Data-Driven Decomposition of Conservative and Non-Conservative Dynamics in Multiscale Systems

Abstract: We present a data-driven framework for reconstructing physics-consistent reduced models of multiscale systems. Our approach addresses a fundamental challenge in model reduction: decomposing the deterministic drift into its conservative (reversible) and non-conservative (irreversible) components while preserving the system's invariant measure and autocorrelation. We leverage the k-means Gaussian-mixture method (KGMM) to robustly estimate the score function -- the gradient of the logarithm of the steady-state probability density. This score function is then combined with a finite-volume discretization of the Perron-Frobenius generator to identify a drift matrix whose symmetric part captures the conservative dynamics and whose antisymmetric part encodes the minimal irreversible circulation required by the empirical data. The result is a reduced Langevin model that preserves the stationary density, reproduces short-time correlations by construction, and maintains computational efficiency. We validate our framework on three increasingly complex systems: a one-dimensional nonlinear stochastic system with multiplicative noise, a two-dimensional asymmetric four-well potential system with non-gradient drift, and the stochastic Lorenz-63 attractor. In all cases, our surrogate models accurately recover both the stationary distribution and autocorrelation structure of the original dynamics, while providing an interpretable separation of conservative and non-conservative forces.