Researchers have developed a new method called Target-Guided Selective Reweighting PINN (TGSR-PINN) to improve the transfer learning capabilities of physics-informed neural networks (PINNs) for inverse problems. This approach addresses challenges like negative transfer, where models trained on one set of physical parameters perform poorly on another due to differing mechanisms or noise. TGSR-PINN uses a target-evidence-driven strategy to correct representations by scoring neurons and applying selective soft decay to weights and biases of underperforming neurons. Experiments show this method enhances parameter recovery while maintaining accuracy in various complex physics tasks, including advection-diffusion and cross-PDE-family transfers. AI
IMPACT Improves transfer learning for physics-informed neural networks, potentially enabling more accurate parameter recovery in complex scientific simulations.
RANK_REASON The cluster contains a research paper detailing a novel method for physics-informed neural networks.
- 2D advection-diffusion
- Allen--Cahn
- Burgers
- Gaussian mixture model
- physics-informed neural networks
- reaction--diffusion
- Target-Guided Selective Reweighting PINN
- TGSR-PINN
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
