Artificial intelligence (AI) is increasingly deployed in diverse applications, with a growing need for robust and energy-efficient hardware capable of functioning in extreme environments like inclement weather, deep sea, and space. Neuromorphic hardware offers a promising solution. Optical computing, leveraging light's speed and passive propagation, presents advantages in speed and energy consumption compared to electronic counterparts, achieving femtojoules-per-bit. Optical parameters (refractive index, transmittance) are generally less sensitive to temperature and humidity changes than electronic ones, making photonic components suitable for extreme conditions. Diffractive neural networks (DNNs) are wave-based optical neural networks that mimic the human nervous system in 3D, enabling direct parallel processing of optical image data. However, existing DNNs, often 3D printed from organic materials, lack robustness and long-term stability, limiting their compactness and accuracy. Furthermore, the spatial separation of diffractive layers and operation in the terahertz band result in centimeter-scale devices unsuitable for on-chip integration. While some Si wafer-integrated DNNs exist, they lose the parallel processing benefits of 3D structures. This work addresses these limitations by creating a 3D-integrated DNN using stable materials, paving the way for robust AI applications in extreme environments.

Existing research demonstrates the potential of diffractive neural networks (DNNs) for various AI tasks, including image recognition, optical computing, phase retrieval, and terahertz pulse shaping. These networks typically use cascaded diffractive layers, often created via 3D printing. However, organic materials used in these fabrication methods compromise robustness and lifetime. Many DNNs use spatially separated layers, operating in the terahertz band leading to large, non-integrated systems. While on-chip DNNs integrated on Si wafers have been reported, these designs often sacrifice the inherent advantages of three-dimensional wave propagation and parallel processing. This paper builds upon this existing work by addressing the critical need for a robust, compact, and long-lasting on-chip DNN suitable for deployment in extreme environments.

A bilayer DNN chip was fabricated using double-sided lithography to engrave two binary phase-modulated diffractive layers onto a single-crystal quartz substrate. Each layer contained over one million neurons (8 µm × 8 µm). The design used a 500 µm quartz substrate, resulting in a fixed layer spacing. The angular spectrum diffraction method was used to simulate light propagation during training, performed using TensorFlow 2.0 framework. The Modified National Institute of Standards and Technology (MNIST) handwritten digit database was used for training, employing a subset of 1000 images due to training time and alignment complexity considerations. The output amplitude field was optimized to follow a Gaussian distribution, enhancing signal concentration and reducing noise. The trained phase values were binarized, and the resulting patterns were etched onto the quartz substrate using photolithography and plasma dry etching. Handwritten digit recognition was tested experimentally, employing a 4f system for beam expansion and using spatial filtering to remove unwanted diffraction orders. The output intensity distribution was captured using a CMOS camera. Robustness analysis involved simulations examining the impacts of fabrication errors (layer misalignment, substrate thickness variations) and testing conditions (wavelength shift, input layer misalignment). Lifetime testing involved accelerated aging at high temperatures (1400 °C), assessing performance degradation and estimating room temperature lifetime using the Arrhenius equation.

The fabricated DNN chip achieved an 82% recognition accuracy for ten handwritten digits (0-9) in experimental testing. Simulation results using the same parameters indicated a higher accuracy of 85.4%. The chip demonstrated high performance stability at high temperatures. After annealing at 1400 °C for 2 hours, the recognition accuracy for two digits (0 and 1) remained at 100%, despite increased surface roughness. The room-temperature lifetime was estimated to be 1.84 × 10²³ trillion years based on accelerated aging experiments. Analysis showed that the chip exhibited robustness against reasonable fabrication errors (layer misalignment within 1-2 µm, thickness error up to ±50 µm). The impact of wavelength shift could be mitigated through calibration. Simulation showed a higher accuracy (96.1%) and lower loss (0.198) using the training set. Binary phase modulation, while impacting diffraction efficiency (causing the visibility of the unmodulated input image in the output), was chosen to maintain high accuracy, as a tradeoff parameter. The bilayer DNN performed significantly better (85.4% accuracy) than a monolayer DNN (91.2% accuracy) under the same conditions. The study also validated the chip's applicability to other tasks such as fashion product recognition and phase imaging through simulations. The chip's robustness was further analyzed for tolerance to damaged neurons where the performance remained high (highest performance when damaged area <20%).

The results demonstrate the successful development of a highly robust and long-lasting on-chip DNN suitable for extreme environments. The achieved accuracy, while not exceeding state-of-the-art electronic DNNs, is significant given the chip's unique features. The use of quartz as the substrate and the integrated, bilayer design directly address the limitations of previously reported DNNs. The high-temperature stability and remarkably long lifetime are crucial for applications where device reliability and longevity are paramount. The fact that the chip demonstrated function in additional AI tasks beyond the specific training provides further evidence of its versatility and potential impact.

This work successfully demonstrated a millimeter-scale bilayer DNN chip integrated on a quartz substrate, offering superior robustness and an exceptionally long lifetime for AI applications in extreme environments. The 82% accuracy in handwritten digit recognition, coupled with its robustness against fabrication and operational errors, highlights the potential of this technology. Future research could focus on increasing the number of layers using bonding techniques to improve accuracy, incorporating nonlinear activation functions to boost performance, and integrating the DNN chip with other components (cameras, electrical neural networks) to create a fully integrated system.

The current design uses only two diffractive layers, and the neurons are not fully connected, limiting the achievable accuracy. The binary phase modulation, while simplifying fabrication and alignment, reduces diffraction efficiency. The impact of unexpected damage beyond conventional aging during long-term operation remains to be fully investigated. The study focused primarily on handwritten digit recognition, and while simulations showed potential in other tasks, further experimental validation is necessary.