Wavefront-aberration-tolerant diffractive deep neural networks using volume holographic optical elements.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computational Science

Background:

• Increasing demand for AI computational performance drives research into novel computing architectures.
• Diffractive deep neural networks (D2NNs) offer high-speed AI computation using optical modulation via diffractive optical elements (DOEs).
• Volume holographic optical elements (vHOEs) provide unique wavelength and angle selectivity but are susceptible to fabrication errors causing wavefront aberrations.

Purpose of the Study:

• To develop and demonstrate a training method for D2NNs that can adapt to unknown wavefront aberrations.
• To evaluate the performance of D2NNs utilizing vHOEs with the proposed adaptive training method.
• To explore the potential of vHOE-based D2NNs for advanced optical computing applications.

Main Methods:

• Proposed a novel training methodology to enable D2NNs to adapt to unknown wavefront aberrations.
• Fabricated and integrated volume holographic optical elements (vHOEs) into a D2NN architecture.
• Conducted optical experiments for handwritten digit classification to validate the proposed method.

Main Results:

• The adaptive training method significantly improved classification accuracy by approximately 58 percentage points in optical experiments.
• Demonstrated a functional D2NN utilizing vHOEs capable of handling complex optical aberrations.
• Validated the effectiveness of the proposed method in real-world optical experimental settings.

Conclusions:

• The developed training method effectively compensates for wavefront aberrations in vHOE-based D2NNs.
• This research paves the way for more robust and accurate optical AI computing systems.
• The findings suggest promising applications in multi-wavelength parallel optical computing, bioimaging, and optical communication.