| Abstract | Topology-optimized silicon-photonic devices promise ultra-compact footprints and tailored spectral responses, but fabrication-induced deviations often degrade their performance in practice. In this work, we experimentally demonstrate the effectiveness of deep learning-based models in mitigating this issue by integrating adaptive, feature-specific corrections into the design layout. We apply this technique to a two-channel C-band wavelength-division demultiplexer with a 20-nm channel spacing and a flat-top passband within a highly compact 3 × 5 μm² footprint. We achieve significant improvement in optical performance compared to devices fabricated without layout corrections. The center-wavelength shift is reduced from 23.1 nm to 3.8 nm, and insertion loss drops from 3.7 dB to 1.0 dB, all while preserving strong crosstalk suppression of 22 dB, increased from 14.2 dB, and out-of-band rejection of 18.5 dB, increased from 17.5 dB. Additionally, our approach preserves the critical flat-top channel profiles, with a 1-dB bandwidth of 13 nm for channel 1 and 15 nm for channel 2 in the corrected fabricated device. In contrast, traditional correction methods, which attempt to add uniform dilation or erosion, reduce the center-wavelength shift but fail to maintain the passband shape. These results validate deep learning-driven corrections as a powerful approach for integrating high-performance, ultra-compact silicon photonic devices and pave the way toward commercially viable photonic systems. |
|---|
