Xingyuan Xu

The rise of edge-centric 6G networks demands real-time intelligent processing of heterogeneous radio-frequency (RF) signals. Complex-valued blind source separation (CV-BSS) provides a physical-layer capability, enabling the blind extraction of source components from overlapping RF signals. Here, we demonstrate an integrated photonic vector processor (IPVP) that implements real-time CV-BSS via in-physics computation directly on raw RF signals. The IPVP operates across multiple RF bands up to 24 GHz and achieves blind, in-hardware adaptation to unknown sources and channels via in situ learning. Experiments demonstrate the efficacy of IPVP in multiple-input multiple-output (MIMO) communications and multi-band radar interference suppression. The IPVP reduces the EVMs of the two recovered signals from over 50% to 6.18% and 5.3%, respectively, while enhancing the SIR by up to 26 dB. This work demonstrates, to our knowledge, the first photonic processor capable of real-time CV-BSS at RF, highlighting the potential of integrated photonic hardware for ultralow-latency, edge-intelligent signal processing.

Photonic hardware offers a promising route to high‐speed and energy‐efficient convolution computing beyond the bandwidth and power limitations of conventional electronic platforms. Here, we propose and experimentally demonstrate a cross‐scale integrated photonic convolver (CSIPC) that enables cross‐scale feature extraction through a single training process. Implemented on a hierarchically coupled finite‐impulse‐response (FIR) photonic chip, the CSIPC exploits the propagation of conjugate‐symmetric interference‐weight relations through the coupling network, together with the spatiotemporal interleaving induced by serial‐to‐parallel conversion. A single set of trained chip weights can be mapped at different output ports into a family of equivalent convolution kernels with distinct effective tap numbers and temporal intervals. By selecting the output port and tuning the input symbol rate, 2 × 4, 2 × 2, and 2 × 1 two‐dimensional convolution kernels can be realized on the same chip. We experimentally validate multifunctional cross‐scale image feature extraction and achieve a maximum single‐mode equivalent computing throughput of 0.276 TOPS. These results provide a route toward cross‐scale photonic convolution using a single device and a single training process, while improving functional integration density and enabling the reuse of trained weights and hardware resources.

Dilated convolution is pivotal for dense prediction tasks such as semantic segmentation, as it expands the receptive field to capture multi‐scale contextual information without increasing parameter complexity. However, its electronic implementation often faces bottlenecks due to non‐contiguous memory access and high computational load. Here, we report a Reconfigurable Optical Dilated Convolution Accelerator (RODCA) based on a wavelength‐dispersion co‐tuning mechanism. By dynamically controlling optical frequency comb spacing and dispersion, RODCA enables reconfigurable dilation rates while improving spectral efficiency. The system achieves an equivalent computing throughput of 13.858 Tera operations per second (TOPS). We experimentally validate the RODCA using Laplacian kernels and semantic segmentation tasks, where it achieves a pixel accuracy of 88.46% and a mean Intersection over Union (mIoU) of 76.08%, closely matching in‐silico results. This work demonstrates RODCA's capability for efficient multi‐scale feature extraction, highlighting its potential for next‐generation photonic computing in complex AI applications.

Holographic communication enables immersive metaverse interaction but transmitting dynamic media like point clouds demands unprecedented bandwidth, a key challenge for current networks. We propose an optical semantic holographic communication architecture integrating deep learning algorithms, coherence-cloned Kerr microcombs, and hollow-core optical fiber to meet bandwidth, fidelity, and low-latency requirements, achieving the first 100 km coherent transmission using coherence-cloned Kerr microcombs. Deep learning-enabled semantic joint coding-modulation (JCM) improves capacity by 65.3% and enhances channel robustness compared with conventional digital protocols. For phase-sensitive constellations of the JCM-generated discrete-time continuous-amplitude symbols, the microcombs act as phase-matched carriers/oscillators for stable coherent detection. The ultra-low nonlinearity of the hollow-core fiber preserves semantic integrity, further suppresses link-phase noise, and synergizes with microcombs for physical-layer phase noise control. Our system achieves 2.5875 trillion 3D points/s with superior quality, setting a new artificial intelligence-native metaverse benchmark.

The exponential growth of cloud computing and artificial intelligence (AI) applications has driven an urgent need for high-bandwidth, energy-efficient hardware architectures in data centers. With Moore’s Law nearing its limits, optical neuromorphic computing hardware offers a promising alternative, providing ultra-high speeds and minimal energy consumption due to its analog architecture. Here, we propose the microcomb-enabled parallel optical convolution streaming processor (OCSP) with time, space, and wavelength three-dimensional multiplexing, operating at data rates of 50 GBaud or higher, achieving a convolution computing speed of up to 4 trillion operations per second (TOPS). Moreover, the OCSP uses a robust self-calibration mechanism to achieve accurate optical phase calibration and set-up of its convolution function. This innovative approach leverages time-space interleaving passive periodic interference architecture, incorporating wavelength-division-multiplexing technology, and is verified experimentally for parallel image feature extraction and recognition tasks. Our OCSP offers a practical pathway for seamlessly integrating photonic computing units into data center interconnects, unlocking photonic computing’s potential for scalable, low-latency AI workloads.

Microwave photonic (MWP) filters are essential components in microwave systems due to their wide bandwidth, low loss, and immunity to electromagnetic interference. Steep roll‐off is critical for precise spectral shaping and effective interference suppression, yet conventional MWP filters face challenges in simultaneously achieving steep transitions and high reconfigurability. Here, reconfigurable MWP filters with steep roll‐off based on a transversal filter structure using an optical microcomb source are demonstrated. Four different types of single‐band MWP filters with ultrahigh roll‐off rates up to ∼32.6 dB/GHz and a minimum 20‐dB shape factor of ∼1.15 are achieved. In addition, simply through programming tap coefficients, band‐pass filters with tunable centre frequencies ranging from 5 to 15 GHz and dual‐band MWP filters with various filter responses are demonstrated without changing any hardware, where steep roll‐off is also validated. The reconfigurable MWP filters with steep roll‐off offer a stable and versatile solution for applications requiring high spectral selectivity, such as next‐generation wireless networks, high‐resolution radar imaging, and advanced biomedical sensing.

Transposed convolution, crucial in large‐parameter generative models ranging from content creation to autonomous driving, imposes substantial demands on GPU memory and energy consumption in electronic processors. Electronic processors, fundamentally limited by the von Neumann architecture and further hindered by silicon‐based quantum tunneling effects, struggle to meet the stringent real‐time requirements of modern generative workloads. In contrast, optical computing—exploiting ultra‐wide bandwidth and ultra‐low power consumption—offers a promising alternative for high‐speed transposed convolution in next‐generation AI. Here, we introduce a high‐speed and reconfigurable photonic transposed convolution accelerator (PTCA). By interleaving wavelength, temporal, and spatial dimensions and leveraging an integrated Kerr microcomb for data‐dimension expansion, the PTCA achieves tera operations per second (TOPS) with 100% bit efficiency. Experiments demonstrate a processing speed of 1.026 TOPS, making it, to the best of our knowledge, the fastest reconfigurable PTCA to date. In Fashion‐MNIST reconstruction tasks, this system achieves a mean squared error (MSE) of 0.0062 without any additional post‐processing by electronic fully connected layers. Our work thus establishes a high‐speed, reconfigurable photonic paradigm for accelerating future generative AI.