Reproduce “HyperPINN: Learning parameterized differential equations with physics-informed hypernetworks” on OpenResearch

About this paper

Filipe de Avila Belbute-Peres, Yi-fan Chen, Fei Sha

Researchers from Carnegie Mellon University and Google Research developed HyperPINN, an architecture that leverages hypernetworks to efficiently learn solutions to parameterized differential equations. This method generates specific solution networks for varying physical parameters, achieving high accuracy comparable to larger models while maintaining significantly faster inference times for the main solution network.

Key insights

• Offloading the complexity of learning the parameter space to a dedicated hypernetwork allows the repeatedly evaluated main network to remain small and computationally efficient.
• Integrating meta-learning (hypernetworks) with physics-informed methods provides a scalable approach for efficiently learning families of differential equation solutions.
• The approach is versatile, demonstrated effectively with both partial differential equations (1D Burgers' equation) using standard PINN loss and ordinary differential equations (Lorenz system) using a multistep neural network loss.

Summary by alphaXiv.