Compact, Large-Scale Photonic Neurons by Modulation-and-Weight Microring Resonators

TL;DR

This work introduces a compact, large-scale photonic neuron in which modulation and weighting are performed within each microring, using coexistent carrier and thermal tuning to minimize spectral alignment constraints and footprint. The architecture enables configurable feedforward and recurrent operation, including a simple electrical feedback path that provides memory for temporal processing. Demonstrated with a 10-MRR array, it achieves a $3\times3$ convolution with RMSE $<5\%$, a per-weight footprint of $80\times 45~\mu\mathrm{m}^2$, and an average weight tuning power of $0.186~\mathrm{mW}$, yielding $4.67~\mathrm{TOPS/s/mm^2}$ and $105~\mathrm{TOPS/W}$ on-chip tuning efficiency. These results position modulation-and-weighting MRR banks as scalable building blocks for large-scale neuromorphic photonic systems with high density, low power, and functional flexibility.

Abstract

Neuromorphic photonics promises sub-nanosecond latency, ultrawide bandwidth, and high parallelism, but practical scalability is constrained by fabrication tolerances, spectral alignment, and tuning energy. Here, we present a large-scale, compact, and reconfigurable photonic neuron in which each microring performs modulation and weighting simultaneously. By exploiting both carrier and thermal tuning within a single device, this architecture reduces footprint, relaxes spectral alignment requirements to just two optical components, and yields a steep transfer response that lowers tuning energy. The proposed neuron supports multiple operating configurations, allowing its dynamical behavior to be adapted to different computational tasks. In particular, a short electrical feedback path enables recurrent operation, providing tunable short- and long-term memory for temporal processing. Using a 10-microring resonator array, we demonstrate both spatial and temporal computing, including a 3$\times$3 convolution for image processing with an error of $<$5\% and high-frequency financial time-series prediction. Each modulation-weighting element occupies 80$\times$45 \SI{}{\micro\meter^2} and consumes an average of \SI{0.186}{\milli\watt}, corresponding to a compute density of \SI{4.67}{TOPS/s/\milli\meter^2}. Excluding electronic power, the on-chip tuning efficiency reaches approximately \SI{105}{TOPs/\watt}, which is comparable to state-of-the-art implementations. These results indicate that modulation-and-weighting microring resonator banks provide a scalable building block for large-scale neuromorphic photonic systems, offering a favorable combination of compact footprint, low power consumption, and functional flexibility.

Figures (5)

• Figure 1: (a) Neural network architecture composed of numerous small units that perform weighting, summation, and nonlinear activation. Photonic implementations typically require lasers and multiple integrated photonic devices for each function. (b) A typical MRR-based photonic neuron, comprising a bank of multiple lasers at different wavelengths, a ring modulator bank, an MRR-based weight bank, and a balanced PD. (c) The proposed approach, which consolidates both modulation and weighting in a single MRR bank and employs a single-ended PD, thereby reducing complexity and footprint. (d) Three possible configurations of the proposed MRR bank, enabling neuron types in feedforward, recurrent (short-term memory), and combined long- and short-term memory modes.
• Figure 2: Architecture and characterization of the large-scale photonic processor. (a) System schematic. (b) Optical spectra and (c) Current-tuning curve of the ten MRRs. (d) Micrograph of the dual-modulation microring. $i_{ht}$ denotes the thermal tuning current applied to the heater. $v_n$ and $v_p$ stand for the voltage applied to the p-doped and n-doped regions of the PN junction for modulation, respectively. (e) Micrograph of the photonic chip. (f) Photograph of the fully packaged chip.
• Figure 3: (a) Schematic illustration of photonic image convolution. (b) Grayscale and (c) color image outputs produced by convolving each original image (upper right corner) with three different kernels, whose values are indicated in the leftmost column. In addition to the blurring effect that is achieved by averaging neighboring pixels, other kernels perform edge detection by computing local derivatives in specific directions. The Laplacian kernel applies an isotropic derivative, enhancing high-frequency features, such as the thin vertical and horizontal lines of the fence behind the tiger.
• Figure 4: Results of high-frequency trading. (a) Schematic illustration of the photonic high-frequency trading processor. (b–d) correspond to three different stock symbols, which are AAPL, TSLA, and GOOG, respectively. In each panel, the top row shows the stock price over one trading day. The middle row displays the simulated photonic neuron output, generated by convolving 10 consecutive price points; green and red dots denote sell and buy signals, respectively. Thresholds for buy/sell decisions are determined during PSO-based training. The bottom row illustrates the resulting cumulative profit for each stock, demonstrating robust gains irrespective of price trend.
• Figure 5: Performance comparison of the photonic HFT using three different neuron types, which are (a) basic convolution, (b) convolution with one historical feedback path, and (c) convolution with two historical feedback paths with short and long delays.