Recent breakthroughs in deep learning have been fueled by powerful computers, yet their reliance on energy-intensive hardware limits further miniaturization. All-optical computing, which uses light itself to carry and process information, offers a revolutionary alternative with advantages like ultra-low power consumption, lightning speed, and massive parallelism. Inspired by this, researchers developed Optical Diffractive Neural Networks (ODNNs) that perform computations at near-light speed using layered diffraction patterns. While successful in tasks like image recognition, traditional ODNNs hit a major roadblock: to achieve higher integration for complex tasks, their neuron structures must shrink to scales comparable to the wavelength of light. This makes them exquisitely sensitive to microscopic fabrication flaws and alignment errors, creating a frustrating trade-off where high-performance designs become too fragile for practical manufacturing.

A research team led by Professor Ming Zhao at Huazhong University of Science and Technology (HUST), China, has developed a novel solution to this critical challenge. They propose a Hybrid Optical Diffractive Neural Network (H-ODNN) featuring diffractive layers with variable neuron sizes, unlike the uniform, tiny neurons in conventional designs. This architecture is inspired by the variable hidden layers found in software-based neural networks. The team rigorously tested 20 different network designs under various misalignment conditions using tasks like digit recognition. Remarkably, the H-ODNN structure matches or even surpasses the performance of vaccinated networks, and it does not require the complex structure or time-consuming re-training process, thus significantly reducing the development time. Crucially, the inclusion of larger, more fault-tolerant neuron structures in key layers significantly relaxes manufacturing precision requirements, promising higher production yields. This breakthrough paves a more practical and reliable way toward miniaturized, high-speed optical computing systems for next-generation AI.

The work entitled “Hybrid structure for enhancing robustness in optical diffractive neural networks without vaccination training” was published in Frontiers of Optoelectronics (published on Feb. 11, 2026).

DOI:10.2738/foe.2026.0012