Development of regularization method for physics problems using PINNs and data augmentation using equation discovery technique
Authors:
Open Access
- 1 January 2025
- journal article
- Published by Japan Society of Mechanical Engineers in Transactions of the JSME (in Japanese)
- Vol. 91 (949) , 25-00102
Abstract
This study proposes a regularization method for high-accuracy prediction by machine learning under small data conditions by combining Physics-Informed Neural Networks (PINNs) and data augmentation based on equation discovery using Sparse Identification of Nonlinear Dynamical Systems (SINDy). PINNs incorporate governing equations into the loss function to enforce physical consistency, while SINDy extracts mathematical models from data, enabling interpolation with continuity and physical interpretability. Validation using crack propagation data (25 points) demonstrated that the combination of SINDybased augmentation and PINNs reduced the sum of squared errors (SSE) by approximately 28% compared to linear interpolation. The effectiveness of the proposed method was also confirmed for different datasets (59 points). These results indicate that the method is particularly effective in improving learning accuracy for engineering problems where only small amounts of data are available.Keywords
This publication has 10 references indexed in Scilit:
- Prediction of plural crack propagation by machine learning limited small dataset2023
- Prediction of fatigue crack growth using convolutional neural network (2nd Report, Prediction of crack propagation on different levels)Transactions of the JSME (in Japanese), 2023
- Constructing an Surrogate Model for Pedestrian Protection Using Machine LearningThe Proceedings of The Computational Mechanics Conference, 2022
- Physics-informed neural networks (PINNs) for fluid mechanics: a reviewActa Mechanica Sinica, 2021
- Thermodynamics-based Artificial Neural Networks for constitutive modelingJournal of the Mechanics and Physics of Solids, 2021
- AI Feynman: A physics-inspired method for symbolic regressionScience Advances, 2020
- Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equationsJournal of Computational Physics, 2019
- Discovering governing equations from data by sparse identification of nonlinear dynamical systemsProceedings of the National Academy of Sciences, 2016
- The Elements of Statistical Learning2009
- Fatigue Crack Growth Simulation Using S-FEMTRANSACTIONS OF THE JAPAN SOCIETY OF MECHANICAL ENGINEERS Series A, 2008