Muon with Spectral Guidance: Efficient Optimization for Scientific Machine Learning
Published in arXiv preprint, 2026
We extend the Muon optimizer with spectral guidance, exploiting the spectrum of the gradient covariance to choose step sizes that respect curvature anisotropies common in scientific-ML problems (stiff PDEs, multi-scale features). The result is faster and more stable training on PINN- and operator-learning workloads.
Recommended citation: Lu, B., Zhang, J., & Lin, G. (2026). "Muon with Spectral Guidance: Efficient Optimization for Scientific Machine Learning." arXiv:2602.16167.
Download Paper