Compact eternal diffractive neural network chip for extreme environments

The study addresses the need for robust, low-power, and high-speed neuromorphic hardware capable of operating in extreme environments such as inclement weather, deep sea, and space. Photonics is highlighted as an attractive platform due to light-speed passive propagation, low energy per operation, and environmental stability of optical parameters. Diffractive neural networks (DNNs) offer 3D wave-based processing that enables parallel handling of 2D optical images, but existing implementations often rely on 3D-printed organic materials with limited robustness and are centimeter-scale with separated layers (often terahertz band), hindering on-chip integration and long-term stability. Some on-chip diffractive neural networks on silicon exist but flatten 3D architectures into 2D, losing the parallel 3D advantage. The research question is how to realize a compact, truly 3D-integrated, robust DNN with long lifetime suitable for extreme environments. The paper proposes a bilayer DNN chip on a quartz substrate using double-sided lithography to achieve binary phase diffractive layers on both wafer surfaces, enabling stable interlayer propagation within quartz and on-chip 3D integration.

The paper reviews diffractive neural networks (DNNs) and their applications in image recognition, optical computing, phase retrieval, adaptive focusing, and terahertz pulse shaping, emphasizing their 3D wave-based processing advantage over waveguide-based optical neural networks. Existing DNNs commonly use 3D printing with organic materials, leading to limited robustness and centimeter-scale, separated layers (often in the THz regime), which impede compact integration and stability. On-chip implementations on silicon have been reported, but they reduce 3D networks to 2D designs, sacrificing parallel processing benefits for 2D optical images. The authors identify a gap for a stable-material, 3D-integrated DNN suitable for extreme conditions.

Design and training: The DNN comprises two diffractive layers engraved on opposite sides of a single-crystal quartz wafer (thickness 500 µm), creating a fixed interlayer spacing of 500 µm. Each diffractive layer implements binary phase modulation (0 and approximately π/2 at 532 nm) via pixel height encoding. Neurons: 1024 × 1024 neurons per layer with 8 µm × 8 µm pixel size (>1 million neurons per layer). Due to the fixed spacing, the network is not fully connected; one neuron in layer 1 couples to roughly a 7 × 7 neighborhood in layer 2 via zero-order diffraction. Optical propagation model and training: Angular spectrum diffraction is used for forward modeling and error backpropagation. Training is implemented in TensorFlow 2.0. The MNIST dataset is used; for practical training time and alignment considerations, a training subset of 1000 images is selected (full-dataset training >50 h and yields higher spatial frequency phase masks that complicate alignment). Phase values are trained and then binarized to match fabrication. Output detection consists of 10 predefined regions corresponding to digits 0–9. The amplitude field in these regions is optimized toward a Gaussian distribution to concentrate intensity, allowing shorter camera exposure and lower laser power. Loss-function modifications to suppress zero-order images were considered but rejected due to accuracy trade-offs. Fabrication: Double-sided photolithography with plasma dry etching on commercial single-crystal quartz. Photolithography used a SUSS MA6 UV aligner with double-sided alignment (overlay accuracy ~1–2 µm). Etching used SENTECH ICP with SF6. Achieved etch depth ~284 nm, corresponding to near π/2 phase shift at 532 nm for n ≈ 1.46. SEM confirms pattern fidelity; profilometry and AFM provide height and roughness characterization. Experimental setup: A 532 nm laser passes through a half-wave plate and polarized beam splitter for power control, then a 4f beam expander (lenses L1, L2). A phase-only SLM with double Fourier transform and a spatial filter generates amplitude-only MNIST digit inputs by selecting the first diffraction order while blocking 0th and −1st orders. The input plane is ~5 cm before the DNN chip; the output plane is ~16.4 cm after. A CMOS camera records the output intensity. Alignment between the input layer and DNN chip is performed by simultaneously observing the unmodulated digit image (indicating input plane) and the 10 diffractive output spots (indicating DNN output) on the CMOS and adjusting their relative positions. Robustness simulations: The effects of (i) interlayer overlay errors (1–2 pixels), (ii) substrate thickness deviations (±50 µm), (iii) incident wavelength shifts (affecting phase and numerical aperture), and (iv) input-to-chip misalignment were quantified. Wavelength-induced scaling of output spot positions can be compensated by repositioning the output plane and detection regions. Lifetime testing: Bilayer DNNs trained for binary classification (digits 0 and 1) were fabricated to achieve 100% accuracy. Accelerated aging was performed at 1400 °C in air for 2 h in a box furnace. Surface roughness increased from ~0.97 nm to ~20.1 nm, but recognition accuracy for the 2-class task remained 100% due to intensity integration within target regions. Annealing for 3 h caused catastrophic damage from Al2O3–SiO2 reactions with the boat. Lifetimes were extrapolated via Arrhenius analysis from high-temperature data to room temperature.

- Device and architecture: A compact bilayer diffractive neural network chip integrated on both sides of a 500 µm-thick single-crystal quartz wafer, achieving >1,000,000 neurons per layer (1024 × 1024; 8 µm pitch) with binary phase modulation (~0 and π/2 at 532 nm). Etch depth ~284 nm realized the intended phase shift. - MNIST performance: Simulations achieved 96.1% training accuracy (loss ~0.198) on the training set and 85.4% accuracy on a 1000-image test set. Experimental inference on 50 test images achieved 82% accuracy across 10 classes. A comparable monolayer DNN reached only ~91.2% accuracy (loss ~0.437) in simulation, highlighting the benefit of multiple layers even with partial connectivity. - Trade-off observation: Binary phase modulation and constraints yield visible zero-order (unmodulated) digit images at the output; suppressing this via modified losses reduced accuracy, indicating a diffraction-efficiency vs accuracy trade-off. - Connectivity effect: Due to the 500 µm layer spacing, neurons are not fully connected; approximate 7 × 7 target coupling in the second layer via zero-order diffraction. Simulations show increasing layer number, spacing, or reducing neuron size improves accuracy by increasing degrees of freedom. - Robustness: Overlay misalignments of 1–2 pixels (8–16 µm) have minimal accuracy impact. Substrate thickness errors up to ±50 µm (~10%) reduce accuracy by ~2%. Wavelength shifts significantly affect accuracy via NA-induced magnification/demagnification but can be mitigated by re-aligning the output plane and detection regions. Input-plane misalignment with respect to the chip degrades accuracy; binary phase helps reduce alignment difficulty. - Generalization to other tasks (simulation): Fashion-MNIST achieved ~92.2% training accuracy and ~80.1% test accuracy. Phase imaging task successfully converted phase-encoded inputs to intensity outputs, demonstrating non-recognition functionality. - High-temperature stability and lifetime: After 2 h at 1400 °C, surface roughness increased (from ~0.97 nm to ~20.1 nm) but the 2-class digit recognizer maintained 100% accuracy for 50 images; samples annealed for 3 h were destroyed. Arrhenius extrapolation indicates an ultralong room-temperature lifetime, reported as τ ≈ 1.84 × 10^35 years at 300 K, with very long projected lifetimes even at elevated temperatures (e.g., 500 K). Simulations indicate performance can tolerate up to ~20% area damage without significant degradation.

The bilayer quartz-integrated DNN demonstrates that true 3D wave-based optical inference can be realized on a compact, robust platform with long projected service life, addressing key challenges for AI hardware in extreme environments. The chip’s robustness stems from stable material properties (quartz), fixed interlayer spacing inside the substrate, and high-fidelity double-sided lithography. The observed advantages over a monolayer architecture confirm that even partially connected multi-layer diffractive networks offer improved accuracy by increasing the effective degrees of freedom. Robustness analyses quantify tolerances to overlay, thickness, and alignment errors, and provide guidance for test-time calibration (e.g., output plane/detection region repositioning for wavelength shifts). The device’s adaptability to recognition and non-recognition tasks (e.g., phase imaging) suggests broader applicability. The main performance constraints are limited interlayer connectivity (from fixed thickness and pixel size), the small number of diffractive layers, and lack of nonlinear activation. The authors discuss feasible paths to higher performance: reducing pixel size, increasing interlayer spacing or number of layers via bonding techniques (e.g., laser bonding of quartz), and incorporating nonlinear absorbing layers (e.g., 2D materials, perovskites) between diffractive layers to realize deeper networks with activation. System-level integration with sources (VCSEL arrays), detectors, and electronic neural processors is proposed to realize fully integrated optoelectronic inference systems for extreme environments.

The work introduces a compact, robust, bilayer diffractive neural network chip fabricated on both sides of a quartz wafer, enabling 3D optical inference at visible wavelengths with over a million neurons per layer. The chip achieves up to 82% experimental accuracy on 10-class MNIST and shows strong robustness to fabrication and alignment errors. Accelerated aging tests and Arrhenius extrapolation indicate an ultralong projected lifetime at room temperature, supporting deployment in extreme environments. Simulations further demonstrate task versatility (Fashion-MNIST classification and phase imaging). Future directions include enhancing connectivity (smaller pixels, optimized spacing), increasing the number of layers via bonding, introducing nonlinear materials to implement optical activation for deeper networks, and integrating sources/detectors to build fully integrated 3D diffractive systems.

- Partial interlayer connectivity due to fixed 500 µm wafer thickness and 8 µm pixel size limits performance; increasing spacing or reducing pixel size would help but requires advanced fabrication. - Only two diffractive layers are implemented; additional layers are expected to improve accuracy but add fabrication complexity. - No nonlinear activation between diffractive layers, which limits performance relative to deep electronic neural networks; integrating nonlinear materials is proposed but not demonstrated. - Sensitivity to incident wavelength necessitates calibration of output plane and detection regions for off-design operation. - Experimental evaluation used a relatively small test set (50 images) for the 10-class task; full-dataset training is time-consuming and increases spatial frequency content, complicating alignment. - High-temperature anneal beyond 2 h at 1400 °C caused catastrophic damage due to reactions with the Al2O3 boat, indicating processing constraints for extreme thermal exposures.