Abstract
We introduce a generative modeling framework for thermodynamic computing, inwhich structured data is synthesized from noise by the natural time evolutionof a physical system governed by Langevin dynamics. While conventionaldiffusion models use neural networks to perform denoising, here the informationneeded to generate structure from noise is encoded by the dynamics of athermodynamic system. Training proceeds by maximizing the probability withwhich the computer generates the reverse of a noising trajectory, which ensuresthat the computer generates data with minimal heat emission. We demonstratethis framework within a digital simulation of a thermodynamic computer. Ifrealized in analog hardware, such a system would function as a generative modelthat produces structured samples without the need for artificially-injectednoise or active control of denoising.