Machine-learning-driven optimization of vertical fiber-to-chip coupling system for co-packaged optics and in-package optical I/O applications

Efficient light coupling in photonic integrated circuits (PICs) is vital for co-packaged optics (CPO) and in-package optical I/O (OIO). Grating couplers in PICs offer a viable solution for vertical coupling to a fiber attachment above the chip but are sensitive to fabrication and alignment variations. Using micro-optics, like collimating micro-lenses, can mitigate these issues but adds system complexity. Designing such systems necessitates simulating light propagation across different scales, from the sub-wavelength feature size of the grating coupler to the millimeter scale of the micro-optics and fiber attachment above. We propose an automated optimization workflow driven by state-of-the-art machine learning algorithms that combines optical simulation techniques suitable for these scales. This workflow also includes robustness analysis to capture variations from fiber assembly, fabrication, and material errors, providing insights into key variations affecting yield in the manufacture process.