A 103-TOPS/mm$^2$ Integrated Photonic Computing Engine Enabling Next-Generation Reservoir Computing

TL;DR

This work demonstrates the first integrated photonic NG-RC on a silicon chip using a passive star coupler and on-chip delay lines, achieving 60 Gbaud operation and a computing density of 103 TOPS/mm^2. By omitting training requirements for the reservoir and leveraging photodiode-generated quadratic nonlinearity, the system delivers high speed, low energy, and improved fabrication tolerance with a compact footprint. The approach is validated on multiple tasks, including Lorenz forecasting, NARMA10, and COVID-19 image classification, showcasing strong predictive and classification performance. The results establish a practical pathway toward ultrafast on-chip photonic reservoir computing and scalable, high-density photonic computing engines.

Abstract

Reservoir computing (RC) is a leading machine learning algorithm for information processing due to its rich expressiveness. A new RC paradigm has recently emerged, showcasing superior performance and delivering more interpretable results with shorter training data sets and training times, representing the next generation of RC computing. This work presents the first realization of a high-speed next-generation RC system on an integrated photonic chip. Our experimental results demonstrate state-of-the-art forecasting and classification performances under various machine learning tasks and achieve the fastest speeds of 60 Gbaud and a computing density of 103 tera operations/second/mm$^2$ (TOPS/mm$^2$). The passive system, composed of a simple star coupler with on-chip delay lines, offers several advantages over traditional RC systems, including no speed limitations, compact footprint, extremely high fabrication error tolerance, fewer metaparameters, and greater interpretability. This work lays the foundation for ultrafast on-chip photonic RC, representing significant progress toward developing next-generation high-speed photonic computing and signal processing.

Figures (7)

• Figure 1: Basic principle of reservoir computing. a, The framework of traditional reservoir computing systems, including the input layer, the reservoir layer, and the readout layer. b, The flow of next-generation RC, which is exactly equivalent to traditional RC. NG-RC employs a constant $C$, input data $O_{linear}$, and the quadratic functionals of input data $O_{nonlinear}$ to generate its output. c, The schematic diagram of our photonic RC, which is perfectly equivalent to NG-RC. The proposed photonic reservoir computing is realized using on-chip delay lines and a star coupler.
• Figure 2: Experimental setup. PC, polarization controller; AWG, arbitrary waveform generator; EDFA, erbium-doped fiber amplifier; PD, photodiode. The unmodulated light and modulated signal as the constant and input data are injected into the photonic reservoir computing respectively.
• Figure 3: Experiment result of Lorenz. a, The ground truth three-dimensional diagram of the Lorenz system. b, The three-dimensional prediction result of our photonic system. c, The ground truth (blue) and prediction (red) variables of Lorenz during the training phase. d, The ground truth (blue) and prediction (red) variables of Lorenz during the testing phase. The NMSE of z variable between ground truth and prediction is $1.43 \times 10^{-2}$.
• Figure 4: The comparison of reported works. The performance is evaluated with a focus on operation speed and the Normalized Mean Squared Error of the NARMA10 model. These works all have about 50 nonlinear nodes.
• Figure 5: Experiment results of Covid-19 classification. a, The preprocessing of the Covid-19 images. b, The confusion matrices for Covid-19 classification. The accuracy of the Covid-19 classification is 92.1% c, The receiver operating characteristic curve of Covid-19 classification. The area under the curve of this task is 0.93.