On-chip Multimode Opto-electronic Neural Network

Technical Summary: On-chip Multimode Opto-electronic Neural Network

Problem Statement
While opto-electronic computing offers a promising path to overcome the energy and throughput limitations of electronic processors in the post-Moore era, practical deployment remains hindered by architectural fragility and complexity. Existing approaches face significant trade-offs:

• Wavelength-Division Multiplexing (WDM): Typically relies on microring resonators (MRRs) and multiple narrow-linewidth lasers. These systems require delicate wavelength management to counteract thermal crosstalk and temperature drift, increasing power consumption and control complexity.
• Coherent Space-Division Multiplexing (SDM): Based on Mach-Zehnder interferometers (MZIs), these systems demand precise phase control across entire optical paths. They are highly sensitive to accumulated phase noise and require complex matrix decomposition algorithms for weight mapping, making scalability difficult.
• Mode-Division Multiplexing (MDM): While MDM has been introduced to enhance parallelism, previous demonstrations have been confined to linear computational regimes, failing to implement complete, nonlinear neural networks on a single chip.

Methodology
The authors demonstrate a monolithic Multimode Opto-electronic Neural Network (MOENN) fabricated on a silicon-on-insulator (SOI) platform. The architecture addresses the limitations of prior works through the following design principles:

1. Single-Wavelength, Multi-Mode Operation: Instead of using multiple wavelengths or phase-sensitive coherent detection, the system encodes independent information onto orthogonal waveguide eigenmodes (e.g., TE₀, TE₁, TE₂) within a single wavelength. This inherently eliminates spectral crosstalk and phase noise sensitivity.
2. Integrated Components: The chip monolithically integrates:
   • Input Encoders: High-speed Mach-Zehnder modulators (MZMs) with >15 GHz bandwidth.
   • Programmable Fan-in/Fan-out: Cascaded asymmetric directional couplers (ADCs) for low-loss, low-crosstalk mode multiplexing.
   • Linear Weighting: Broadband optical attenuators (PIN diodes) that modulate intensity to apply weights from 0 to 1.
   • Nonlinear Activation: A multimode photodetector (Germanium-on-Silicon) coupled with a carrier-injection MRR. This Optical-Electrical-Optical (O-E-O) link performs weighted summation and applies nonlinear activation functions (e.g., Sigmoid, ReLU) by modulating the MRR transmission based on the photocurrent.
3. In-situ Training: To bypass the need for precise analytical modeling of the analog hardware (which is prone to thermal crosstalk and noise), the system employs a genetic algorithm (GA) for hardware-in-the-loop training. Control parameters are encoded into feature vectors, and the network is optimized iteratively based on classification accuracy on the training set, without requiring gradient calculations.

Key Contributions

• First Multimode OENN: This work presents the first demonstration of a complete, nonlinear neural network on a multimode photonic platform, moving beyond linear MDM demonstrations.
• Robust Architecture: By utilizing orthogonal eigenmodes and single-wavelength operation, the system achieves robustness against spectral crosstalk and phase noise, simplifying control requirements compared to WDM and coherent SDM schemes.
• Monolithic Integration: The successful integration of encoders, mode-division units, and nonlinear activation functions on a compact 5.0 × 1.8 mm² chip.
• Versatile Training: The demonstration of a gradient-free training protocol that effectively compensates for fabrication errors and environmental deviations.

Results
The fabricated MOENN chip was evaluated on three distinct tasks:

1. Nonlinear Decision Boundary (Two-Class Dataset): A simple network (2 input, 4 hidden, 2 output neurons) successfully resolved a linearly inseparable synthetic dataset. The system achieved 99% accuracy, with the decision boundary converging within 20 iterations.
2. Iris Classification Benchmark: A three-layer photonic neural network was mapped to the two-mode MOENN using a weight pruning strategy (factorizing an 8×8 matrix into four 2×2 sub-networks). The system achieved 92.10% accuracy on the test set and 92.86% on the training set.
3. ECG-Based Emotion Recognition: The three-mode MOENN was reconfigured to execute a one-dimensional Convolutional Neural Network (1D-CNN) for classifying heart electrical activity into four emotional states (baseline, stress, amusement, meditation). By performing convolution and nonlinear activation in a time-multiplexed manner on-chip, the system achieved 90.7% accuracy on the test set and 90.9% on the training set.

Significance and Claims
The authors claim this work establishes a new paradigm for opto-electronic computing characterized by simple control and excellent robustness. Key implications highlighted include:

• Scalability: The architecture avoids the scaling penalties associated with phase control in MZIs or wavelength management in WDM systems.
• Deployability: The system's immunity to environmental deviations makes it suitable for real-world applications where thermal stability is difficult to maintain.
• Extensibility: The MOENN can be seamlessly extended to a hybrid MDM-WDM framework, potentially reducing the number of required laser sources by a factor equal to the number of mode channels.
• Performance: Operating at 100 MHz, the three-mode chip delivers a throughput of 3.6 GOPS with an energy efficiency of 684 fJ/OP, demonstrating a viable path toward scalable, deployable photonic intelligence.