On the Generalization of PINNs outside the training domain and the Hyperparameters influencing it
This addresses the lack of practical studies on PINN generalization, which is relevant for architectures combining PINNs with domain decomposition or adaptive training strategies.
The paper empirically analyzes the generalization of Physics-Informed Neural Networks (PINNs) outside their training domain, investigating scenarios for consistent predictions and the influence of algorithmic setups, revealing insightful and sometimes counterintuitive results.
Physics-Informed Neural Networks (PINNs) are Neural Network architectures trained to emulate solutions of differential equations without the necessity of solution data. They are currently ubiquitous in the scientific literature due to their flexible and promising settings. However, very little of the available research provides practical studies that aim for a better quantitative understanding of such architecture and its functioning. In this paper, we perform an empirical analysis of the behavior of PINN predictions outside their training domain. The primary goal is to investigate the scenarios in which a PINN can provide consistent predictions outside the training area. Thereinafter, we assess whether the algorithmic setup of PINNs can influence their potential for generalization and showcase the respective effect on the prediction. The results obtained in this study returns insightful and at times counterintuitive perspectives which can be highly relevant for architectures which combines PINNs with domain decomposition and/or adaptive training strategies.