Weak Adversarial Neural Pushforward Mappings for Fokker-Planck Equations

Abstract: This paper presents a novel method for solving Fokker-Planck equations by learning neural samplers via a weak adversarial framework. We represent the solution distribution through a neural pushforward map, bypassing the limitations of density-based methods. A key innovation is our use of computationally efficient plane-wave test functions, whose derivatives are explicitly computed -- a treatment distinct from prior work. This approach handles distributions without densities and naturally enforces probability conservation.