Universal Construction of Generalized Lyapunov Functions for Nonlinear Dynamical Systems Using Physics-Informed Neural Networks

Abstract

A scalar potential landscape is one of the most useful ways to understand the stability and transition of a dynamical system. For non-gradient dynamics, however, the construction of a global Lyapunov-type scalar for nonlinear flows with recurrent structures remains a major obstacle. We introduce the generalized Lyapunov function, a scalar function that is non-increasing along deterministic trajectories, as a unifying notion of nonequilibrium potential. Ordinary Lyapunov functions, Freidlin--Wentzell quasi-potentials, and Ao-type potentials are recovered as special representatives. We then propose a data-free physics-informed neural-network framework in which the Lyapunov inequality and a weak divergence-scale compatibility condition are directly embedded into the loss function. The method is tested on linear systems, the Hopf normal form, the van der Pol oscillator, and a three-dimensional Hopf-link flow with two linked limit cycles. The learned landscapes agree with available analytical benchmarks and reveal the invariant sets as low-potential or constant-potential structures, providing a practical route to potential-landscape construction for nonlinear non-gradient systems.