Double-layer all-optical inference via sequential nonlinear optical diffraction
Phys. Rev. Applied 25, 064067 – Published 23 June, 2026
DOI: https://doi.org/10.1103/cd67-l2b5
Abstract
We demonstrate a double-layer all-optical inference platform for classification using cascaded nonlinear diffraction. The system utilizes a dual-pass configuration through a nonlinear optical crystal, shaping a superposition of Laguerre-Gaussian modes across two layers to steer optical intensity into class-specific detector regions. By optimizing via an in situ genetic algorithm with 10 trainable degrees of freedom, we achieve 88.65% accuracy on the ten-class Modified National Institute of Standards and Technology (MNIST) dataset, leveraging crystal temperature as a critical hyperparameter to optimize noncollinear mode mixing. This work establishes a robust pathway for high-accuracy, low-complexity all-optical computing through structured light and cascaded nonlinear optical interactions.
