MOENN: Integrated Multimode Opto-Electronic Neural Network

• MOENN is an integrated opto-electronic neural network that uses spatially orthogonal TE modes on a silicon-on-insulator platform for high-density computation.
• It features monolithic integration of input encoding, mode fan-in/out, programmable weighting, and nonlinear activation for reconfigurable operations.
• The architecture employs a gradient-free genetic algorithm for training, overcoming challenges of conventional systems and enabling scalable, energy-efficient performance.

The Monolithically Integrated Multimode Opto-electronic Neural Network (MOENN) is an on-chip computing architecture that leverages the spatial orthogonality of photonic waveguide eigenmodes to realize robust, energy-efficient neural networks. Distinct from prior wavelength- or phase-sensitive systems, MOENN utilizes guided-wave transverse electric (TE) modes as independent information channels, achieving high-density, single-wavelength operation immune to spectral crosstalk and phase noise. All functional blocks, including input encoding, mode-division fan-in/fan-out, weighting, and nonlinear multimode activation, are monolithically integrated on a silicon-on-insulator (SOI) platform. The network is trained in situ using a gradient-free genetic algorithm, enabling high classification performance on several benchmarks and supporting reconfigurable topologies, such as @@@@1@@@@ (Xiang et al., 22 Jan 2026).

1. Underlying Physical Principles

MOENN departs from conventional opto-electronic architectures by encoding each neuron’s signal into distinct, rigorously orthogonal TE modes of a multimode bus waveguide. In contrast, wavelength-division multiplexing (WDM) requires multiple narrow-band lasers and microring-based weight banks, introducing complexity and thermal management issues. Coherent space-division multiplexing (SDM) demands precise phase control and calibration. By employing modal-division multiplexing (MDM), MOENN enables all network channels to share a single optical carrier (λ₀), and since detection is intensity-based, the system is intrinsically phase-noise tolerant (Xiang et al., 22 Jan 2026).

2. Chip Architecture and Monolithic Integration

The MOENN chip integrates four functional blocks:

• Input Encoders: Mach–Zehnder modulators (MZMs) (>15 GHz bandwidth) map electrical feature vectors onto orthogonal TEM modes using push–pull driving.
• Mode-Division Fan-In/Fan-Out: Cascaded asymmetric directional couplers (ADCs) multiplex modes onto a common bus, split them into branches, and recover each mode for weighting; measured insertion losses are <0.5 dB (TE1), <0.8 dB (TE2), <1 dB (TE3), and modal crosstalk is <–12 dB from 1510–1555 nm.
• Programmable Weighting: P–i–N diode optical attenuators provide intensity modulation from 0 dB to 6 dB (w ∈ [0,1]) with ≈0.135 dB/mA attenuation efficiency.
• Nonlinear Multimode Activation: Weighted modes are recombined in a multimode germanium-on-silicon photodetector (R≈0.85 A/W, >2.8 GHz bandwidth) whose output photocurrent drives a carrier-injection microring resonator (MRR); by tuning Δλ, the MRR supplies flexible nonlinear transfer functions (sigmoid, ReLU, RBF), with activation speeds up to ~100 MHz.

3. Mathematical Framework and Neural Computation

Let E(ℓ) ∈ ℂ^M denote the field envelopes at layer ℓ for M modes. Linear synaptic weighting is given by:

$E^{(\ell+1)} = W^{(\ell)} E^{(\ell)}$

where W ∈ ℝ^{M×M} is implemented via demux, PIN attenuation, and remux. Input intensities x ∈ ℝ^M are summed by the detector:

$I_{\text{pd}} = R \sum_{i=1}^{M} |E_i|^2$

Activation is realized via the microring response:

$$y = f(I_{\text{pd}}) ≡ T_{\text{MRR}}(I_{\text{pd}})$$

where T_MRR(·) is a current-tuned Lorentzian-like nonlinear transmission. The neuron function is thus:

$y = f\bigl( R \sum_i w_i x_i \bigr)$

This formalism generalizes seamlessly to multilayer and convolutional layouts.

4. In-Situ Genetic Algorithm Training

MOENN dispenses with backpropagation, instead utilizing an in-situ genetic algorithm (GA) to directly tune on-chip PIN currents. All N weights are arranged as a chromosome, loaded onto the chip, and evaluated on the training set with fitness given by classification accuracy. GA operations include elite selection (top 5% retained), parent selection (top 20%), crossover (randomized cut-point recombination), and mutation (Gaussian perturbation of 15% of the population). Evolution continues until fitness exceeds threshold or epoch limit. The cost metric minimized is:

$$J = 1 - \frac{1}{N_{\text{train}}} \sum_{n=1}^{N_{\text{train}}} \delta(y_n, \hat{y}_n)$$

where δ is an indicator for correct classification.

5. Network Performance and Benchmarked Applications

MOENN has demonstrated high efficacy in various tasks:

Task	Topology	Modes	Training Accuracy	Test Accuracy
Two-Class Toy Dataset	2–4–2	2	99%	—
Iris Classification	4–4–3 (factorized)	2	92.86%	92.10%
ECG Emotion Recognition	Conv1D + FC (20 neu.)	3	90.9%	90.7%

Operating at 100 MHz per mode, a 3-mode MOENN achieves 3.6 GOPS, 0.4 GOPS/mm² in a 5 × 1.8 mm² footprint, and energy efficiency of ≈684 fJ/OP. Versatility is exemplified by reconfiguring the network into a convolutional topology encompassing kernels, activation, batch normalization, max-pooling, and a final FC layer (Xiang et al., 22 Jan 2026).

6. Comparative Evaluation and Scalability

MOENN provides operational simplification versus WDM-based OENN (single laser, no thermal tuning) and coherent SDM (no mesh calibration, phase insensitivity). Unlike previous MDM demonstrations, it supports full nonlinear activation on-chip via integrated detection and MRR. Control is reduced to direct PIN current mapping, obviating complex matrix decomposition. Bandwidth can potentially be raised into the GHz by mitigating opto-electronic link parasitics. Scalability is underpinned by demonstration of up to 15 TE modes and possible hybridization with WDM, enabling multiplicative channel expansion. Advances in (de)multiplexers, splitters, and switches will allow growth to larger layer sizes and deeper networks.

7. Broader Context and Future Directions

MOENN establishes a new paradigm for robust, scalable, and energy-efficient photonic neural computation. Its monolithic integration, phase-insensitive architecture, and genetic algorithm training enable practical deployment in real-world applications. The approach generalizes to additional multiplexing strategies (modal, wavelength, polarization) for exponential increase in channel density. As supporting silicon photonic components mature, deeper and larger multilayer networks will be realizable. Extensions could include all-optical, ultradense diffractive neural networks leveraging metasurface-based polarization multiplexing, as demonstrated in related CMOS-compatible platforms (Luo et al., 2021). Such architectures are suitable for vision and sensing tasks requiring multi-channel parallelism and low latency.

In summary, by exploiting guided-wave eigenmode orthogonality within a fully integrated SOI framework, MOENN delivers robust, high-performance neural inference with simple electronic control, supporting scalable opto-electronic intelligence for a variety of machine learning applications.