Fig. 1. Basic architecture of analog photonic computing system

Fig. 2. Implementation schemes of analog photonic computing under different dimensions. (a) Spatial dimension; (b) wavelength dimension; (c) time dimension; (d) 3D spatial light dimension

Fig. 3. Generalized analog photonic computing for different computing functions

Fig. 4. Optical matrix computing architectures. (a) MZI architecture based on orthogonal matrix decomposition^[17]; (b) MZI architecture suitable for complex number computing^[18]; (c) optical dot product kernel architecture^[20]; (d) MZI based cross-bar architecture^[25]; (e) MZI architecture based on IQ modulator^[26]; (f) MZI architecture based on time integration^[28]; (g) MRR weight bank based on broadcast-and-weight^[39]; (h) MRR based cross-bar architecture^[40]; (i) architecture based on optical frequency comb and PCM^[41]; (j) matrix inversion architecture based on Richardson method^[37]

Fig. 5. Optical tensor convolution architectures. (a) Convolution architecture based on wavelength-time stretching principle^[44]; (b) convolution architecture based on MRR and fixed delay line^[47]; (c) convolution architecture based on optical frequency comb and fixed delay line^[48]; (d) convolution architecture based on RF multiplexing^[49]; (e) convolution architecture based on multimode interferometer and fixed delay line^[50]; (f) convolution architecture based on waveguide mode converter^[51]

Fig. 6. Programmable signal processor architectures. (a) Quadrangular architecture^[56]; (b) hexagonal architecture^[57]; (c) triangular architecture^[58]

Fig. 7. Optical neuromorphic computing architectures. (a) Gain selection scheme based on VCSEL-SA^[60]; (b) injection locking scheme based on microdisk laser^[65]; (c) mode selection scheme based on VCSEL^[71]; (d) material phase change threshold scheme based on GST MRR^[72]; (e) optical-electrical hybrid scheme^[73]

Fig. 8. Optical nonlinear activation function schemes. (a) Electro absorption effect^[78]; (b) microring modulator^[79]; (c) injection-locked DFB scheme^[82]; (d) carrier dispersion effect scheme^[83]; (e) thermo-optic effect scheme^[84]; (f) structural nonlinear scheme^[85]

Fig. 9. High-precision computation and parameter calibration techniques. (a) Feedforward calibration and balanced detection technique^[86]; (b) full-link jitter monitoring scheme^[89]; (c) Kramers-Kronig phase retrieval algorithm^[91]; (d) fully analog closed-loop PID controller^[92]; (e) precise control of pulse amplitude and duration parameters^[93]; (f) integration of PCM (Sb₂Se₃) with PIN diode-based microring resonator^[97]

Fig. 10. Specialized analog photonic computing for different computing functions

Fig. 11. Specified analog photonic computing for vision sensing systems. (a) Programmable 4f system^[101]; (b) optical lens neural network^[102]; (c) spatial diffractive neural network^[103]; (d) all-analog opto-electronic chip based on spatial diffractive neural network^[16]; (e) end-to-end photonic deep neural network chip^[79]; (f) integrated diffractive optical network chip^[119]

Fig. 12. Specified analog photonic computing for optical communication systems. (a) Nonlinear dispersion compensation technique for optical communication^[120]; (b) dynamic compensation technique for nonlinear dispersion in optical fiber^[121]; (c) determining optimal bidirectional optical communication channels of arbitrary scattering optical systems^[122]

Fig. 13. Specified analog photonic computing for RF systems. (a) Wireless channel estimation technique^[125]; (b) blind source separation technique^[130]; (c) terahertz topology photonic integration technique^[131]; (d) super-resolution direction-of-arrival estimation technique^[132]; (e) radar signal pulse compression technique^[133]; (f) radar signal feature extraction technique^[134]

Fig. 14. Specified analog photonic computing for coherent Ising machines. (a) OPO network Ising machine^[141]; (b) large-scale OPO network Ising machine^[142]; (c) integrated MZI mesh Ising machine^[144]; (d) orthogonal optical spatial Ising machine^[145]

Fig. 15. Onlinetraining methods of analog photonic computing. (a) Physics-aware training (PAT) algorithm^[150]; (b) hybrid training method^[151]; (c) asymmetric estimation training (AsyT) method^[154]; (d) pump-probe scheme for gradient approximation^[149]; (e) MZI-based in situ backpropagation^[157]; (f) bidirectional propagation in linear and nonlinear layers^[158]; (g) simultaneous perturbation stochastic approximation (SPSA) method using randomized directional derivatives^[160]; (h) G-MFO algorithm^[161]; (i) DFA algorithm^[167]; (j) physical local training (PhyLL) algorithm^[168]