Deep convolutional optical neural network driven by micro-ring resonator array and microcomb

In the post-Moore’s Law era, traditional electronic computing architectures are approaching their physical limits in terms of energy efficiency, facing severe computing power bottlenecks and energy consumption challenges. Optical computing, with its characteristics of high parallelism, low latency, and high energy efficiency, is regarded as an important path to break through the bottlenecks of electronic computing. Although optical neural networks (ONNs) have shown great potential, current research is mostly limited to single-layer linear operations or simple fully connected networks, and is severely constrained by the transmission loss of optical waveguides, making it difficult to achieve deep network cascading. Furthermore, the key “sliding window” convolution operation of convolutional neural networks (CNNs), which are the core of deep learning, is difficult to directly implement on optical structures, and existing spatial multiplexing schemes can lead to significant loss and crosstalk. Therefore, this study aims to propose an on-chip integrated deep convolutional ONN system architecture based on a micro-ring resonator (MRR) array and microcomb. This architecture is dedicated to solving the implementation challenges of optical convolution and interlayer signal attenuation. Through frequency-domain parallel processing and optoelectronic-optical nonlinear activation mechanisms, it constructs an ultra-compact, high-computing power, high-energy-efficiency, and deeply scalable on-chip optical computing system to meet the needs of future high-density integration and edge computing.

In terms of architecture design, the system utilizes optical microcombs as a high-computational-density multi-wavelength light source. Through the Waveshaper, image pixel information is encoded onto the intensities of different wavelengths, and power pre-compensation is performed. The core convolution function is implemented using the periodic free spectral range (FSR) characteristics of MRRs. By precisely adjusting MRR parameters, it generates the same transmission response for multiple wavelengths separated by FSR in the frequency domain, thereby converting the time-domain "sliding window" into parallel processing in the frequency domain. With just nine MRRs, it is possible to achieve synchronous weighting of multiple windows using a 3×3 convolution kernel, and differential operations of positive and negative weights are completed through a wavelength division multiplexer and a balanced photodetector (BPD). To address the issue of cascading loss in optical signals, an interlayer nonlinear activation unit for optoelectronic-optical (OEO) conversion is designed. The signal is converted into current by the BPD, amplified by a transimpedance amplifier (TIA), and then the unidirectional conduction characteristic of diodes is utilized to physically simulate the ReLU activation function. Ultimately, the processed signal is fed back to a new light source by a microring modulator, achieving signal regeneration and deep propagation. In the verification phase, a physical link model is established using the ANSYS Lumerical INTERCONNECT module to verify the edge extraction capability of the MRR array and the electrical signal rectification function of the nonlinear unit. Furthermore, system-level numerical simulation is conducted using the PyTorch library. Under the constraints of weight precision (retained to two decimal places) and range ([-1,1]), a deep network comprising multiple hidden layers is constructed. The Optical Handwritten Digit Dataset is used for training and inference testing, and random noise is introduced to evaluate the system's anti-interference capability.

Simulation indicates that the optical convolution unit can accurately extract features of images, and the optoelectronic-optical nonlinear activation unit successfully achieves signal amplification and rectification (ReLU effect), effectively solving the problem of optical transmission loss limiting network depth. In the classification task of the Optical Handwritten Digit Dataset, the system demonstrates excellent inference ability: for the four-class task of 0-3, the basic architecture achieves a high accuracy rate of 98.61%; after expanding the network to a dual convolution kernel, three hidden layers, and 18×18 neuron scale, for the ten-class task of 0-9, the system still maintains an accuracy rate of 95.83%. Error analysis further verifies the robustness of the system. After introducing random noise simulating thermal crosstalk and process errors during the inference process, applying ±5% weight error results in a maximum accuracy decrease of only 2.95%, and applying ±10% error results in an accuracy decrease of about 5.9%, indicating that the system can maintain high computational reliability in the face of actual hardware interference.

We propose and validate an on-chip integrated architecture for deep convolutional optical neural networks (ONNs) based on MRR arrays and microcombs. This research breaks the limitations of traditional ONNs in convolutional operations and depth scaling. By creatively utilizing the FSR characteristics of MRRs to convert time-domain sliding windows into frequency-domain parallel processing, the computational density and integration level are greatly enhanced. Combined with a nonlinear activation scheme involving optoelectronic conversion, it effectively overcomes signal link loss, making it possible to cascade multiple layers of deep networks. Experimental results show that the system exhibits high accuracy and good noise resistance in handwritten digit classification tasks.