All-optical classification using a compact single-layer dual-wavelength differential diffractive network

Diffractive deep neuron networks (D²NNs), which compute with photons instead of electrons, hold the potential to accelerate the advancement of artificial intelligence by delivering orders-of-magnitude improvements in computational speed, massive parallelism, and ultralow energy consumption. However, conventional multi-layer D²NNs are fundamentally limited by inter-layer misalignments, which increase optical architecture complexity and critically impair performance, especially at visible wavelengths where optical inter-layer alignment is highly sensitive. Here, we demonstrate a compact, single-layer dual-wavelength differential D²NNs that combines wavelength multiplexing and differential detection, achieving high classification accuracy with only a single diffractive modulation layer and effectively mitigating inter-layer mechanical alignment errors. By harnessing complementary spatial-frequency information encoded at two distinct wavelengths, the network effectively circumvents the intrinsic non-negativity constraint imposed by intensity-only optical detection. Numerical results demonstrate high classification accuracies of 98.59% on MNIST and 90.4% on Fashion-MNIST using only 40k tunable parameters, demonstrating improved performance compared with conventional five-layer cascaded D²NNs operating under single-wavelength illumination, which achieve 91.33% and 83.67% accuracy, respectively, despite utilizing 200k trainable parameters. In addition, we further demonstrate that the proposed method maintains robust performance, achieving 97.95% accuracy on MNIST and 88.7% on Fashion-MNIST, even when the number of trainable parameters is reduced to just 10k, and exhibits superior resilience against random phase perturbations. Compared with conventional multi-layer cascaded D²NNs operating under single-wavelength illumination, this compact single-layer design scheme not only outperforms cascaded D²NNs with single wavelength in classification accuracy, but also alleviates the performance degradation caused by the inter-layer mechanical misalignment, thereby offering a more robust and scalable framework for practical photonic computation applications.