Stochastic Physics-Constrained Operator Inference for Complex Systems

Abstract

Quadratically nonlinear systems driven by additive noise arise throughout science and engineering, both as stochastic dynamical systems in their own right and as reduced-order models (ROMs) of turbulent flows. We propose stochastic physics-constrained operator inference (SPOpInf), a non-intrusive method that learns a stochastic differential equation, i.e., a quadratic--linear drift and a diffusion amplitude, from data. Unlike operator inference (OpInf), which fits a deterministic drift by least squares, SPOpInf maximizes an Euler--Maruyama increment likelihood in which the residual covariance is itself an unknown, so the deterministic and stochastic components are calibrated jointly; it enforces the energy-preserving quadratic structure exactly, through linear equality constraints imposed inside the optimization; and it works in an upper-triangular parametrization that removes the structural non-identifiability of the full quadratic tensor. The resulting optimization is solved by an expectation--maximization (EM)-based iterative learning algorithm that updates the drift coefficients and the noise levels in turn: each iteration requires only an inexpensive weighted least-squares solve for the drift and a residual-variance update for the noise. With five stochastic benchmark systems that show intermittency and chaos, SPOpInf recovers the deterministic operators more accurately than OpInf, satisfies the imposed constraints to machine precision, accurately estimates the noise amplitudes, and reproduces long-time statistics. Applied to a stochastic Burgers equation and a quasi-geostrophic double-gyre flow, the same estimator yields non-intrusive stochastic ROMs that remain stable over long times where unconstrained alternatives blow up.