{"id":"arxiv/2601.23185","url":"https://arxiv.org/abs/2601.23185","tags":["math.NA"],"score":0,"title":"Preconditioning and Numerical Stability in Neural Network Training for Parametric PDEs","domain":"math.NA","source":"arxiv","authors":["Markus Bachmayr","Wolfgang Dahmen","Chenguang Duan","Mathias Oster"],"reviews":0,"subject":"math.NA","summary":"In the context of training neural network-based approximations of solutions of parameter-dependent PDEs, we investigate the effect of preconditioning via well-conditioned frame representations of operators and demonstrate a significant improvement on the performance of standard training methods. We also observe that standard representations of preconditioned matrices are insufficient for obtaining numerical stability and propose a generally applicable form of stable representations that enables computations with single- and half-precision floating point numbers without loss of precision.","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-01-30","reactions":{}}