Diffractive Meta-Neural Networks (DMNNs)

• Diffractive Meta-Neural Networks (DMNNs) are optical neural systems that compute via light diffraction using trainable metasurface layers.
• They leverage diverse architectures including single-layer multifunctional designs, multilayer cascades, and channel multiplexing for task-specific optical processing.
• DMNN training integrates fabrication-, coherence-, and robustness-aware strategies to bridge the gap between simulation and practical hardware deployment.

Diffractive Meta-Neural Networks (DMNNs) are diffractive optical neural systems in which the trainable optical layers are realized by metasurfaces or closely related meta-optical structures, so that computation is carried by wave propagation, interference, and detector readout rather than electronic multiply-accumulate operations. The terminology is not fully standardized: related papers use “diffractive neural network,” “D2NN,” “MDNN,” “A-DNN,” “PDNN,” or “metasurface-enabled diffractive neural network,” but they describe a common family of physically parameterized, trainable wave processors whose “weights” are embedded in diffractive or metasurface layers (Luo et al., 2021, Tian et al., 23 Jun 2025, Behroozinia et al., 2024). In current literature, this family spans visible, terahertz, and broadband implementations; multifunctional single-layer metasurfaces; multilayer cascades; polarization- and wavelength-multiplexed processors; and adjacent acoustic and planar-RF analogues that extend the same design logic to other wave domains (He et al., 15 Jun 2026, Luo et al., 2019, Weng et al., 2019, Teng et al., 30 Nov 2025).

1. Concept, scope, and terminology

A DMNN is best understood as a physically instantiated neural architecture in which each optical layer is a spatial array of trainable modulation elements and the inter-layer connectivity is created by diffraction. In the broader diffractive-neural literature, a standard diffractive deep neural network consists of cascaded diffractive layers, often phase-only, trained by backpropagation; metasurface-based variants replace larger diffractive pixels with subwavelength meta-atoms or meta-unit cells, thereby adding compactness, higher areal density, and access to polarization or dispersion engineering (Luo et al., 2019, Luo et al., 2021). Some papers reserve their own acronyms: the visible on-chip multitask classifier is explicitly called an “MDNN” rather than a DMNN (Luo et al., 2021), the layer-permutable system is called an “A-DNN” (Tian et al., 23 Jun 2025), and the laterally composable architecture is called a “PDNN” (Tian et al., 25 Jan 2026). This suggests that “DMNN” functions mainly as an umbrella label rather than a universally adopted formal name.

The architectural scope is correspondingly broad. A DMNN may be a single multifunctional metasurface, as in the spin-multiplexed edge-enhanced classifier (He et al., 15 Jun 2026); a bilayer cascaded metasurface supporting multi-task inference through polarization or wavelength channels (Behroozinia et al., 2024); a multilayer broadband diffractive processor that performs deterministic spectral filtering or wavelength de-multiplexing (Luo et al., 2019); or a hybrid system in which a metasurface convolutional front end feeds a diffractive decoder (Liang et al., 5 Dec 2025). Related work further extends the same principles to acoustic metamaterial “meta-neurons” (Weng et al., 2019) and planar RF diffractive networks based on coupled transmission-line layers (Teng et al., 30 Nov 2025). A plausible implication is that the defining feature of DMNNs is less the fabrication platform than the combination of trainable wavefront modulation, physical propagation, and task-specific optical inference.

2. Physical and mathematical foundations

The common forward model is a layered propagation operator in which each neuron or pixel applies a local complex transmission and free-space diffraction couples that modulation to downstream layers. A representative formulation writes the local transmission of neuron $i$ on layer $l$ as

$$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$

with wavelength-dependent amplitude $a_i^{\,l}$ and phase $\phi_i^{\,l}$, and models propagation between layers with the Rayleigh–Sommerfeld kernel (Luo et al., 2019). In a related scalar form, the secondary wave emitted by a neuron is

$$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$

with

$r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$

and the field at the output plane is read through the intensity

$I^{M+1} = \left|U^{M+1}(x_{M+1},y_{M+1})\right|^2.$

These expressions make explicit that DMNNs are linear in complex field propagation before detection, while the detector introduces square-law readout (Tian et al., 23 Jun 2025).

Metasurface implementations enrich this baseline with channel-selective local physics. The visible on-chip MDNN employs birefringent rectangular TiO$_2$ nanopillars whose Jones response depends on incident polarization (Luo et al., 2021). The polarized OAM network uses rectangular micro-structure meta-material elements with two local phase responses and a rotation angle, thereby replacing scalar modulation with local anisotropic vector-field processing (Zhang et al., 2022). The edge-enhanced single-layer classifier adds a second, nonlocal operator: the co-polarized channel of a nonlocal Huygens’ metasurface performs momentum-space filtering for real-time edge detection, while the cross-polarized channel supplies Pancharatnam–Berry phase modulation for classification (He et al., 15 Jun 2026). That pairing of local trainable phase control and nonlocal spatial-frequency filtering is one of the clearest departures from earlier scalar diffractive-mask models.

Coherence conditions further modify the effective forward model. Under partial spatial or temporal coherence, neither the fully coherent field model nor the fully incoherent intensity model is generally adequate. The coherence-aware framework therefore starts from the mutual coherence function

$$\Gamma(x_1,y_1,x_2,y_2,\tau) = \left\langle E(x_1,y_1,t+\tau)E^*(x_2,y_2,t) \right\rangle$$

and the complex degree of coherence

$l$0

then trains the network by summing intensities over source points and wavelengths (Kleiner et al., 2024). The paper shows that when the spatial coherence length on the object is comparable to the minimum preserved feature size, coherent and incoherent approximations can both fail. This directly affects DMNN deployment in active-illumination settings such as reflected-light microscopy, autonomous vehicles, and smartphones (Kleiner et al., 2024).

3. Architectural patterns and multiplexing strategies

One major branch of DMNN development uses metasurfaces to compress the optical stack into ultrathin, channel-multiplexed hardware. The visible on-chip MDNN integrates a polarization-multiplexed metasurface directly with a Sony IMX686 CMOS sensor, operates at $l$1, uses a $l$2 period and $l$3 TiO$l$4 nanopillars, and reaches an artificial-neuron areal density of $l$5 per channel (Luo et al., 2021). The same physical neuron array supports two tasks through orthogonal linear polarizations, with the output plane partitioned into detector regions. The tri-channel bilayer metasurface processor extends this idea from polarization to wavelength multiplexing, using 450, 550, and 650 nm channels and two cascaded metasurfaces separated by $l$6, with each optical neuron defined as a $l$7 supercell of TiO$l$8 nanofins (Behroozinia et al., 2024).

A second branch emphasizes physical reconfigurability without active pixel tuning. The arrangeable DNN realizes task switching by physically permuting the order of pre-trained metasurface layers: configuration $l$9 performs one task and $$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$0 another, exploiting the noncommutativity of cascaded diffractive operators (Tian et al., 23 Jun 2025). The partitionable DNN instead uses horizontal composition: a single diffractive aperture is partitioned into four subnetworks, each quadrant functioning independently under selective illumination, while the full aperture forms an additional task-specific network when all quadrants are active (Tian et al., 25 Jan 2026). These systems suggest that “reconfigurability” in DMNNs need not mean pixel-level programmability; it may instead arise from layer ordering, selective activation, or modular assembly.

A third pattern is multifunctional single-layer design. The edge-enhanced metasurface classifier receives right circularly polarized light, emits a co-polarized RCP output for edge detection and a cross-polarized LCP output for classification, and uses coupled quasi-bound states in the continuum and magnetic dipole resonances in crescent-shaped silicon nanopillars to raise polarization conversion efficiency to approximately $$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$1 (He et al., 15 Jun 2026). Because the LCP phase varies linearly over $$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$2 with nanopillar rotation while the RCP phase remains nearly flat, the classification and preprocessing branches are decoupled in the same patterned structure. This suggests a route to multi-stage optical processing without multilayer alignment overhead.

4. Training, inference, and hardware-aware co-design

Most DMNNs are trained digitally with a differentiable wave-propagation model and then instantiated physically. The broadband THz framework explicitly learns the thickness profiles of three transmissive diffractive layers over $$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$3–$$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$4 THz by optimizing wavelength-dependent complex transmission, using $$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$5 sampled frequencies, random batches of $$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$6, 200 epochs, and Adam with learning rate $$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$7 (Luo et al., 2019). The adaptive two-layer multifocal design likewise learns physical thicknesses $$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$8 through the bounded parameterization

$$t_i^{\,l}(x_i,y_i,z_i,\lambda)=a_i^{\,l}(x_i,y_i,z_i,\lambda)\exp\!\big(j\phi_i^{\,l}(x_i,y_i,z_i,\lambda)\big),$$9

then optimizes multi-wavelength focusing efficiencies with adaptive weights that evolve according to the residual task errors (Chen et al., 2022). In both cases, dispersive material response is treated as part of the computational resource rather than a nuisance.

Metasurface DMNNs add a second design stage: mapping idealized optical coefficients to physically realizable meta-atoms. A library-based strategy first trains ideal phase maps, then assigns each neuron a discrete meta-atom whose multi-channel response best matches the desired phases. The three-task wavelength-multiplexed bilayer metasurface uses a library of rectangular TiO$a_i^{\,l}$0 nanofins with $a_i^{\,l}$1 sampled in 5 nm increments, excludes cells with transmission below 0.5, and chooses the geometry minimizing a weighted sum of phase errors across tasks (Behroozinia et al., 2024). The same paper introduces an end-to-end alternative in which three surrogate ANNs map $a_i^{\,l}$2 directly to transmittance, $a_i^{\,l}$3, and $a_i^{\,l}$4 at 450, 550, and 650 nm, and the structural parameters are optimized jointly with the optical loss using

$a_i^{\,l}$5

with $a_i^{\,l}$6, $a_i^{\,l}$7, and $a_i^{\,l}$8 (Behroozinia et al., 2024). This directly addresses the model-to-hardware gap created by post hoc library quantization.

Robustness-aware training has become equally central. For transverse-shift tolerance, the robust visible-wavelength DNN defines a shifted transmission

$a_i^{\,l}$9

treats the loss as a random function of the DOE shifts, and minimizes its expectation

$\phi_i^{\,l}$0

through Monte Carlo gradient estimates over random displacements (Soshnikov et al., 2024). For coherence uncertainty, the coherence-aware framework trains either a nonblind network for one specified $\phi_i^{\,l}$1 pair or a coherence-blind network by randomly choosing a coherence condition for each training batch (Kleiner et al., 2024). Output readout can also be optimized rather than fixed: in mode sorting, detector masks $\phi_i^{\,l}$2 are included in the trainable parameter set, improving the efficiency–crosstalk tradeoff relative to fixed-output-region design (Bearne et al., 27 Aug 2025). Taken together, these methods indicate that modern DMNN training increasingly treats illumination statistics, fabrication constraints, detector geometry, and layer misalignment as first-class optimization variables.

5. Representative functions and reported performance

The demonstrated functions of DMNNs now extend well beyond single-task monochromatic classification. They include on-chip visible multitask recognition (Luo et al., 2021), wavelength-selective multi-task inference (Behroozinia et al., 2024), layer-reorderable or partitionable task switching (Tian et al., 23 Jun 2025, Tian et al., 25 Jan 2026), broadband spectral filtering and wavelength routing (Luo et al., 2019), multifocal and spectrally tailored flat optics (Chen et al., 2022), vector-beam classification and OAM multiplexing (Zhang et al., 2022), optical mode sorting with trainable detector regions (Bearne et al., 27 Aug 2025), all-optical imaging through random diffusers (Li et al., 2022), beam shaping for manufacturing (Jacob et al., 17 Sep 2025), and nonlinear diffractive inference with second-harmonic generation (Braasch et al., 26 Mar 2026). A plausible implication is that the field is shifting from “optical classifier” demonstrations toward a broader notion of task-specific meta-optical computing.

Work	Configuration	Reported result
Edge-enhanced single-layer metasurface DMNN (He et al., 15 Jun 2026)	Spin-multiplexed edge detection + classification on MNIST	Accuracy increased from $\phi_i^{\,l}$3 to $\phi_i^{\,l}$4; polarization conversion efficiency approximately $\phi_i^{\,l}$5
On-chip visible MDNN (Luo et al., 2021)	Polarization-multiplexed metasurface at 532 nm integrated with CMOS	Both networks achieved $\phi_i^{\,l}$6 accuracy with two hidden layers in simulation; experimental dual-target dual-channel results showed $\phi_i^{\,l}$7 match with simulation
Arrangeable metasurface A-DNN (Tian et al., 23 Jun 2025)	Two tasks by layer permutation $\phi_i^{\,l}$8 and $\phi_i^{\,l}$9	Test accuracies $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$0 on MNIST and $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$1 on FashionMNIST for the shared A-DNN; 50% hardware-efficiency improvement relative to separate D2NNs
Multiplexed bilayer metasurface WM-DNN (Behroozinia et al., 2024)	Three tasks at 450/550/650 nm	Library design maintained $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$2 accuracy for all tasks; end-to-end redesign improved FMNIST by 4.7% and KMNIST by 4.2%
MAODCNN hybrid metasurface architecture (Liang et al., 5 Dec 2025)	Optical convolution layer + diffractive metasurface decoder	MAODCNN reached $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$3 versus $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$4 on MNIST and $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$5 versus $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$6 on Fashion-MNIST in the reported comparison

A separate line of evidence concerns robustness rather than raw accuracy. Under random diffusers, a four-layer diffractive network trained at $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$7 achieved representative test PCC values around $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$8, while deeper systems improved to about $$w_i^l(x,y,z)=\frac{z-z_i}{r_i^2}\left(\frac{1}{2\pi r_i}+\frac{1}{j\lambda}\right)\exp\!\left(\frac{j2\pi r_i}{\lambda}\right),$$9 for five layers (Li et al., 2022). Under transverse layer shifts, a robust two-DOE visible DNN preserved $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$0 accuracy and $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$1 minimum contrast when both DOEs were shifted by two pixels, whereas the non-robust counterpart dropped to $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$2 accuracy and $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$3 contrast under the same perturbation (Soshnikov et al., 2024). For partial coherence, nonblind two-layer MNIST classifiers ranged from $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$4 to $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$5, while coherence-blind classifiers ranged from $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$6 to $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$7, showing that illumination statistics can be as consequential as optical-layer design (Kleiner et al., 2024).

6. Limitations, misconceptions, and research directions

Several limitations recur across the literature. First, many of the most technically ambitious DMNN results remain simulation-based. The edge-enhanced single-layer metasurface reports simulated classification and simulated edge images rather than fabricated-device measurements (He et al., 15 Jun 2026); the metasurface all-optical diffractive CNN is also numerically validated (Liang et al., 5 Dec 2025); and the three-task multiplexed bilayer metasurface remains a numerical study despite its fabrication-aware co-design (Behroozinia et al., 2024). Where experiments exist, simulation–hardware gaps remain substantial: the arrangeable metasurface A-DNN reports five-class experimental accuracies of $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$8 for digits and $r_i=\sqrt{(x-x_i)^2+(y-y_i)^2+(z-z_i)^2},$9 for fashions versus simulated $I^{M+1} = \left|U^{M+1}(x_{M+1},y_{M+1})\right|^2.$0 and $I^{M+1} = \left|U^{M+1}(x_{M+1},y_{M+1})\right|^2.$1, attributing much of the gap to forward-design errors in metasurface realization (Tian et al., 23 Jun 2025).

Second, single-layer compactness does not remove the representational bottleneck. The edge-enhanced classifier improves MNIST accuracy from $I^{M+1} = \left|U^{M+1}(x_{M+1},y_{M+1})\right|^2.$2 to $I^{M+1} = \left|U^{M+1}(x_{M+1},y_{M+1})\right|^2.$3, but the paper explicitly notes that the task is still MNIST and that $I^{M+1} = \left|U^{M+1}(x_{M+1},y_{M+1})\right|^2.$4 remains far below electronic deep networks (He et al., 15 Jun 2026). This bears on a common misconception: more compact all-optical inference does not automatically imply competitive benchmark performance. A related misconception concerns optical depth. The SHG study emphasizes that passive linear diffractive layers remain linear in the field, so one undepleted second-harmonic stage can enhance classification accuracy and class contrast, but “a single undepleted SHG layer does not by itself guarantee universal approximation capability” (Braasch et al., 26 Mar 2026). The placement of the nonlinear layer is itself critical.

Third, real deployment is governed by nonideal physics beyond nominal forward propagation. Coherence mismatch can reduce a nonblind classifier to $I^{M+1} = \left|U^{M+1}(x_{M+1},y_{M+1})\right|^2.$5 accuracy under strong spatial-coherence mismatch (Kleiner et al., 2024). Alignment errors can destroy nominally trained multilayer systems unless expected shifts are built into the training objective (Soshnikov et al., 2024). Resonant metasurfaces are often wavelength-specific, and several of the highest-performing channel-multiplexed designs rely on carefully engineered anisotropy, q-BIC/MDR overlap, or surrogate-modeled geometry-to-response mappings (He et al., 15 Jun 2026, Behroozinia et al., 2024). This suggests that practical DMNN design is increasingly inseparable from fabrication-aware, channel-aware, and illumination-aware optimization.

Current research directions therefore converge on a few themes: deeper yet still alignable meta-optical stacks; multifunctionality through polarization, wavelength, spin, or module arrangement; direct geometry optimization rather than post hoc library matching; optical preprocessing integrated into the same meta-neural substrate; robustness to coherence and assembly uncertainty; and experimentally viable nonlinear layers such as SHG or other metasurface-compatible mechanisms (Behroozinia et al., 2024, Braasch et al., 26 Mar 2026). A plausible implication is that the most consequential advances in DMNNs may come less from isolated improvements in phase-mask design than from architectural co-design across sources, channels, metasurface unit cells, propagation geometry, and detector readout.