Compact topology-optimized wavelength demultiplexers with flat-top response enabled by deep learning-based...

Area of Science:

• Photonics
• Materials Science
• Artificial Intelligence

Background:

• Topology-optimized silicon-photonic devices offer miniaturization and custom spectral control.
• Fabrication imperfections commonly degrade device performance, limiting practical applications.
• Existing correction methods struggle to maintain desired spectral characteristics.

Purpose of the Study:

• To demonstrate deep learning-based layout corrections for mitigating fabrication deviations in silicon photonics.
• To improve the optical performance of topology-optimized devices, specifically a wavelength-division demultiplexer.
• To validate the effectiveness of adaptive, feature-specific corrections compared to traditional methods.

Main Methods:

• Development and application of deep learning models for adaptive, feature-specific layout corrections.
• Fabrication and characterization of a two-channel C-band wavelength-division demultiplexer (3 µm × 5 µm).
• Comparison of optical performance metrics (center-wavelength shift, insertion loss, crosstalk, out-of-band rejection, passband shape) before and after correction.

Main Results:

• Reduced center-wavelength shift from 23.1 nm to 3.8 nm.
• Decreased insertion loss from 3.7 dB to 1.0 dB.
• Improved crosstalk suppression to 22 dB and out-of-band rejection to 18.5 dB.
• Preserved flat-top passband profiles with 1-dB bandwidths of 13 nm and 15 nm.
• Outperformed traditional correction methods in maintaining passband shape.

Conclusions:

• Deep learning-driven layout corrections effectively mitigate fabrication-induced performance degradation in silicon photonics.
• This approach enables high-performance, ultra-compact photonic devices with tailored spectral responses.
• The validated technique paves the way for commercially viable, advanced photonic systems.