WorksPhysics and Astronomy
Physics-informed neural network for model-based global computation of the asymptotic phase of limit-cycle oscillators
Abstract
The asymptotic phase is the central object of phase reduction theory for limit-cycle oscillators. It is a scalar coordinate defined throughout the basin of attraction of a limit cycle that advances at a constant natural frequency along every trajectory. Computing the phase from a given mathematical model requires solving a first-order partial differential equation (PDE) over the entire basin, which is challenging due to high computation costs and complex boundary conditions. In this study, we propose a physics-informed neural network to efficiently calculate the asymptotic phase of limit-cycle oscillators. By incorporating three types of loss functions, we train a physics-informed neural network (PINN) to satisfy the governing PDE for the asymptotic phase and the adjoint equation for the phase gradient. The trained network delivers the phase function accurately and, via automatic differentiation, allows the calculation of the phase gradient governing the oscillator's response to perturbations, which enables rapid entrainment closed-loop control using strong inputs. We validate the approach using three types of limit-cycle oscillators and demonstrate that the PINN can facilitate fast and global phase-locking control.