Diffractive Neural Networks

• Diffractive Neural Networks are optical systems that harness diffractive elements and phase/amplitude masks to perform complex, parallel machine learning tasks.
• They employ multilayer architectures with gradient-based backpropagation to train physical diffraction parameters for precise imaging and classification.
• Recent advances enable reconfigurable and robust designs through modular partitioning, multiplexing, and noise-resilient training for diverse optical AI applications.

Diffractive Neural Networks (DNNs) are a class of physical neural networks that employ diffractive optical elements to execute machine learning tasks by controlling and exploiting the wave propagation and diffraction of coherent light. These networks realize computational operations entirely in the optical domain, achieving ultrafast parallel processing and extremely low power consumption. DNNs encode learnable functions within phase or amplitude masks that modulate incident electromagnetic fields, with the network’s overall mapping determined by the sequential interaction of these layers through scalar diffraction. Since their introduction for applications such as all-optical image classification and lens design, DNNs have evolved to encompass reconfigurable, multifunctional, and robust architectures with increasing optical complexity and task versatility.

1. Physical Principles of Diffractive Neural Networks

Diffractive neural networks operate on the foundation of scalar diffraction theory, commonly formulated via the Rayleigh–Sommerfeld, Huygens–Fresnel, or angular-spectrum methods. A thin transmission mask with complex amplitude $t(x,y)$ modulates an input complex field $u_\mathrm{in}(x,y)$, followed by free-space propagation governed by the diffraction kernel $h$:

$$u_\mathrm{out}(x,y) = \iint h(x-x',y-y')\, t(x',y')\, u_\mathrm{in}(x',y')\, dx' dy'$$

The intensity at the detector is $I_\mathrm{out}(x,y) = |u_\mathrm{out}(x,y)|^2$. In multilayer architectures, each diffractive layer consists of a 2D grid of “neurons” or pixels, where each neuron applies a phase-only or complex modulation to the incident field. The interlayer free-space propagation—modeling all-optical “connections”—is linear and energy-preserving, implying that the full diffractive stack is (ideally) unitary (Zheng et al., 2018). Diffraction enables DNNs to realize high-dimensional linear transformations and, with the inclusion of optical nonlinearities, nonlinear mappings (Braasch et al., 26 Mar 2026, Dong et al., 18 Apr 2025).

2. Architecture and Training Methodology

A basic diffractive network comprises $M$ stacked layers, each with $N \times N$ pixels. Each pixel (neuron) at layer $l$ and position $(i,j)$ imparts a transmission term $$t_l(x_{ij}, y_{ij}) = a \exp[j\phi_l(x_{ij},y_{ij})]$$, with phase-only configurations setting $u_\mathrm{in}(x,y)$0. The network output may be an image (holographic reconstruction) or a classification decision, typically inferred via the energy deposited in pre-defined detector regions assigned to each class.

Training proceeds via gradient-based backpropagation through the entire physical propagation pipeline, utilizing differentiable convolutional operators and automatic differentiation frameworks (e.g., PyTorch). The loss is a task-dependent functional: mean square error for imaging, cross-entropy for classification. Hardware-aware regularization such as phase quantization and pixel sub-meshing can be incorporated to model fabrication limitations.

Pseudocode (high level):

$u_\mathrm{in}(x,y)$4 (Tian et al., 25 Jan 2026)

3. Modular, Multifunctional, and Reconfigurable Diffractive Networks

Recent advancements address the fixed-function nature of conventional DNNs by introducing modular and reconfigurable designs:

• Partitionable Diffractive Neural Networks (PDNNs): Each diffractive layer is partitioned laterally into $u_\mathrm{in}(x,y)$1 submodules (e.g., quadrants), each independently trainable for different functions. Horizontal stacking and selective illumination (amplitude masks) allow a single physical device to be reconfigured—without refabrication—to implement multiple distinct imaging or classification tasks. Further, rotated submodule arrangements lead to new composite functionalities. Joint training of the submodules is accomplished by optimizing a combined loss:

$u_\mathrm{in}(x,y)$2

(Tian et al., 25 Jan 2026)

• Arrangeable Multitask Networks (A-DNN): Independent metasurface layers are trained under multitask loss for several tasks and physically reordered at inference to realize different functions. This approach provides dynamic reconfiguration at low cost and improved hardware efficiency over separate networks (Tian et al., 23 Jun 2025).
• Illumination-Controlled Functionality: Task selection via angular-spectrum encoding employs amplitude masks and a $u_\mathrm{in}(x,y)$3 lens system to sculpt the illumination's angular distribution, passively switching network functionality in the paraxial regime (±0.43° numerical aperture), without mechanical reconfiguration (Kleiner et al., 8 Jan 2026). This method scales by the number of distinct angular “bins” available, and can be combined with polarization or wavelength encoding for further multiplexing.

4. Robustness, Tolerances, and Generalization

Physical realization of DNNs necessitates robustness to fabrication errors and environmental perturbations:

• Transverse and Longitudinal Misalignment: Shift-aware training introduces random displacements of diffractive elements during optimization, achieving networks that retain high accuracy up to ~17 wavelengths of misalignment (Soshnikov et al., 2024). “Vaccination” methods inject 3D misalignment errors in simulation, training DNNs whose inference accuracy remains stable across a wide range of layer-to-layer shifts (Mengu et al., 2020).
• Weight-Noise-Injection: Additive Gaussian noise to the phase weights during training drives the solution to flat minima in the loss landscape, promoting resilience against fabrication imperfections and external disturbances. This approach achieves >80% MNIST accuracy even under strong noise, outperforming conventional DNNs which collapse near random guessing in strongly perturbed regimes (Shi, 2020).
• Coherence Adaptivity: Degree of spatial and temporal coherence in illumination critically impacts diffractive network performance. “Coherence-blind” networks, trained under a range of coherence conditions, generalize gracefully to illumination drift, whereas nonblind networks can degrade rapidly under mismatch. A general training framework enables end-to-end optimization for arbitrary partial coherence (Kleiner et al., 2024).

5. Task Versatility: Multifunction, Multiplexing, and Parallelization

Diffractive neural networks are being extended to ever more complex, simultaneous task execution:

• Wavelength and Polarization Multiplexing: Cascaded metasurfaces and polarization-sensitive elements enable parallel, channel-separated processing of multiple tasks (e.g., MNIST, FashionMNIST, KMNIST) within one physical system. Dual-channel DNNs using wavelength or polarization multiplexing achieve near-single-task accuracy per channel, while tri-channel systems maintain >80% accuracy for all tasks; end-to-end joint optimization of metaatom parameters substantially narrows the remaining performance gap (Behroozinia et al., 2024).
• Broadband and Multi-Spectral Operation: Adaptive diffractive networks enable dual-band and ultra-narrowband focusing, PSF shaping, and super-oscillatory spot generation across multiple illumination bands. Gradient-based inverse design incorporates full diffraction models and multi-objective loss functions, producing efficiency surpassing single-layer limits and direct metrology compatibility (Chen et al., 2022).
• Mode Sorting and Recognition: The unitary property of DNNs allows lossless mapping of orthogonal input modes (e.g., Hermite–Gaussian, Laguerre–Gaussian) to spatially separated detector regions, with simulation and experiment confirming high efficiency and minimal crosstalk; output detection regions can themselves be made trainable for efficiency-crosstalk optimization (Zheng et al., 2018, Bearne et al., 27 Aug 2025).

6. Scaling, Nonlinearity, and Hardware Realizations

• Scalability: Device size and neuron pitch are physically limited, but nanofabrication advances (e.g., TiO₂ metasurfaces, deep lithography) enable large-scale arrays and submicron pitch layers. All-optical convolutional architectures, using physically implemented convolution kernels with diffractive decoding, approach the structure of electronic CNNs, with accuracy scaling upward with layer count (Liang et al., 5 Dec 2025).
• Nonlinear Activation: Integration of second-order nonlinearities (e.g., second-harmonic generation, SHG) and χ^3 Kerr effects enables incorporation of all-optical activation functions. Performance—classification accuracy, class contrast—is highly dependent on the placement of nonlinear layers relative to phase modulation and free-space propagation. SHG in the intermediate layers (after propagation but before detection) yields optimal accuracy and contrast. Experimentally, photonic DNNs with integrated Kerr nonlinearities (in silicon) realize classification accuracy improvements that increase with depth and complexity, moving toward functional parity with digital deep nets (Braasch et al., 26 Mar 2026, Dong et al., 18 Apr 2025).
• Reconfigurability: Dynamic task switching is supported by spatial light modulators (SLM) or through physical manipulations (rotation, stacking, or layer arrangement). Digital updating of phase masks on SLMs allows online or real-time reprogramming of the network for different inference tasks (Dong et al., 18 Apr 2025, Tian et al., 23 Jun 2025).

7. Performance Benchmarks and Applications

Quantitative benchmarks demonstrate the state-of-the-art capabilities of diffractive neural networks:

Task/System	Layers/Type	Accuracy (Dataset)	Notes
PDNN (full network)	4 submodules, THz exp.	100% (MNIST 1-vs-3), 13–16% eff.	Multifunctional, reconfigurable (Tian et al., 25 Jan 2026)
Dual-wavelength D²NN	1 layer	98.59% (MNIST), 90.40% (Fashion)	Surpasses 5-layer baseline (Wang et al., 23 Jul 2025)
Metasurface dual-channel	2 layers	97.7% (MNIST), 88.0% (FMNIST)	Simultaneous classification (Behroozinia et al., 2024)
A-DNN (metasurface)	2 layers	93.1% (MNIST), 87.2% (FMNIST)	Reconfigurable order (Tian et al., 23 Jun 2025)
SHG Single layer	SHG after mask+propagation	95.2% (MNIST), 81.6% (Fashion-MNIST)	Optimal nonlinear placement (Braasch et al., 26 Mar 2026)
Misalignment-robust	Single DOE	91% (MNIST), up to 17λ shift	After robust training (Soshnikov et al., 2024)

Diffractive neural networks have demonstrated all-optical classification, imaging, mode recognition, multi-task, and multi-modal processing—providing a path toward high-speed, low-energy, parallel, and reconfigurable optical AI hardware. Future work encompasses hybrid optoelectronic architectures, broader nonlinearity integration, on-chip photonic implementation, and scaling to complex vision and inference tasks.