Zaijun Chen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Zaijun Chen has a total of 68 co-authors. Here are the top co-authors by number of shared publications:
- Dirk Englund
Massachusetts Institute of Technology, Cambridge, USA
7 shared publications - Ryan Hamerly
Research Laboratory of Electronics, MIT, Cambridge, USA
7 shared publications - Alexander Sludds
Massachusetts Institute of Technology
5 shared publications Kaiwen Xue
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA 90089, USA
5 shared publicationsLian Zhou
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA 90089, USA
5 shared publications- Liane Bernstein
MIT Research Laboratory of Electronics, Cambridge, MA, USA
4 shared publications
Latest Publications
- Open Access
- 13 June 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (24) , eadt3558
RF-photonic deep learning processor with Shannon-limited data movement
Edholm’s law predicts exponential growth in data rate and spectrum bandwidth for communications. Owing to exponentially increasing deep neural network computing demands and the slowing of Moore’s law, new computing paradigms are required for future advanced communications like 6G. Optical neural networks (ONNs) are promising accelerators but struggle with scalability and system overhead. Here, we introduce our multiplicative analog frequency transform optical neural network (MAFT-ONN), an artificial intelligence hardware accelerator that experimentally computes fully analog deep learning on raw radio frequency (RF) signals, performing modulation classification that quickly converges to 95% accuracy. MAFT-ONN also exhibits scalability with nearly 4 million fully analog operations for MNIST digit classification. Because of the Shannon capacity–limited analog data movement, MAFT-ONN is also hundreds of times faster than traditional RF receivers.
- 13 June 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (24) , eadt3558
Open AccessRF-photonic deep learning processor with Shannon-limited data movement
Edholm’s law predicts exponential growth in data rate and spectrum bandwidth for communications. Owing to exponentially increasing deep neural network computing demands and the slowing of Moore’s law, new computing paradigms are required for future advanced communications like 6G. Optical neural networks (ONNs) are promising accelerators but struggle with scalability and system overhead. Here, we introduce our multiplicative analog frequency transform optical neural network (MAFT-ONN), an artificial intelligence hardware accelerator that experimentally computes fully analog deep learning on raw radio frequency (RF) signals, performing modulation classification that quickly converges to 95% accuracy. MAFT-ONN also exhibits scalability with nearly 4 million fully analog operations for MNIST digit classification. Because of the Shannon capacity–limited analog data movement, MAFT-ONN is also hundreds of times faster than traditional RF receivers.
- Top Cited PapersOpen Access
- 6 June 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (23) , eadu0228
Hypermultiplexed integrated photonics–based optical tensor processor
The escalating data volume and complexity resulting from the rapid expansion of artificial intelligence (AI), Internet of Things (IoT), and 5G/6G mobile networks is creating an urgent need for energy-efficient, scalable computing hardware. Here, we demonstrate a hypermultiplexed tensor optical processor that can perform trillions of operations per second using space-time-wavelength three-dimensional optical parallelism, enabling O(N 2 ) operations per clock cycle with O(N) modulator devices. The system is built with wafer-fabricated III/V micrometer-scale lasers and high-speed thin-film lithium niobate electro-optics for encoding at tens of femtojoules per symbol. Lasing threshold incorporates analog inline rectifier (ReLU) nonlinearity for low-latency activation. The system scalability is verified with machine learning models of 405,000 parameters. A combination of high clock rates, energy-efficient processing, and programmability unlocks the potential of light for low-energy AI accelerators for applications ranging from training of large AI models to real-time decision-making in edge deployment.
- 6 June 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (23) , eadu0228
Top Cited PapersOpen AccessHypermultiplexed integrated photonics–based optical tensor processor
The escalating data volume and complexity resulting from the rapid expansion of artificial intelligence (AI), Internet of Things (IoT), and 5G/6G mobile networks is creating an urgent need for energy-efficient, scalable computing hardware. Here, we demonstrate a hypermultiplexed tensor optical processor that can perform trillions of operations per second using space-time-wavelength three-dimensional optical parallelism, enabling O(N 2 ) operations per clock cycle with O(N) modulator devices. The system is built with wafer-fabricated III/V micrometer-scale lasers and high-speed thin-film lithium niobate electro-optics for encoding at tens of femtojoules per symbol. Lasing threshold incorporates analog inline rectifier (ReLU) nonlinearity for low-latency activation. The system scalability is verified with machine learning models of 405,000 parameters. A combination of high clock rates, energy-efficient processing, and programmability unlocks the potential of light for low-energy AI accelerators for applications ranging from training of large AI models to real-time decision-making in edge deployment.
- 1 January 2025
- Published by Optica Publishing Group
- p. SS162_3
Scalable, Low-energy Homodyne Computing Crossbar based on TFLN and SiN/Si Photonics
We develop time-multiplexed 4x4 SiN/Si homodyne crossbars and amplitude-only TFLN electro-optics in a heterogenous integrated platform for photonic AI processing. Matrix-multiplication achieves >5-bit precision at GS/s and 10s fJ/symbol encoding.
- 1 January 2025
- Published by Optica Publishing Group
- p. SS162_3
Scalable, Low-energy Homodyne Computing Crossbar based on TFLN and SiN/Si Photonics
We develop time-multiplexed 4x4 SiN/Si homodyne crossbars and amplitude-only TFLN electro-optics in a heterogenous integrated platform for photonic AI processing. Matrix-multiplication achieves >5-bit precision at GS/s and 10s fJ/symbol encoding.
- 1 January 2025
- Published by Optica Publishing Group
- p. SS126_2
VCSEL Optical Neural Networks for High-throughput AI Training
We demonstrate large-scale VCSEL-based computing for convolution neural networks with spatial fanout and high clockrates. We achieved 9 parallel channels with >5 bits precision at 100-million shot-per-second, with in-system training over 93.4% accuracy.
- 1 January 2025
- Published by Optica Publishing Group
- p. SS126_2
VCSEL Optical Neural Networks for High-throughput AI Training
We demonstrate large-scale VCSEL-based computing for convolution neural networks with spatial fanout and high clockrates. We achieved 9 parallel channels with >5 bits precision at 100-million shot-per-second, with in-system training over 93.4% accuracy.
- 1 January 2025
- Published by Optica Publishing Group
- p. JPS200_210
Single-shot in-sensor optical spectral AI processing
We demonstrate an analogue frequency-comb-based optical smart sensor with in-memory computing for real-time image classification and blood analysis. Our system successfully predicts early-stage heart-attack samples with 95.1% accuracy with single-pulse latency of 10 ns.
- 1 January 2025
- Published by Optica Publishing Group
- p. JPS200_210
Single-shot in-sensor optical spectral AI processing
We demonstrate an analogue frequency-comb-based optical smart sensor with in-memory computing for real-time image classification and blood analysis. Our system successfully predicts early-stage heart-attack samples with 95.1% accuracy with single-pulse latency of 10 ns.
- Open Access
- 23 July 2024
- Published by Springer Nature in Research Square
Hypermultiplexed Integrated-Photonics-based Tensor Optical Processor
The escalating data volume and complexity resulting from the rapid expansion of artificial intelligence (AI), internet of things (IoT) and 5G/6G mobile networks is creating an urgent need for energy-efficient, scalable computing hardware. Here we demonstrate a hypermultiplexed integrated-photonics-based tensor optical processor (HITOP) that performs trillions of operations per second (TeraOPS) at the energy cost of 25 femtojoule per operation (25 fJ/OP). Based on space-time-wavelength three-dimensional (3D) data streaming, HITOP is built with arrays of wafer-fabricated III/V-based micron-scale lasers (spanning ~1 THz) incorporating thin-film Lithium-Niobate electro-optic (EO) photonics. Multidimensional parallelism allows matrix-matrix multiplications (N3 operations) using O(N) encoding devices, facilitating scalable on-chip integration with O(N2) computations per clockcycle. With each device activating 10 billion parameters per second, the HITOP scalability is validated in machine learning models with 405,000 parameters, which is 25,000 times more than previous integrated optical systems. A combination of high clockrates (10 GS/s), parallel processing and real-time reprogrammability unlocks the full potential of light for next-generation AI accelerators in applications ranging from training with trillions of parameters, real-time decision making in autonomous vehicles and robotics, dynamic optimization in smart manufacturing, to complex simulation for climate modeling and drug discovery.
- 23 July 2024
- Published by Springer Nature in Research Square
Open AccessHypermultiplexed Integrated-Photonics-based Tensor Optical Processor
The escalating data volume and complexity resulting from the rapid expansion of artificial intelligence (AI), internet of things (IoT) and 5G/6G mobile networks is creating an urgent need for energy-efficient, scalable computing hardware. Here we demonstrate a hypermultiplexed integrated-photonics-based tensor optical processor (HITOP) that performs trillions of operations per second (TeraOPS) at the energy cost of 25 femtojoule per operation (25 fJ/OP). Based on space-time-wavelength three-dimensional (3D) data streaming, HITOP is built with arrays of wafer-fabricated III/V-based micron-scale lasers (spanning ~1 THz) incorporating thin-film Lithium-Niobate electro-optic (EO) photonics. Multidimensional parallelism allows matrix-matrix multiplications (N3 operations) using O(N) encoding devices, facilitating scalable on-chip integration with O(N2) computations per clockcycle. With each device activating 10 billion parameters per second, the HITOP scalability is validated in machine learning models with 405,000 parameters, which is 25,000 times more than previous integrated optical systems. A combination of high clockrates (10 GS/s), parallel processing and real-time reprogrammability unlocks the full potential of light for next-generation AI accelerators in applications ranging from training with trillions of parameters, real-time decision making in autonomous vehicles and robotics, dynamic optimization in smart manufacturing, to complex simulation for climate modeling and drug discovery.
- 27 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 42 (22) , 7795-7806
Netcast: Low-Power Edge Computing With WDM-Defined Optical Neural Networks
This paper analyzes the performance and energy efficiency of Netcast, a recently proposed optical neural-network architecture designed for edge computing. Netcast performs deep neural network inference by dividing the computational task into two steps, which are split between the (cloud) server and (edge) client: (1) the server employs a wavelength-multiplexed modulator array to encode the network's weights onto an optical signal in an analog time-wavelength basis, and (2) the client obtains the desired matrix-vector product through modulation and time-integrated detection. The simultaneous use of wavelength multiplexing, broadband modulation, and integration detection allows large neural networks to be run at the client by effectively pushing the energy and memory requirements back to the server. The performance and energy efficiency are fundamentally limited by crosstalk and detector noise, respectively. We derive analytic expressions for these limits and perform numerical simulations to verify these bounds.
- 27 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 42 (22) , 7795-7806
Netcast: Low-Power Edge Computing With WDM-Defined Optical Neural Networks
This paper analyzes the performance and energy efficiency of Netcast, a recently proposed optical neural-network architecture designed for edge computing. Netcast performs deep neural network inference by dividing the computational task into two steps, which are split between the (cloud) server and (edge) client: (1) the server employs a wavelength-multiplexed modulator array to encode the network's weights onto an optical signal in an analog time-wavelength basis, and (2) the client obtains the desired matrix-vector product through modulation and time-integrated detection. The simultaneous use of wavelength multiplexing, broadband modulation, and integration detection allows large neural networks to be run at the client by effectively pushing the energy and memory requirements back to the server. The performance and energy efficiency are fundamentally limited by crosstalk and detector noise, respectively. We derive analytic expressions for these limits and perform numerical simulations to verify these bounds.
- Top Cited PapersOpen Access
- 6 March 2024
- Published by Springer Nature in Nature
- Vol. 627 (8003) , 289-294
Near-ultraviolet photon-counting dual-comb spectroscopy
Ultraviolet spectroscopy provides unique insights into the structure of matter with applications ranging from fundamental tests to photochemistry in the Earth’s atmosphere and astronomical observations from space telescopes1–8. At longer wavelengths, dual-comb spectroscopy, using two interfering laser frequency combs, has become a powerful technique capable of simultaneously providing a broad spectral range and very high resolution9. Here we demonstrate a photon-counting approach that can extend the unique advantages of this method into ultraviolet regions where nonlinear frequency conversion tends to be very inefficient. Our spectrometer, based on two frequency combs with slightly different repetition frequencies, provides a wide-span, high-resolution frequency calibration within the accuracy of an atomic clock, and overall consistency of the spectra. We demonstrate a signal-to-noise ratio at the quantum limit and an optimal use of the measurement time, provided by the multiplexed recording of all spectral data on a single photon-counter10. Our initial experiments are performed in the near-ultraviolet and in the visible spectral ranges with alkali-atom vapour, with a power per comb line as low as a femtowatt. This crucial step towards precision broadband spectroscopy at short wavelengths paves the way for extreme-ultraviolet dual-comb spectroscopy, and, more generally, opens up a new realm of applications for photon-level diagnostics, as encountered, for example, when driving single atoms or molecules.
- 6 March 2024
- Published by Springer Nature in Nature
- Vol. 627 (8003) , 289-294
Top Cited PapersOpen AccessNear-ultraviolet photon-counting dual-comb spectroscopy
Ultraviolet spectroscopy provides unique insights into the structure of matter with applications ranging from fundamental tests to photochemistry in the Earth’s atmosphere and astronomical observations from space telescopes1–8. At longer wavelengths, dual-comb spectroscopy, using two interfering laser frequency combs, has become a powerful technique capable of simultaneously providing a broad spectral range and very high resolution9. Here we demonstrate a photon-counting approach that can extend the unique advantages of this method into ultraviolet regions where nonlinear frequency conversion tends to be very inefficient. Our spectrometer, based on two frequency combs with slightly different repetition frequencies, provides a wide-span, high-resolution frequency calibration within the accuracy of an atomic clock, and overall consistency of the spectra. We demonstrate a signal-to-noise ratio at the quantum limit and an optimal use of the measurement time, provided by the multiplexed recording of all spectral data on a single photon-counter10. Our initial experiments are performed in the near-ultraviolet and in the visible spectral ranges with alkali-atom vapour, with a power per comb line as low as a femtowatt. This crucial step towards precision broadband spectroscopy at short wavelengths paves the way for extreme-ultraviolet dual-comb spectroscopy, and, more generally, opens up a new realm of applications for photon-level diagnostics, as encountered, for example, when driving single atoms or molecules.
- Top Cited Papers
- 17 July 2023
- Published by Springer Nature in Nature Photonics
- Vol. 17 (8) , 723-730
Deep learning with coherent VCSEL neural networks
- 17 July 2023
- Published by Springer Nature in Nature Photonics
- Vol. 17 (8) , 723-730
Top Cited PapersDeep learning with coherent VCSEL neural networks
- Top Cited Papers
- 21 October 2022
- Published by American Association for the Advancement of Science (AAAS) in Science
- Vol. 378 (6617) , 270-276
Delocalized photonic deep learning on the internet’s edge
Advanced machine learning models are currently impossible to run on edge devices such as smart sensors and unmanned aerial vehicles owing to constraints on power, processing, and memory. We introduce an approach to machine learning inference based on delocalized analog processing across networks. In this approach, named Netcast, cloud-based “smart transceivers” stream weight data to edge devices, enabling ultraefficient photonic inference. We demonstrate image recognition at ultralow optical energy of 40 attojoules per multiply (100 watts) cloud computers.
- 21 October 2022
- Published by American Association for the Advancement of Science (AAAS) in Science
- Vol. 378 (6617) , 270-276
Top Cited PapersDelocalized photonic deep learning on the internet’s edge
Advanced machine learning models are currently impossible to run on edge devices such as smart sensors and unmanned aerial vehicles owing to constraints on power, processing, and memory. We introduce an approach to machine learning inference based on delocalized analog processing across networks. In this approach, named Netcast, cloud-based “smart transceivers” stream weight data to edge devices, enabling ultraefficient photonic inference. We demonstrate image recognition at ultralow optical energy of 40 attojoules per multiply (100 watts) cloud computers.