Laser & Optoelectronics Progress, Volume. 62, Issue 17, 1739002(2025)
Physical Architecture and Application for Optical Computing (Invited)
Fig. 1. Evolution of computation required for training large models of AI over time[5]
Fig. 2. Computational principles of free-space diffractive architectures. (a) Construction of specific vectorial PSF based on convolutional computation methods[25]; (b) angular spectrum calculation employing non-uniform fast Fourier transform (NUFFT) algorithm[28]; (c) V-shaped metallic structures with distinct resonant frequencies for validating the generalized Snell's law[45]; (d) full-color holograms based on 8th-order geometric phase[48]; (e) 8 bit depth hologram construction using reflective Malus meta-surface[53]
Fig. 3. Implementation and optimization of D2NN. (a) D2NN performing classification tasks including handwritten digit recognition[29]; (b) optimization of D2NN's expressive capability through Fresnel number control[32]; (c) D2NN processing high bit-depth grayscale image computational tasks[33]; (d) F-D2NN achieving saliency extraction[34]
Fig. 4. Two typical types of programmable units for on-chip optical computing. (a) Schematic of a single-channel MRR structure[78]; (b) reconfigurable optical switch based on MRR[78]; (c) transmission spectra in block/pass modes[78]; (d) schematic of a dual-channel MRR structure[95]; (e) programmable optical information processing chip composed of six MRR[95]; (f) reconfigurable logic computing chip based on a multi-wavelength architecture[81]; (g) schematic of a MZI structure[102]; (h) on-chip optical neural network supporting
Fig. 5. Implementation of nonlinear activation methods in free-space diffractive optical computing. (a) Unit cell structure of surface-normal photodetector[123]; (b) implementation of incoherent neuron modulation array using transparent phototransistor and liquid crystal modulator[124]; (c) nonlinear activation implementation based on perovskite quantum dots film[129]; (d) minimal unit structure and operation schematic of electrically tunable nonlinear polaritonic meta-surface[130]
Fig. 6. Implementation of nonlinear activation methods in on-chip integrated optical computing. (a) EAM integrated on silicon waveguide[135]; (b) photodiode array electronic chip serving as nonlinear layer for optoelectronic computing[137]; all-optical nonlinear implementation approaches using electromagnetically induced transparency structure (c) and reverse saturable absorption through fullerene (d)[138]; (e) hybrid structure with MRR and MZI implementing Clamped ReLU and Sigmoid nonlinear activation functions[139]
Fig. 7. Differential operators in non-neural-network optical computing tasks. (a) Implementation of unitary matrix construction with halved complexity using beam splitters and phase shifters[99]; (b) three-dimensional data processing architecture based on PCM[117]; (c) planar photonic chip enabling high-order optical differentiation[147]; (d) quantitative phase gradient imaging through meta-surface implementation[151]
Fig. 11. Multi-task reconfigurable optical neural networks and backpropagation-free neural network training. (a) Reconfigurable diffractive processing unit constructing a diffractive neural network[21]; (b) backpropagation-free deep physical neural network[168]; (c) large-scale photonic chip Taichi with co-designed diffractive and interferometric components[108]; (d) multi-task photonic chip based on in-memory optical computing[171]
Get Citation
Copy Citation Text
Shengbing Guo, Wenzhe Liu, Jiajun Wang, Minjia Zheng, Lei Shi. Physical Architecture and Application for Optical Computing (Invited)[J]. Laser & Optoelectronics Progress, 2025, 62(17): 1739002
Category: AI for Optics
Received: May. 21, 2025
Accepted: Jun. 30, 2025
Published Online: Sep. 11, 2025
The Author Email: Jiajun Wang (jiajunwang@fudan.edu.cn), Minjia Zheng (mjzheng@fudan.edu.cn), Lei Shi (lshi@fudan.edu.cn)
CSTR:32186.14.LOP251285