Simulation-based Inference for Fabrication Variations in Silicon Photonics

Performance of silicon photonic devices is significantly affected by the fine-scaled fabrication variations in their geometry such as width, height, and etch depth of waveguides. While a direct measurement is limited by metrology capabilities, a more viable approach is to infer geometry inversely from transmission spectra that are widely available from test structures during production. In this work, we investigate simulation-based inference (SBI) for calibrating the geometry of photonic ring resonators from finite-difference time-domain (FDTD) simulated data, and further explore the broader potential of SBI for uncertainty quantification of fabrication variations in photonics. Across different SBI models, we obtain amortized posteriors that localize nanometer-scale geometry from raw spectra, while the overconfidence identified by posterior diagnostics can be effectively mitigated through SBI ensembling. We further deploy the estimators to a high-fidelity simulator using only a few hundred additional simulations via multi-fidelity fine-tuning and sequential acquisition.