Cosine Similarity-Based Adaptive Weighting for Physics-Informed Neural Networks
This study investigates the application of adaptive loss weighting in the training of physics-informed neural networks (PINNs) to dynamically balance data, boundary, and physics losses. By employing cosine similarity of gradients, the proposed approach mitigates gradient dominance, ensuring more effective optimization. Numerical experiments on the Burgers equation demonstrate accelerated error convergence and enhanced training efficiency, underscoring its potential for tackling complex and high-dimensional physical systems.
Authors
- Jingze Wang (ORCID: https://orcid.org/0009-0005-0748-3142)
- Jixiao Wang (ORCID: https://orcid.org/0000-0002-5027-2456)
- Yunpeng Li (ORCID: https://orcid.org/0000-0001-5053-1393)
- Weisheng Zhang (ORCID: https://orcid.org/0000-0001-8608-4687)
Institutions
- Dalian University of Technology (CN)
- China General Nuclear Power Corporation (China) (CN)
- State Key Laboratory of Structural Analysis for Industrial Equipment
Publication Details
- Published
- 2024-11-29
- DOI
- https://doi.org/10.1109/icmsp64464.2024.10867237
- Primary Topic
- Model Reduction and Neural Networks
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China
- National Key Research and Development Program of China