Compact diffractive optical neural networks with programmable binary nonlinearity and physics-aware adaptive readout
Abstract
Diffractive optical neural networks (DONNs) promise ultra-high-speed computing but face persistent challenges in realizing optical nonlinearity and maintaining experimental robustness. We demonstrate the c-DONN, a compact architecture utilizing a digital micromirror device (DMD), to implement trainable binary nonlinearity. Furthermore, we propose a physics-aware adaptive readout mechanism that significantly mitigates alignment errors and systemic noise without requiring time-consuming in situ retraining. Our prototype system achieves a classification accuracy of over 95% on the MNIST dataset and exceeds 92% in fingerprint recognition tasks. By addressing critical hurdles in nonlinearity and robustness, this work establishes a practical engineering framework for next-generation photonic intelligent systems in biometric and edge computing applications.
Fields of Study & Subjects
Neural Networks and Reservoir Computing
Neural Networks and Applications
Optical Network Technologies
Publication details
- Journal / source
- Photonics Research
- Publisher
- Optica Publishing Group
- Volume
- 14
- Issue
- 7
- Pages
- 2793–2793
- Publication date
- 23 April 2026
- Language
- EN