CNNs on manifolds tackle boundary value problems with improved accuracy

Researchers have developed novel convolutional neural network (CNN) methods for approximating functions and solving elliptic boundary value problems on compact Riemannian manifolds. These methods demonstrate improved approximation rates that depend on the manifold's intrinsic dimension rather than its ambient dimension, helping to overcome the curse of dimensionality. The proposed physics-informed CNN (PICNN) framework specifically addresses boundary value problems by introducing a spectral boundary loss, which enhances accuracy, convergence, and stability compared to standard physics-informed neural networks (PINNs). AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces novel CNN techniques for manifold approximation and boundary value problems, potentially improving scientific computing accuracy.

RANK_REASON This is a research paper detailing new methods for CNN approximation on manifolds and for solving boundary value problems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• CNN
• Riemannian manifolds
• PINN
• PICNN
• Fast Fourier Transforms
• Laplace-Beltrami operator

• paper
• other