{"id":"arxiv/2603.20816","url":"https://arxiv.org/abs/2603.20816","tags":["math.NA"],"score":0,"title":"Preserving Conservation Laws in the Time-Evolving Natural Gradient Method via Relaxation and Projection Techniques","domain":"math.NA","source":"arxiv","authors":["Zihao Shi","Dongling Wang"],"reviews":0,"subject":"math.NA","summary":"Neural networks have demonstrated significant potential in solving partial differential equations (PDEs). While global approaches such as Physics-Informed Neural Networks (PINNs) offer promising capabilities, they often lack inherent temporal causality, which can limit their accuracy and stability for time-dependent problems. In contrast, local training frameworks that progressively update network parameters over time are naturally suited for evolving PDEs. However, a critical challenge remains: many physical systems possess intrinsic invariants -- such as energy or mass -- that must be…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-03-21","reactions":{}}