Abstract
Deep generative models hold great promise for representing complex physicalsystems, but their deployment is currently limited by the lack of guarantees onthe physical plausibility of the generated outputs. Ensuring that knownphysical constraints are enforced is therefore critical when applyinggenerative models to scientific and engineering problems. We address thislimitation by developing a principled framework for sampling from a targetdistribution while rigorously satisfying physical constraints. Leveraging thevariational formulation of Langevin dynamics, we propose Split AugmentedLangevin (SAL), a novel primal-dual sampling algorithm that enforcesconstraints progressively through variable splitting, with convergenceguarantees. While the method is developed theoretically for Langevin dynamics,we demonstrate its effective applicability to diffusion models. In particular,we use constrained diffusion models to generate physical fields satisfyingenergy and mass conservation laws. We apply our method to diffusion-based dataassimilation on a complex physical system, where enforcing physical constraintssubstantially improves both forecast accuracy and the preservation of criticalconserved quantities. We also demonstrate the potential of SAL for challengingfeasibility problems in optimal control.