Multimode Opto-Electronic Neural Network

• MOENN is a neural computing architecture that uses multiple orthogonal optical modes combined with electronic control to achieve high-speed, energy-efficient operations.
• It employs diverse platforms such as multimode VCSEL arrays, silicon photonic chips, multimode fibers, and phase-change metasurfaces for dense matrix–vector processing.
• The approach leverages nonlinear dynamics and programmable activations to enable scalable, edge-AI applications while mitigating conventional scaling challenges.

A Multimode Opto-Electronic Neural Network (MOENN) utilizes the parallelism and high-bandwidth signal processing capacities of multimodal optical platforms—such as multimode lasers, fibers, or waveguides—combined with electronic control and readout to implement neural network operations at speeds and energy efficiencies unattainable with conventional purely electronic or single-mode photonic systems. MOENNs leverage multiple orthogonal optical modes (spatial, polarization, or temporal) as information channels, enabling large matrix–vector operations, efficient mode-multiplexed neural layers, and robust nonlinear activations, while mitigating the scaling and crosstalk limits of wavelength-division multiplexed photonic neural networks.

1. Physical and Device Architectures

MOENNs are realized in multiple physical platforms, including multimode VCSEL lasers, integrated silicon photonic chips, multimode fibers, and phase-change metasurface arrays.

• Multimode VCSEL arrays: Each spatial pixel in the near-field emission of a large-area vertical-cavity surface-emitting laser (LA-VCSEL) acts as a “photonic neuron” (Porte et al., 2020). Boolean or analog input patterns are encoded optically and mapped via a multimode fiber to the VCSEL, whose spatially distributed modes implement dense local and nonlocal recurrent coupling.
• Integrated silicon photonic MOENN: Orthogonal waveguide eigenmodes (e.g., TE₀, TE₁, TE₂) in a silicon bus waveguide, coupled in and out via asymmetric directional couplers and mode-selective weighting/attenuation, realize truly parallel, channel-resolved operations at a single wavelength (Xiang et al., 22 Jan 2026). PIN attenuators set weights, and a multimode Ge photodetector integrates weighted signals before optical nonlinear activation via a microring resonator.
• Multimode fiber MOENNs: High-dimensional, nonlinear optical transforms are performed by the propagation of modulated ultrashort pulses or picosecond pulses through graded-index multimode fibers (GRIN MMF) (Kesgin et al., 2024, Oguz et al., 2022). Nonlinear self-phase modulation, spatio-temporal chaos, and modal mixing underpin analog reservoir computation.
• Phase-change metasurface MOENN: Arrays of programmable Ge-Sb-Te (GST) metasurfaces atop multimode Si₃N₄ waveguides function as locally addressable mode converters, assigning analog-valued weights by adjusting material crystallinity (Wu et al., 2020). Downstream on-chip demultiplexers and photodiodes implement readout and summation, with electronic control over the weight programming.

Key device components for scalable MOENNs include:

• Programmable mode-division fan-in/-out units (ADCs, multimode splitters),
• Mode-selective optical attenuators,
• Ultra-compact mode-converting bends,
• Thermo-optically tunable microrings,
• Nonvolatile phase-change memory arrays for in-memory analog weight storage,
• Hybrid optical/electronic nonlinear activation modules.

2. Mathematical Models and Computational Principles

The principal mathematical structure common to MOENNs is a high-dimensional state or mode expansion, with linear optical mixing/weighting and (optionally) nonlinear transformation:

• State evolution: For systems like the multimode VCSEL, the photonic state vector $x(t)\in\mathbb{R}^n$ evolves under the coupled rate equation

$$\tau\,\frac{dx}{dt} = -x(t) + f\left(W_{\text{int}} x(t) + W_{\text{in}} u(t)\right),$$

where $W_{\text{in}}$ is the input weight matrix (often imposed by a complex transmission matrix or mode multiplexer), $W_{\text{int}}$ is a fixed recurrent coupling (arising from diffraction and carrier diffusion), and $f$ is a nonlinear gain–saturation function (Porte et al., 2020).

• Weighted matrix–vector product in mode space: In integrated platforms, the input vector is mapped onto $M$ modes and $N$ WDM channels, yielding an extended input $x_{m,i}$ and weight tensor $w_{k,(m,i)}$. The output neuron $k$ calculates

$$y_k = \sum_{m=0}^{M-1} \sum_{i=1}^N w_{k,(m,i)}\,x_{m,i},$$

in a single optical pass through a weight bank (Jia et al., 2024).

• Nonlinear transformations: Implemented via optical Kerr effect in multimode fibers (mode-multiplicative nonlinear Schrödinger evolution), or electronic/thermo-optic tuning in microring resonators for differentiable activations (e.g., ReLU, sigmoid) (Xiang et al., 22 Jan 2026, Oguz et al., 2022).
• Readout: Optical or opto-electronic. In most platforms, a linear or binary-weighted summation across all modes, spatial pixels, or photodiodes, forms the output y; for some, differences in mode powers (e.g., $P_{\text{TE}_0} - P_{\text{TE}_1}$) directly yield the matrix–vector multiply components (Wu et al., 2020).

3. Training Methods and Programming

MOENNs are typically trained using approaches adapted to the available degrees of freedom and hardware constraints:

• Reservoir computing: Fixed optical transform; only electronic or optical readout weights (e.g., DMD mirror states, photodiode weights) are trained. Closed-form ridge regression or coordinate-descent "bit-flip" optimization is used for discrete-readout platforms (Porte et al., 2020).
• Programming physical parameters: In MMF-based and phase-change systems, small vectors of analog programming parameters (e.g., SLM phase masks, GST crystallinity levels) are optimized in situ using black-box (e.g., genetic algorithm, surrogate-based) optimization, evaluating network-level accuracy for each setting (Oguz et al., 2022, Xiang et al., 22 Jan 2026).
• On-chip in-memory learning: For phase-change MOENNs, the nonvolatile weight states can be repeatedly updated at low energy for on-chip (co-packaged) learning using electronic backpropagation (Wu et al., 2020).
• Activation tuning: Activation function forms (sigmoid, ReLU, radial-basis) in integrated MOENNs are programmed by setting current and heater values on microring modulators, often via direct optimization of task accuracy (Xiang et al., 22 Jan 2026).

4. Performance Metrics and Experimental Demonstrations

MOENNs have demonstrated benchmark-level performance in both conventional and specialized tasks. Typical metrics and results include:

Task/Metric	Platform	Value/Result
2-bit header recognition	VCSEL MOENN (Porte et al., 2020)	SER < 0.9×10⁻³
2-bit XOR	VCSEL MOENN (Porte et al., 2020)	SER ≈2.9×10⁻²
MNIST (1 vs 2)	GST-PMM MOENN (Wu et al., 2020)	91% (experiment) vs 90% (simulation); >88% single-mode
Iris flower classification	On-chip MOENN (Xiang et al., 22 Jan 2026)	92.1% test accuracy; converges in ~80 GA generations
ECG-based emotion classification	On-chip MOENN (Xiang et al., 22 Jan 2026)	90.7% test accuracy (700 samples, 1D CNN with mode-multiplexed kernels)
Fashion-MNIST classification	MMF MOENN (Oguz et al., 2022)	77–79% with 99.5% parameter reduction vs digital; digital LeNet-5: 79.1%
EuroSAT (geospatial imagery)	MMF chaos MOENN (Kesgin et al., 2024)	84.61% (with chaos), baseline 26.46%, ResNet-18: 80.1%
Energy per MAC/operation	MDM MOENN (Jia et al., 2024)	≈50 fJ optical (excluding electronics), up to 0.5 pJ/FLOP (high-speed DMD MMF)
Areal computational density	GST-PMM MOENN (Wu et al., 2020)	>25 TOPS/mm² (4 WDM lines)

Bandwidths exceeding 20 GHz per mode, sub-pJ/operation energy use, and large fan-in/fan-out due to mode-division multiplexing and parallel multimode processing have been experimentally verified.

5. Nonlinear Dynamics, Chaos, and High-Dimensionality

Nonlinear multimode propagation, especially in graded-index fibers or semiconductor multimode cavities, supports enhanced data separability and capacity by exploiting spatiotemporal chaos or high-dimensional modal scattering. Important phenomena include:

• Spatiotemporal chaos in multimode fibers: Highly nonlinear regimes, induced for example by injecting orbital angular momentum phase masks or increasing pulse peak power, result in “chaotic” propagation regimes where different input classes become linearly separable only after fiber-induced randomization (Kesgin et al., 2024).
• Optimality at “edge of chaos”: Empirical performance is maximized by tuning parameters (pulse power, OAM charge, fiber length) to intermediate regimes—sufficient nonlinear mixing for class separation, but not so strong as to lose input distinction. For instance, Breast MNIST: accuracy rises from 75% (no chaos) to 83.3% (chaotic regime), matching large digital CNNs with 99.9% fewer parameters (Kesgin et al., 2024).
• Analytic and simulation models: The nonlinear Schrödinger or paraxial wave equation, with mode expansion and Kerr nonlinearity, govern the evolution. Multi-mode VCSELs and MMF channels can be simulated via split-step Fourier methods or measured directly.

6. Scalability and Integration

MOENNs provide substantial scaling advantages over both electronic and single-mode photonic platforms:

• Hardware parallelism: Node counts >130 demonstrated with LA-VCSEL (Porte et al., 2020), >1000 expected with larger apertures; up to 15 independent modes in silicon MOENNs (Xiang et al., 22 Jan 2026); mode and channel multiplicity in phase-change and MDM architectures (Jia et al., 2024, Wu et al., 2020).
• Hybrid MDM–WDM: Doubling or multiplying channel count without additional hardware by combining multiple spatial modes and multiple wavelengths (Jia et al., 2024).
• Integration: Monolithic SOI integration demonstrated for all major functions—input encoding, mode mixing, weighting, activation—with footprints down to ∼5 mm² for 3-mode MOENN and >25 TOPS/mm² in phase-change metasurface MOENN (Wu et al., 2020, Xiang et al., 22 Jan 2026).
• Co-integration challenges: Issues include thermal crosstalk (especially in densely packed microrings or PCM arrays), intermodal loss, waveguide sidewall roughness, and electrical–photonic interface parasitics.

Mitigation strategies encompass aggressive inverse-design, co-packaged electronic control loops, and thermal isolation.

7. Prospects, Limitations, and Future Directions

MOENNs represent a promising class of opto-electronic and photonic neural computing hardware with competitive accuracy, scalable channel count, and favorable energy efficiency. Outstanding areas include:

• Bandwidth increases: Current electronic and optical modulators limit speed; progressing to GHz-scale or all-optical control will enable ultrahigh-throughput operation.
• Scaling to deep architectures: Cascaded fiber layers or stacked on-chip waveguide cells enable multilayer MOENNs, but demand robust optical weight update and calibration schemes (Oguz et al., 2022).
• In-memory photonic inference: Phase-change MOENNs and monolithic platforms eliminate DRAM fetches and compute entire matrix–vector products or convolutions in situ, reducing energy and latency (Wu et al., 2020).
• Hardware-in-the-loop and in-situ training: Genetic, surrogate-based, or backprop-assisted training adapted to physical device constraints and noise profiles (Xiang et al., 22 Jan 2026, Oguz et al., 2022).
• Application scope: Demonstrated on biomedical, geospatial, and general pattern-recognition benchmarks; high energy efficiency and parallelism position MOENNs for edge-AI, scientific sensing, and embedded photonic intelligence.

Scaling MOENNs to large neural nets will require high-order mode control (M > 4), integrated multi-line laser sources or microcombs for WDM, advanced fabrication for low cross-talk, and fully on-chip nonlinear normalization and memory. Integration of III–V microlasers and CMOS learning engines with multimode photonic tensor cores is anticipated for next-generation photonic intelligence hardware (Jia et al., 2024).

(Porte et al., 2020, Jia et al., 2024, Wu et al., 2020, Xiang et al., 22 Jan 2026, Oguz et al., 2022, Kesgin et al., 2024)