Physics Informed Neural Networks: Reducing Data Size Requirements via Hybrid Learning
Author:
- 1 December 2022
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 178-179
Abstract
Training physics informed neural networks (PINN) in a supervised manner requires enormous amounts of data. This is a strain on hard disk space and RAM. To decrease the amount of data required, and to increase accuracy, we can train a PINN using both supervised and unsupervised learning. In this method, we use a smaller dataset to nudge the PINN in the correct direction; then, the unsupervised training takes control to smooth out the final results and increase accuracy. While some research has been done using this hybrid approach, there is a gap in literature regarding the size of the requisite training set and the correct method to weigh the supervised and unsupervised losses. In this research, we used a hybrid approach to model the Allen-Cahn equation. We found that even applying a tiny weight to the unsupervised loss of 0.00255 was sufficient to dramatically increase the accuracy of our PINN.Keywords
This publication has 3 references indexed in Scilit:
- Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s NextJournal of Scientific Computing, 2022
- Physics-informed neural networks approach for 1D and 2D Gray-Scott systemsAdvanced Modeling and Simulation in Engineering Sciences, 2022
- Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equationsJournal of Computational Physics, 2019