C-DONN: compact diffractive optical neural network with deep learning regression.

Area of Science:

• Photonics
• Optical Computing
• Artificial Intelligence Hardware

Background:

• On-chip diffractive optical neural networks (DONNs) offer significant computation capacity but face integration limitations.
• Current methods approximate light propagation in metalines, hindering further miniaturization and performance.
• Silicon-on-insulator (SOI) platforms are standard for integrated photonic devices.

Purpose of the Study:

• To propose a new method for improving the integration level of on-chip DONNs.
• To accurately characterize light propagation in subwavelength metalines without approximations.
• To demonstrate a compact-DONN (C-DONN) with enhanced performance.

Main Methods:

• Development of a deep mapping regression model (DMRM) to characterize light propagation in metalines.
• Elimination of approximate characterization conditions (slot groups, extra length).
• Design and benchmarking of a compact-DONN (C-DONN) using the DMRM.

Main Results:

• Achieved an integration level exceeding 60,000 for on-chip DONNs.
• Eliminated the need for approximate physical models in metaline characterization.
• Demonstrated a C-DONN with 93.3% testing accuracy on the Iris plants dataset.

Conclusions:

• The DMRM provides an accurate and efficient method for characterizing light propagation in DONN metalines.
• This approach significantly enhances the integration level of on-chip DONNs.
• The proposed method offers a viable pathway for developing large-scale, high-performance optical neural networks on-chip.