PearSAN: A Machine Learning Method for Inverse Design Using Pearson Correlated Surrogate Annealing

ABSTRACT Inverse design in nanophotonics is often bottlenecked by the curse of dimensionality, particularly for metasurfaces with many degrees of freedom, where the optimization landscape is highly non‐convex. In this work, we introduce Pearson Correlated Surrogate Annealing (PearSAN): a machine learning‐assisted framework that efficiently navigates these complex spaces. PearSAN uses a surrogate model designed to capture relative performance trends within the latent space of a generative model. This allows the algorithm to rapidly converge toward optimal solutions without requiring exhaustive physical simulations. As a showcase example, PearSAN is applied to the design of thermophotovoltaic (TPV) metasurfaces by matching the working bands of a thermal radiator and a photovoltaic (PV) cell. PearSAN achieves a state‐of‐the‐art design efficiency of , outperforming conventional baselines in both convergence speed and final device performance. These results establish PearSAN as a robust tool for photonics applications requiring precise spectral engineering.