Guangwei Cong

As Moore’s law approaches its limit, the search for alternative computing platforms beyond conventional digital electronics is intensifying. Photonic interferometric processors offer a route to ultrahigh-throughput computing by exploiting the spatial, temporal, and wavelength degrees of freedom of light. However, scaling such systems has remained a major challenge due to the lack of robust calibration strategies for large interferometric meshes subject to fabrication disorder and crosstalk. Here, we present a scalable silicon photonic interferometric complex-valued matrix processor. This is achieved by introducing a newly developed disorder-aware calibration framework that directly infers circuit parameters and crosstalk from input–output measurements, eliminating the need for auxiliary monitoring structures. This approach enables high-fidelity operation of large-scale interferometric meshes beyond the regime accessible with conventional calibration methods. We experimentally demonstrate a 16 × 16 processor achieving near-ideal complex-valued matrix operations, including MNIST classification without off-chip electronic neural networks. Importantly, we further demonstrate wavelength-insensitive operation across 86 channels over a 2.2-THz bandwidth without recalibration, enabling simultaneous scaling in both spatial and wavelength domains within a single processor. Our results establish a scalable architecture for interferometric photonic computing and open a path toward ultrahigh-throughput matrix–matrix multiplication using wavelength-division multiplexing.

Optical neural networks (ONNs) are the technology that implements neural computations directly in the optical domain. ONNs offer unique merits such as ultra-low latency, high energy efficiency, and intrinsic parallelism. Photonic neural networks (PNNs) are a particular form of ONNs that is implemented on photonic integrated circuits promising mass-producibility, robustness, and compactness in addition to the merits of ONNs. Despite the recent advancement of photonic integrated circuit technologies such as silicon photonics and the numerous proof-of-concept demonstrations, PNNs still face key challenges, including scalability and the implementation of nonlinear operations essential for high-performance, efficient AI computation. In this tutorial paper, we begin by reviewing the historical development of optical computing and recent advances in PNNs. We identify current research directions, key challenges, and possible solutions. We then present our recent approaches that enable streaming processing without relying on traditional activation functions, instead we propose utilizing new models that align with the physical nature of PNNs, such as optimal combinations of electro-optic data encoding and cascaded linear operations. Finally, we outline promising future directions for this rapidly evolving field.

Micro-transfer printing (µ-TP) enables high-density integration of pre-structure-defined device coupons, facilitating multiple photonic functions on a single chip after the CMOS back-end-of-line process. This study demonstrates µ-TP of a pre-structured thin-film lithium-niobate (TFLN) coupon with two parallel etched TFLN waveguides onto a silicon photonics (SiPh) platform, achieving a 3σ alignment accuracy of approximately 0.5 μm. It incorporates low-loss design-optimized inter-layer couplers with a minimum transition loss of less than 0.1 dB and traveling-wave coplanar electrodes. The Si/TFLN hybrid Mach–Zehnder modulator achieves an on-chip device loss of 1.0 dB and a 3-dB EO bandwidth exceeding 67 GHz. This represents the first µ-TP integration of pre-structured TFLN waveguides with low-loss inter-layer coupling on a SiPh platform, marking a significant advancement in scalable, high-performance photonic integration.

The heterogeneous integration of thin-film lithium niobate (TFLN) provides energy-efficient broadband optical modulation of photonic platforms. Micro-transfer printing (µ-TP) yields dense integration of heterogeneous materials and cost-effective photonic integrated circuits. However, the reports so far have considered unpatterned slab membranes for the structure of the µ-TP integrated TFLN, and the short interaction length increases the driving voltage. In this study, the µ-TP of pre-structure-defined TFLN Mach-Zehender (MZ) interferometer was demonstrated for up to 1.1 cm with propagation loss of 0.7 dB/cm. The TFLN MZ modulator integrated on a silicon substrate using µ-TP exhibited the lowest driving voltage V_π of 2.7 V and a modulation bandwidth of 28 GHz with a low insertion loss of 0.97 dB. In addition, modulation frequencies exceeding 50 GHz and V_π of 8.9 V were demonstrated in a device with 3-mm-long electrodes.

Integrated photonic neural networks (PNNs) usually adopt traditional convolutional neural network (CNN) or multilayer perceptron (MLP) network models. These models consist of horizontally cascaded deep layer architectures interleaved by nonlinear activation functions. However, there are practical challenges for on-chip realizing such architectures, including the optical loss and the lack of efficient on-chip optical activation nonlinearity. Here, we propose a vertically hierarchical photonic neural network leveraging electro-optical element-wise multiplication to extract an element-wise feature in a polynomial projection space, which enables high-accuracy classification. For this network architecture, the light propagates through only two fully connected linear layers; thus, vertical extension to the deep layer is not limited by optical loss. This electro-photonic network can perform equivalently to or outperform optical CNN and MLP models even without interleaving deep layers by activation functions, benchmarking ∼97.9%, ∼87.7%, and ∼90.3% average blind-testing accuracies, for the whole test sets of MNIST handwritten digits, Fashion-MNIST images, and KMNIST Japanese cursive characters, respectively. It also demonstrates a >99% accuracy for boundary prediction of 12-labeled clusters. This work presents a different PNN architecture, which offers both high performance and better amenability to an integrated photonics platform.

Controlling the intensity of emitted light and charge current is the basis of transferring and processing information¹. By contrast, robust information storage and magnetic random-access memories are implemented using the spin of the carrier and the associated magnetization in ferromagnets². The missing link between the respective disciplines of photonics, electronics and spintronics is to modulate the circular polarization of the emitted light, rather than its intensity, by electrically controlled magnetization. Here we demonstrate that this missing link is established at room temperature and zero applied magnetic field in light-emitting diodes^2,3,4,5,6,7, through the transfer of angular momentum between photons, electrons and ferromagnets. With spin–orbit torque^8,9,10,11, a charge current generates also a spin current to electrically switch the magnetization. This switching determines the spin orientation of injected carriers into semiconductors, in which the transfer of angular momentum from the electron spin to photon controls the circular polarization of the emitted light². The spin–photon conversion with the nonvolatile control of magnetization opens paths to seamlessly integrate information transfer, processing and storage. Our results provide substantial advances towards electrically controlled ultrafast modulation of circular polarization and spin injection with magnetization dynamics for the next-generation information and communication technology¹², including space–light data transfer. The same operating principle in scaled-down structures or using two-dimensional materials will enable transformative opportunities for quantum information processing with spin-controlled single-photon sources, as well as for implementing spin-dependent time-resolved spectroscopies.

We experimentally demonstrate long-time (~10 hours) continuous full-C-band automatic working wavelength alignment for silicon photonic cascaded-ring vernier filters against wide input-wavelength and temperature changes using reinforced Q-learning method, without pre-building look-up table and temperature monitor.

We propose a method of supercontinuum light generation enhanced by multimode excitation in a precisely dispersion-engineered deuterated SiN (SiN:D) waveguide. Although a regularly designed SiN-based nonlinear optical waveguide exhibits anomalous dispersion with the fundamental and first-order multimode operation, the center-symmetric light pumping at the input edge has so far inhibited the full potential of the nonlinearity of SiN-based materials. On the basis of numerical analysis and simulation for the SiN:D waveguide, we intentionally applied spatial position offsets to excite the fundamental and higher-order modes to realize bandwidth broadening with flatness. Using this method, we achieved an SNR improvement of up to 18 dB at a wavelength of 0.6 µm with an offset of about 1 µm in the Y-axis direction and found that the contribution was related to the presence of dispersive waves due to the excitation of TE₁₀, and TE₀₁ modes.

Training photonic circuits to learn specific machine learning tasks is the key for photonic AI. We review our work on on-chip training in silicon photonic circuits to implement novel classification computing and its recent progress.