This paper proposes integrating paper-clip spiral waveguides into programmable photonic neural network cores to address challenges in delay-line compactness, phase control precision, and thermal management. The architecture embeds spirals as tunable delay elements within Mach-Zehnder interferometer meshes, utilizing thermo-optic phase-shifting strategies optimized for the spiral geometry. Theoretical analysis indicates 10× improvements in phase-shifting energy efficiency and 5× in delay-line density compared to conventional straight-waveguide implementations. The work establishes a comprehensive experimental validation framework including fabrication protocols for next-generation photonic accelerators.
Key findings
Paper-clip spiral waveguides demonstrate ultra-low delay loss of 0.5 dB ns−1 and propagation losses as low as 0.06 dB cm−1 at 1550 nm.
The proposed architecture theoretically achieves 10× improvement in phase-shifting energy efficiency through optimized thermal management of spiral geometries.
Delay-line density improves by 5× compared to conventional straight-waveguide implementations due to compact 9.5 cm path length in (0.30×3.00) mm² footprint.
Multi-mode waveguide operation (>2µm width) reduces sidewall scattering while maintaining single-mode coupling compatibility.
Complete experimental validation framework encompasses MPW fabrication, characterization protocols, and neural network benchmarking.
Limitations & open questions
Experimental validation pending; current results are theoretical projections based on component-level demonstrations.
Integration challenges remain regarding multi-mode to single-mode waveguide coupling and thermal crosstalk in dense arrays.
Fabrication tolerance and yield risks associated with undercut structures and heater integration in spiral geometries.