CSE Community Seminar | September 26, 2025

Abstract
The (generalized) Fluctuation–Dissipation Theorem (GFDT) links small changes in forcings or initial conditions to predictable changes in statistics. Applications are diverse and include determining climate change with respect to anthropogenic forcing as well as parameter calibration for models. The theorem offers a principled route to quantify how means, higher moments, or extreme events respond to changes in forcing. What held GFDT back in practice was the need to characterize the invariant distribution, particularly its “score-function”, in non-Gaussian systems. I’ll show how recent advances in generative modeling make this tractable: denoising score-matching with convolutional networks for spatial data, and a lightweight clustering-plus-regression approach (KGMM) for low-dimensional dynamics. Together, these tools turn GFDT into a practical, data-driven workhorse that (i) predicts mean and higher-order responses without new simulations, (ii) outperforms Gaussian closures in nonlinear regimes, and (iii) can even deliver parameter Jacobians for direct statistical calibration. I’ll illustrate the workflow on several models, from simple potentials to 2D Navier–Stokes.