Dynamic Network Cube: Diffractive Optical System

• Dynamic Network Cube is a mechanically reconfigurable 3D diffractive neural network that leverages phase-only masks for optical multiplexing.
• It employs mechanical operations such as permutation, translation, and rotation to modulate thousands of non-orthogonal optical channels.
• Deep joint optimization across cascaded phase masks ensures high channel fidelity and low crosstalk for applications like holography and OAM generation.

The Dynamic Network Cube, also referred to as the Diffractive Magic Cube Network (DMCN), is a mechanically reconfigurable, three-dimensional diffractive neural network system designed to achieve ultra-high multiplexing capacity for optical information processing. The architecture leverages phase-only modulation, mechanical operations (permutation, translation, rotation), and deep diffractive neural network (D²NN) optimization to encode thousands of non-orthogonal optical channels under a compact, low-complexity framework devoid of active electronics or biasing, with demonstrated applications in holography, optical focusing, and orbital angular momentum (OAM) beam generation (Feng et al., 2024).

1. Architectural Foundation and Physical Implementation

The DMCN architecture entails three cascaded phase-only layers, each realized as a planar phase mask of dimensions 300 × 300 pixels with a sub-millimeter pitch (12.5 µm in experiments; 400 nm simulated for metasurfaces). These layers are sequentially arranged along the optical axis, separated by adjustable distances $d_1$, $d_2$, and $d_3$, subject to a fixed total propagation distance (30 cm in simulation; several millimeters in the SLM-based experimental setup). The configuration delineates the edges of an abstract cube, channeling light from input through each phase mask in sequence and ultimately producing the modulated field in the output field-of-view (FOV).

In the laboratory realization, a reflective spatial light modulator (LCOS-SLM X13138-01) is partitioned into three independent 300 × 300 sub-regions, each encoding a learned phase mask $\Phi_1(x, y)$, $\Phi_2(x, y)$, or $\Phi_3(x, y)$. A dual-reflection optical cavity with two mirrors ensures that an incident 532 nm laser beam sequentially interrogates each phase mask through three reflections. The mechanical adjustability of mirrors and the SLM facilitates dynamic reconfiguration of $d_1$, $d_2$, $d_3$ (translation). Rotational and permutation functionalities are emulated by loading rotated phase patterns or by reassigning mask positions on the SLM sub-apertures.

2. Mechanical Reconfiguration: Permutation, Translation, and Rotation

Mechanical operations define the core channelization mechanism in DMCN by selectively altering the system transfer matrix:

$$\mathbf{A}_c = W(d_3) \cdot \mathrm{Diag}[\mathrm{Rot}(\Phi_3, \theta_3)] \cdot W(d_2) \cdot \mathrm{Diag}[\mathrm{Rot}(\Phi_2, \theta_2)] \cdot W(d_1) \cdot \mathrm{Diag}[\Phi_1]$$

where $W(d)$ is the angular-spectrum propagation operator over distance $d$, $\mathrm{Diag}[\cdot]$ represents the phase mask as a diagonal transmission matrix, and $\mathrm{Rot}(\Phi, \theta)$ denotes the 2D rotation of the phase mask by angle $\theta$.

The reconfigurable degrees of freedom are:

• Rotation: Layers $l=2,3$ are each rotated by among $n_r$ discrete angles (e.g., $$\theta \in \{0^\circ, 90^\circ, 180^\circ, 270^\circ\}$$). In the continuous form, a phase mask is rotated by transforming sampled coordinates with $R_\theta$.
• Translation: Inter-layer distances $(d_1, d_2, d_3)$ are selected from $n_t$ equally spaced values, maintaining $d_1 + d_2 + d_3 = \mathrm{const}$. Experimental channelization used 29 distinct triplet combinations at 0.5 cm increments.
• Permutation: The assignment of the three phase masks $(\Phi_1, \Phi_2, \Phi_3)$ is permuted, enabling $3! = 6$ (potentially up to 15 with partial permutations) distinct channel orderings.

By composing these independent operations, the theoretical channel count reaches up to $$N_{\text{perm}} \times N_{\text{trans}} \times N_{\text{rot}} \approx 15 \times 29 \times 16 = 4179$$. Subsets of channels are selected and jointly optimized in practice for the intended application domain.

3. Deep Diffractive Neural Network (D²NN) Joint Optimization

DMCN utilizes the D²NN framework to jointly optimize the three base phase masks for all mechanical reconfiguration states. For a plane-wave input $J(x, y) = 1$, the forward propagation for each channel $c$ is expressed as:

$$\begin{align*} u^{(0)} &= J \ u^{(l)} &= \mathcal{F}^{-1}\left\{ \mathcal{F}\{u^{(l-1)}\} \cdot H(f_x, f_y; d_l) \right\} \cdot \exp[j \cdot \Phi_l^{(c)}(x, y)], \quad l = 1, 2, 3 \ u_{\text{out}} &= u^{(3)}|_{\text{FOV}} \end{align*}$$

where $H$ is the angular-spectrum transfer function and $\Phi_l^{(c)}$ denotes the phase mask after applying rotation, permutation, and translation corresponding to channel $c$.

A single set of learnable masks $\{\Phi_1, \Phi_2, \Phi_3\}$ is updated across all channels. The joint loss function is:

$$L = \sum_{c=1}^{N} L_{\text{err}}^{(c)} + \lambda \sum_{c=1}^{N} L_{\text{eff}}^{(c)}$$

with $L_{\text{err}}$ being mean squared error (MSE) for intensity-based tasks or negative complex correlation (NCC) for OAM tasks, and $L_{\text{eff}}$ penalizing channels with transmission efficiency $\eta_c$ below threshold. All-mask joint optimization indirectly suppresses inter-channel crosstalk by selecting minimally correlated channel states. Optimization proceeds via back-propagation and Adam (learning rate 0.1, 4000 epochs), with typical training times of $\approx$15 min on an RTX 4090 GPU.

4. Multiplexing Capacity and Channel Discernibility

The capacity of the DMCN is determined by the independent combinatorics of the three mechanical operations, setting an upper bound:

$$N_{\text{channels}} = N_{\text{perm}} \times N_{\text{trans}} \times N_{\text{rot}}$$

For the described implementation, this yields 4179 channels, with subsets (e.g., $N=144$, $N=108$, $N=60$) optimized for specific modalities (holography, focusing, OAM). Channel fidelity is quantified using the Pearson correlation coefficient (PCC) for intensity-only outputs and complex correlation coefficient (CC) for OAM fields:

• PCC: $$\frac{\sum (G-\mu_G) (O-\mu_O)}{\sqrt{\sum (G-\mu_G)^2 \sum (O-\mu_O)^2}}$$
• CC: Used for OAM channel discrimination.
• Signal-to-Crosstalk Ratio (SCR): Derived from the off-diagonal elements of the CC-matrix, reported as $$SCR_{dB} = 10 \log_{10}(P_{\text{signal}} / P_{\text{crosstalk}})$$.

5. Experimental Validation and Channel Performance

The DMCN experimental setup employs a 532 nm Verdi V6 laser incident on the divided SLM, with sequential mask interrogation facilitated by dual-mirror reflection. System parameters include pixel size (12.5 µm), camera FOV (834 × 834 pixels at 3.45 µm pitch), and translation adjustment in 0.5 cm increments over a 30 cm optical path.

Key experimental results include:

Application	Channel Count	Mean Fidelity	Mean Efficiency	Other Metrics
Holography	144	PCC ≈ 0.98	≈ 5.2%	PSNR ≈ 23 dB
Single/Multi-focus	108	PCC ≈ 0.89	≈ 2.9%	FWHM ≈ 33.5 µm
OAM beam/comb	60	CC ≈ 0.99 (diag)	≈ 3.1%	Crosstalk < 1% (SCR)

No explicit crosstalk penalty is enforced during training; the mechanical configuration and joint optimization suffice to achieve high channel selectivity and low interference between optimized channels.

6. Applications and Prospective Extensions

The DMCN paradigm enables advanced architectures for ultra-high-capacity, mechanically reconfigurable three-dimensional optical networks using phase-only modulation. Absence of active electronics, scalability via mechanical operations, and compatibility with polarization/wavelength multiplexing facilitate applications in optical storage, all-optical computation, multi-mode communications, and 3D photolithography. The inherent “channel privacy”—whereby recovery of intended output requires knowledge of the precise permutation, translation, and rotation sequence—suggests plausible implications for secure holographic encryption.

Extending the cube architecture beyond three layers exponentially increases potential capacity; with $L$ layers, the channel count scales as $(3! \times N_t \times N_r)^L$. This suggests a versatile platform for future dynamic diffractive compute-and-communicate systems, readily adaptable to evolving demands in photonic information science (Feng et al., 2024).