Computer-generated holograms for optical neural networks: on-axis versus off-axis geometry.

Area of Science:

• Optics
• Computer Science
• Holography

Background:

• Optical neural networks (ONNs) utilize computer-generated holograms for interconnections.
• The space-bandwidth product (SBP) of holograms is critical for ONN performance.
• Off-axis holograms offer binary fabrication but have high SBP; on-axis holograms have lower SBP but require complex multilevel phase structures prone to fabrication errors.

Purpose of the Study:

• To investigate a modified on-axis holographic geometry for ONNs.
• To reduce the impact of fabrication errors in multilevel phase structures.
• To compare the performance of off-axis, on-axis, and modified on-axis geometries regarding interconnection accuracy, diffraction efficiency, and sensitivity to fabrication errors.

Main Methods:

• Fabrication and characterization of off-axis, on-axis, and modified on-axis holograms.
• Analysis of space-bandwidth product requirements for each geometry.
• Evaluation of interconnection accuracy and diffraction efficiency.
• Assessment of sensitivity to fabrication errors in multilevel phase structures.

Main Results:

• The modified on-axis geometry retains a lower space-bandwidth product per interconnection compared to off-axis holograms.
• This modified geometry significantly reduces the impact of fabrication errors inherent in multilevel phase structures.
• Performance metrics including interconnection accuracy and diffraction efficiency were quantitatively compared across the three geometries.

Conclusions:

• Modified on-axis holography presents a viable solution for fabricating ONNs with reduced sensitivity to errors.
• This approach balances the benefits of lower space-bandwidth product with improved fabrication tolerance.
• The modified on-axis geometry offers a promising pathway for more robust and scalable optical neural network implementations.