Generative Priors Accelerate Inverse Design for Regularized, High Performance Integrated Photonics

Adjoint-based inverse design has revolutionized integrated photonics, yet existing workflows remain computationally inefficient due to a heavy dependence on high-quality initialization. We propose that archives of past optimization traces, typically discarded, should be repurposed as $\textbf{reusable scientific memory}$. We introduce a physics-compatible generative framework that learns a conditional prior over optimized device layouts to seed downstream adjoint refinement. Our model, a conditional Variational Autoencoder (cVAE), is trained on archived designs and incorporates a differentiable filter-project operator to enforce fabrication-feasible geometries. At inference, the model samples candidate templates which are ranked via a reduced-cost electromagnetic screen before final adjoint optimization. On arbitrary-ratio silicon power splitters with a matched 60-step refinement budget, our cVAE-seeded approach reduces median insertion loss by $3.2\times$ (to $0.127$ dB) with no floor objective function and from $5.07$ dB to $1.94$ dB in the with-floor setting. Crucially, the learned prior also improves both convergence speed by 50\% and the geometry-performance trade-off. Here, we demonstrate $\textbf{amortization of historical optimization}$ into reusable priors transforms restart-heavy global search into efficient, template-guided refinement.