Reproduce “DeepGreen: Deep Learning of Green's Functions for Nonlinear Boundary Value Problems” on OpenResearch

About this paper

Craig R. Gin, Daniel E. Shea, Steven L. Brunton, J. Nathan Kutz

DeepGreen is a deep learning framework that transforms nonlinear boundary value problems into linear systems within a learned latent space. This approach enables the discovery of a 'nonlinear Green's function,' allowing rapid determination of solutions for new forcing functions through efficient matrix operations.

Key insights

• The learned linear operator and its inverse consistently exhibit diagonal dominance, suggesting the network discovers nearly orthonormal basis transformations.
• Analysis revealed that multiple distinct latent space representations can achieve consistent performance, indicating non-uniqueness in the discovered linearizing transformations.
• The linear operator loss term consistently represents the highest component during training, highlighting that enforcing strict linearity in the latent space is the most challenging optimization aspect.

Summary by alphaXiv.