Mechanical diffusion models

Diffusion models have gained considerable attention in generative modeling, driven by their theoretical foundation in non-equilibrium thermodynamics. However, their implementation on digital computers faces challenges of slow sampling speeds and substantial energy consumption. In this theoretical work, we propose an alternative approach utilizing physical systems intrinsically coupled to a thermodynamic bath. Specifically, we investigate nano-electro-mechanical resonators, leveraging their thermal noise-induced fluctuations and engineered nonlinear couplings to create an ultra-low-energy sampling machine for score-based generative modelling.