Improved adaptive optimization algorithm for training physics-informed neural networks

doi:10.1016/j.asoc.2026.115473

Improved adaptive optimization algorithm for training physics-informed neural networks

2026-05-16

0

PRE

AI

DOI：10.1016/j.asoc.2026.115473

Original

Original request for help

Save

Abstract

En 中文

• AdaLB combines belief-based scaling with long-term gradient memory. • A power-series accumulator defines a time-varying second-moment decay. • AdaLB helps stabilize adaptive updates in PINN training. • AdaLB performs well on Burgers and Poisson PINN benchmarks. • AdaLB can provide a stronger first stage for L-BFGS refinement.

Keywords:

AdaLB

adaptive optimization

physics-informed neural networks

gradient memory

L-BFGS refinement

Improved adaptive optimization algorithm for training physics-informed neural networks

Abstract

Journal

Organization