Opto-Electronic Neural Network Achieves 90.7% Accuracy on Silicon-On-Insulator Platform

Scientists are tackling the limitations of current opto-electronic systems with a novel neural network architecture. Jinlong Xiang, Youlve Chen, and Yuchen Yin from Shanghai Jiao Tong University, alongside Yufeng Zhang from Aalto University and Chaojun Xu et al, have demonstrated the first multimode opto-electronic neural network (MOENN) fabricated on a silicon-on-insulator platform. This breakthrough overcomes challenges posed by delicate wavelength control and phase sensitivity, offering a robust and inherently stable system that utilises orthogonal waveguide eigenmodes for independent data transmission. The research is significant because it monolithically integrates all necessary components , from input encoding to nonlinear activation , and showcases exceptional versatility through in-situ training, achieving over 90% accuracy on both Iris classification and electrocardiogram-based emotion recognition, paving the way for scalable and deployable intelligent systems.

The team successfully monolithically integrated all essential functional components onto a single chip, including input encoders, programmable mode-division fan-in/-out units, and crucially, nonlinear multimode activation functions. This breakthrough establishes a new paradigm for opto-electronic computing, offering simplified control and exceptional robustness for scalable, deployable intelligence.

The research establishes a versatile system, validated through in-situ training using a genetic algorithm, which effectively resolved the nonlinear decision boundaries of a two-class dataset. Achieving 92.1% accuracy on the Iris classification benchmark demonstrates the MOENN’s capacity for complex pattern recognition. Furthermore, the team reconfigured the MOENN into a one-dimensional convolutional neural network, attaining an impressive 90.7% accuracy on the challenging task of electrocardiogram-based emotion recognition. This reconfiguration highlights the adaptability of the architecture and its potential for diverse applications beyond standard classification problems.

Experiments show the MOENN operates on the principle of “broadcast-and-weight”, where input signals are encoded and multiplexed onto distinct guided eigenmodes within a shared bus waveguid. Low-loss, low-crosstalk mode multiplexers, specifically asymmetric directional couplers, facilitate this multiplexing process. To create weighted interconnections between neural layers, each mode signal is split into multiple branches, enabling precise control over signal propagation and ultimately, network performance. The system’s design prioritises simplicity and robustness, addressing key limitations of existing opto-electronic architectures.

This work overcomes critical challenges in field-deployable opto-electronic computing systems, which often suffer from sensitivity to environmental deviations like thermal crosstalk and temperature drift. By utilising a single wavelength and eliminating the need for precise phase control, the MOENN significantly reduces system complexity and power consumption. The researchers’ approach bypasses the limitations of wavelength-division multiplexing, which requires multiple lasers, and coherent space-division multiplexing, which is susceptible to phase noise. This innovation provides a compelling pathway towards scalable and practical photonic intelligence, paving the way for future advancements in artificial intelligence hardware.

Monolithic Opto-electronic Neural Network Fabrication and Operation demonstrate

Scientists engineered a monolithic multimode opto-electronic neural network (MOENN) on a silicon-on-insulator platform, overcoming limitations of existing architectures reliant on delicate wavelength management or phase-sensitive detection. The study pioneered a system utilising orthogonal waveguide eigenmodes as independent information carriers, achieving robust single-wavelength operation inherently immune to spectral crosstalk and phase noise. Researchers fabricated a chip monolithically integrating input encoders, programmable mode-division fan-in/-out units, and crucially, nonlinear multimode activation functions, demonstrating a complete neural network on a single platform. Experiments employed asymmetric directional couplers (ADCs) to achieve linear mode-division fan-in, cascading these components to minimise loss and crosstalk between modes.

To establish weighted interconnections, each mode signal was split and subjected to mode-selective weighting via broadband optical attenuators, applying a normalised weight range of 0 to 1. Subsequently, weighted signals were converted back to their original mode and combined, enabling weighted summation on all mode signals using a high-sensitivity multimode photodetector. This innovative approach bypasses the need for pre-amplification and filtering, simplifying the computational process significantly. The team harnessed a carrier-injection microring resonator (MRR) to implement nonlinear activation functions, configuring it via optical-electrical-optical (O-E-O) conversion.

By modifying the detuning between the pump wavelength and the MRR resonance, researchers implemented functions like Sigmoid or ReLU, achieving nonlinear modulation of light transmission. To benchmark learning capacity, the MOENN was trained in-situ using a genetic algorithm, successfully resolving nonlinear decision boundaries and attaining 92.1% accuracy on the Iris classification benchmark. Furthermore, scientists reconfigured the MOENN into a one-dimensional convolutional neural network, achieving 90.7% accuracy on an electrocardiogram-based emotion recognition task, demonstrating versatility for real-world applications. This work establishes a new paradigm of simple control and excellent robustness, providing a compelling path toward scalable, deployable intelligence.

Silicon Photonics Achieves 92.1% Iris Classification Accuracy

Scientists have demonstrated the first multimode opto-electronic neural network (MOENN) fabricated on a silicon-on-insulator platform. This breakthrough architecture utilizes orthogonal waveguide eigenmodes as independent information carriers, achieving robust single-wavelength operation inherently immune to spectral crosstalk and phase noise. The fabricated MOENN chip monolithically integrates input encoders, programmable mode-division fan-in/-out units, and crucially, nonlinear multimode activation functions, paving the way for scalable intelligence. Experiments revealed the system’s versatility through in-situ training via a genetic algorithm, successfully resolving nonlinear decision boundaries of a two-class dataset.

The team measured a 92.1% accuracy on the Iris classification benchmark, demonstrating the MOENN’s capability for inseparable data classification. To further validate its potential, researchers reconfigured the MOENN into a one-dimensional convolutional neural network, attaining a competitive 90.7% accuracy on an electrocardiogram-based emotion recognition task. This reconfiguration highlights the adaptability of the MOENN for complex, real-world applications. Measurements confirm that the mode multiplexer exhibits insertion losses below 0.5 dB, 0.8 dB, and 1.0 dB for the TE1, TE2, and TE3 modes respectively, within a wavelength range of 1510nm to 1555nm.

Furthermore, the maximum modal crosstalk for all three modes remained below -12 dB, ensuring signal integrity. The measured 3-dB bandwidths of the Mach-Zehnder modulators (MZMs) exceeded 15GHz, significantly faster than thermally tuned MZI encoders, enabling high-speed input encoding. The attenuation factor of the P-doped-intrinsic-N-doped optical attenuators was characterized as a function of injected current, allowing for precise weight control. Scientists recorded a photocurrent responsivity of the multimode photodetector for TE0-TE3 channels under a reverse-bias voltage of -2V, confirming efficient signal detection. Nonlinear temporal response measurements of the activation unit, using input pulses with increasing optical power, indicated an operation speed of 100MHz, demonstrating rapid processing capabilities. This work establishes a novel opto-electronic computing paradigm with simple control and excellent robustness, offering a compelling path toward deployable opto-electronic intelligence.

Silicon Photonics Achieves High Accuracy Classification of optical

Scientists have demonstrated the first multimode opto-electronic neural network (MOENN) fabricated on a silicon-on-insulator platform. This innovative architecture utilizes orthogonal waveguide eigenmodes to carry information independently, achieving robust single-wavelength operation that is resistant to spectral crosstalk and phase noise. The MOENN chip monolithically integrates input encoders, programmable mode-division fan-in/-out units, and crucially, nonlinear multimode activation functions, all on a single chip. Researchers successfully trained the system in-situ using a genetic algorithm, enabling it to resolve nonlinear decision boundaries within a two-class dataset.

The MOENN achieved 92.1% accuracy on the Iris classification benchmark and 90.7% accuracy on an electrocardiogram-based emotion recognition task, demonstrating its versatility. The authors acknowledge limitations including the speed of the nonlinear activation units and the potential for further scaling to higher-order modes. Future work will focus on increasing computational performance through these improvements, as well as exploring hybrid multiplexing architectures, combining wavelength-division and mode-division multiplexing, to reduce laser requirements and system complexity. This research establishes a new paradigm for opto-electronic computing, offering a compelling route towards scalable, robust, and energy-efficient intelligent systems.