Improving physics-informed neural networks with meta-learned optimization
This work addresses the challenge of improving accuracy in solving differential equations for researchers in mathematical physics, though it appears incremental as it builds on existing physics-informed neural networks with a novel optimization twist.
The paper tackles the problem of reducing error in physics-informed neural networks for solving differential equations by using meta-learned optimization instead of fixed optimizers, achieving substantial error reduction across several equations including linear advection and Burgers' equation.
We show that the error achievable using physics-informed neural networks for solving systems of differential equations can be substantially reduced when these networks are trained using meta-learned optimization methods rather than to using fixed, hand-crafted optimizers as traditionally done. We choose a learnable optimization method based on a shallow multi-layer perceptron that is meta-trained for specific classes of differential equations. We illustrate meta-trained optimizers for several equations of practical relevance in mathematical physics, including the linear advection equation, Poisson's equation, the Korteweg--de Vries equation and Burgers' equation. We also illustrate that meta-learned optimizers exhibit transfer learning abilities, in that a meta-trained optimizer on one differential equation can also be successfully deployed on another differential equation.