Latent-MoE architecture enhances physics-informed neural networks for complex PDEs

Researchers have developed Latent-MoE, a novel domain-aware Mixture-of-Experts architecture designed to improve the performance of physics-informed neural networks (PINNs) on partial differential equations (PDEs) with complex, multi-regime physics. Standard PINNs struggle with such problems due to issues with their neural tangent kernel (NTK), which can lead to long-range coupling and gradient conflicts during training. Latent-MoE addresses this by interleaving domain-aware MoE blocks within a shared backbone, allowing for localized learning while enabling capacity to flow across different regions. This approach significantly outperforms existing methods on benchmarks with multi-stage time-variable physics, reducing gradient conflict and improving prediction accuracy by over an order of magnitude. AI

IMPACT This new architecture could significantly improve the accuracy and efficiency of AI models used in scientific simulations and engineering.

RANK_REASON The cluster contains a research paper detailing a new architecture for physics-informed neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• FB-PINNs
• Latent-MoE
• neural tangent kernel
• partial differential equations
• physics-informed neural networks
• X-PINNs

• paper
• model release

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