ACS Photonics, Journal Year: 2024, Volume and Issue: unknown
Published: Sept. 18, 2024
Language: Английский
Advanced Materials, Journal Year: 2024, Volume and Issue: unknown
Published: July 14, 2024
Abstract The demand for accurate perception of the physical world leads to a dramatic increase in sensory nodes. However, transmission massive and unstructured data from sensors computing units poses great challenges terms power‐efficiency, bandwidth, storage, time latency, security. To efficiently process data, it is crucial achieve compression structuring at terminals. In‐sensor integrates perception, memory, processing functions within sensors, enabling terminals perform structuring. Here, vision are adopted as an example discuss electronic, optical, optoelectronic hardware visual processing. Particularly, implementations devices in‐sensor that can compress structure multidimensional information examined. underlying resistive switching mechanisms volatile/nonvolatile their operations explored. Finally, perspective on future development provided.
Language: Английский
Citations
23
Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)
Published: March 30, 2024
Language: Английский
Citations
16
ACS Photonics, Journal Year: 2024, Volume and Issue: 11(2), P. 723 - 730
Published: Jan. 10, 2024
With the rapid development of Internet Things, how to efficiently store, transmit, and process massive amounts data has become a major challenge now. Optical neural networks based on nonvolatile phase change materials (PCMs) have breakthrough point due their zero static power consumption, low thermal crosstalk, large-scale, high efficiency. However, current photonic devices cannot meet multilevel requirements in neuromorphic computing limited storage capacity. Here, new way for increasing capacity is paved from perspective modulation crystallization kinetics PCMs. A more progressive transition amorphous crystalline states achieved through grain-refinement phenomenon induced by nitrogen (N) element doping Ge2Sb2Te5 (GST), giving precise access multibit states. By integrating N-doped (N-GST) with waveguide, high-capacity device enabling over 7 bits (∼222 levels) first time. The increased nearly times compared state-of-the-art (∼32 levels). An optical convolutional network successfully established MINIST handwritten digit recognition task mapping synapse weight multiple levels, accuracy up 96.5% achieved. Our work provides strategy integrated demonstrates enormous application potential field large-scale networks.
Language: Английский
Citations
13
Nanophotonics, Journal Year: 2024, Volume and Issue: 13(12), P. 2183 - 2192
Published: Jan. 12, 2024
In the development of silicon photonics, continued downsizing photonic integrated circuits will further increase integration density, which augments functionality chips. Compared with traditional design method, inverse presents a novel approach for achieving compact devices. However, compact, reconfigurable devices that employs modulation method exemplified by thermo-optic effect poses significant challenge due to weak capability. Low-loss phase change materials (PCMs) Sb
Language: Английский
Citations
12
Nature Reviews Electrical Engineering, Journal Year: 2024, Volume and Issue: 1(6), P. 358 - 373
Published: June 6, 2024
Language: Английский
Citations
12
Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)
Published: July 23, 2024
Abstract Multimodal deep learning plays a pivotal role in supporting the processing and of diverse data types within realm artificial intelligence generated content (AIGC). However, most photonic neuromorphic processors for can only handle single modality (either vision or audio) due to lack abundant parameter training optical domain. Here, we propose demonstrate trainable diffractive neural network (TDONN) chip based on on-chip optics with massive tunable elements address these constraints. The TDONN includes one input layer, five hidden layers, output forward propagation is required obtain inference results without frequent optical-electrical conversion. customized stochastic gradient descent algorithm drop-out mechanism are developed neurons realize situ fast convergence achieves potential throughput 217.6 tera-operations per second (TOPS) high computing density (447.7 TOPS/mm 2 ), system-level energy efficiency (7.28 TOPS/W), low latency (30.2 ps). has successfully implemented four-class classification different modalities (vision, audio, touch) achieve 85.7% accuracy multimodal test sets. Our work opens up new avenue integrated processors, providing solution low-power AI large models using technology.
Language: Английский
Citations
12
International Journal of Extreme Manufacturing, Journal Year: 2024, Volume and Issue: 6(3), P. 035501 - 035501
Published: Feb. 22, 2024
Abstract Multi-level programmable photonic integrated circuits (PICs) and optical metasurfaces have gained widespread attention in many fields, such as neuromorphic photonics, communications, quantum information. In this paper, we propose pixelated Si 3 N 4 PICs with record-high 20-level intermediate states at 785 nm wavelength. Such flexibility phase or amplitude modulation is achieved by a Sb 2 S matrix, the footprint of whose elements can be small 1.2 μ m, limited only diffraction limit an in-house developed pulsed laser writing system. We believe our work lays foundation for laser-writing ultra-high-level (20 levels even more) systems based on change materials, which could catalyze diverse applications biosensing, computing, reconfigurable metasurfaces.
Language: Английский
Citations
8
Applied Physics Reviews, Journal Year: 2024, Volume and Issue: 11(1)
Published: Jan. 31, 2024
Every multi-input multi-output linear optical system can be deemed as a matrix multiplier that carries out desired transformation on the input information, such imaging, modulation, and computing. The strong programmability of has been explored proved to able bring more flexibility greater possibilities applications signal processing general digital analog Furthermore, burgeoning integrated photonics with advanced manufacturing light manipulating technology pave way for large-scale reconfigurable photonic coherent matrix. This paper reviews programmable in platform. First, theoretical basis optimizing methods three types (Mach–Zehnder interferometer mesh, multi-plane diffraction, crossbar array) are introduced. Next, we overview configuring method this their processing, neural network, logic operation, recurrent acceleration, quantum computing comprehensively reviewed. Finally, challenges opportunities discussed.
Language: Английский
Citations
7
Optics & Laser Technology, Journal Year: 2024, Volume and Issue: 176, P. 111005 - 111005
Published: April 12, 2024
This theoretical modeling and simulation paper presents designs projected performances of two non-volatile, broadband, on-chip 2 × electro-optical switches based upon the germanium-on-insulator (GeOI) photonic-electronic platform operating at 2.5 µm mid-infrared wavelength. These compact devices facilitate large-scale integration on a "monolithic wafer" where all components are made group-IV semiconductors. The two-waveguide directional coupler (DC) Mach-Zehnder interferometer (MZI). A thin-film graphene Joule-effect micro-heater is assumed planarized GeOI device to change phase (reversably) DC-slot-embedded Sb2Se3 phase-change material (PCM) from crystalline amorphous. MZI has this PCM within its slotted-arm waveguides. Simulations show high-performance bistable or multi-stable cross-bar switching in both devices. DC an active coupling length 17 µm, 130 nm gap, footprint 5 x 31 µm. bandwidth 30 over wavelength range cross bar insertion losses IL less than 0.3 dB, optical crosstalk −15 dB. Results for crossbar attained with 7.8 µm-length slot 51 switch footprint. Stable, multi-level via partial amorphization. Thermal shows that careful control voltage-pulse amplitude V applied (rectangular pulse duration 500 ns) can give 32 levels, example, using 6.18 7.75 Volts. Multi-level shown also PCM-based ring resonators.
Language: Английский
Citations
7
Optica, Journal Year: 2024, Volume and Issue: 11(8), P. 1039 - 1039
Published: July 8, 2024
Integrated photonic neural networks (PNNs) are at the forefront of AI computing, leveraging light’s unique properties, such as large bandwidth, low latency, and potentially power consumption. Nevertheless, integrated optical components inherently sensitive to external disturbances, thermal interference, various device imperfections, which detrimentally affect computing accuracy reliability. Conventional solutions use complicated control methods stabilize devices chip, result in high hardware complexity impractical for large-scale PNNs. To address this, we propose a training approach enable control-free, accurate, energy-efficient without adding complexity. The core idea is train parameters physical network towards its noise-robust region. Our method validated on different PNN architectures applicable solve imperfections thermally tuned PNNs based phase change materials. A notable 4-bit improvement achieved micro-ring resonator-based needing complex or power-hungry temperature stabilization circuits. Additionally, our reduces energy consumption by tenfold. This advancement represents significant step practical, energy-efficient, noise-resilient implementation
Language: Английский
Citations
7