network with unique weights thatunderstand boundary conditions. Then, additional layers are used to improve theaccuracy. This solves the problem of an imbalanced loss function withoutfurther need for hyperparameter tuning. For 2D flow over a cylinder as abenchmark, smart initialization in FE-PINN is 574 times faster thanhyperparameter tuning in vanilla PINN. Even with the optimal loss weight value,FE-PINN outperforms vanilla PINN by speeding up the average training time by1.98. Also, the ability of the proposed approach is shown for an inverseproblem. To find the inlet velocity for a 2D flow over a cylinder, FE-PINN istwice faster than vanilla PINN with the knowledge of optimal weight loss valuefor vanilla PINN. Our results show that FE-PINN not only eliminates thetime-consuming process of loss weighting but also improves convergence speedcompared to vanilla PINN, even when the optimal weight value is used in itsloss function. In conclusion, this framework can be used as a fast and accuratetool for solving a wide range of Partial Differential Equations across variousfields.