TASE-Stabilized Time-Evolving Natural Gradient Methods for Diffusion-Dominated PDEs (Q7353453)

We propose a new method TASE-TENG that combines Time-Evolving Natural Gradient (TENG) methods for neural approximations of time-dependent PDEs with a highly stable explicit operator, which is constructed from a learnable surrogate operator and improves stability in stiff diffusion-dominated problems while retaining accuracy.