New initialization strategy tackles spectral bias in physics-informed neural networks

Researchers have developed a new initialization strategy for Fourier Feature Physics-Informed Neural Networks (PINNs) to address spectral bias, a common issue where certain frequencies of a target function converge slower during training. By analyzing training dynamics in the Neural Tangent Kernel regime, they derived an equation showing that initialization weights significantly impact frequency convergence rates. The proposed method tailors the initial weight distribution to the specific partial differential equation (PDE) being solved, balancing convergence across frequencies and improving prediction accuracy without increasing training costs. AI

IMPACT Improves training efficiency and accuracy for PINNs, potentially enabling more complex scientific simulations.

RANK_REASON Academic paper detailing a novel method for improving neural network training dynamics. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Francisco Sahli Costabal
• Hugging Face
• Neural Tangent Kernel
• physics-informed neural networks

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →