Acta Optica Sinica, Volume. 45, Issue 14, 1420006(2025)
Recent Research Advances of On‐Chip Optical Nonlinear Activation Function Devices (Invited)
Figures & Tables(13)
Fig. 1. Introduction to deep learning neural network. (a) Architecture of deep learning neural network; (b) three classical nonlinear activation functions: Sigmoid function, hyperbolic tangent (Tanh) function, and rectified linear unit (ReLU) function; (c) typical applications of several deep learning neural networks; (d) performance comparison between networks with and without activation functions in classification tasks
Fig. 2. Technical pathways for optically activated functional devices. The green section presents all-optical activation function implementation schemes, categorized based on different material systems; the blue section illustrates electro-optical activation function implementation schemes, classified according to different detection approaches
Fig. 3. Electro-optical activation functions based on detection-modulation schemes. (a) Activation function scheme using photodiodes for optical signal detection and EAM for modulation[66]; (b) activation function scheme employing photodiodes for optical signal detection and MZI modulators for modulation[70]; (c) activation function scheme utilizing photodiodes for optical signal detection and microring resonator modulators for modulation[34]
Fig. 4. Activation function scheme using MoS2 optoelectronic RAM switches combined with anelectrical control unit for optical signal detection and MZI modulators for modulation[71]. (a) Schematic diagram of the activation function structure; (b) implemented activation function curves
Fig. 5. Integrated detection-modulation activation function scheme implemented using graphene-silicon heterogeneously integrated microring resonators[73]. (a) Neural network architecture employing the integrated detection-modulation activation function scheme; (b)(c) relationships of photocurrent and transmittance versus input optical power and bias voltage for the two activation function devices; (d) five activation functions derived from the two devices, corresponding to different photocurrent contour levels
Fig. 6. All-optical nonlinear activation functions implemented using nonlinear effects in silicon materials. (a) All-optical nonlinear activation function achieved through free-carrier dispersion effect[80]; (b) nonlinear activation function realized via Kerr effect through inverse design[83]; (c) nonlinear activation function implemented utilizing free-carrier absorption and two-photon absorption effects[84]
Fig. 8. All-optical nonlinear activation function implemented using second-order nonlinear effects in lithium niobate (LiNbO₃) materials[91]
Fig. 10. All-optical nonlinear activation function implemented using nonlinear effects in erbium-doped optical amplifiers [96]
Fig. 11. All-optical nonlinear activation functions implemented using nonlinear effects in phase-change materials. (a) Nonlinear activation function realized through pulsed optical excitation of GST phase transition[30]; (b) nonlinear activation function achieved via spatially modulated optical excitation of GST phase transition[99]; (c) nonlinear activation function implemented using continuous-wave optical excitation of VO₂ phase transition[97]
Fig. 12. All-optical nonlinear activation functions implemented using nonlinear effects in two-dimensional materials. (a) Nonlinear activation function realized via graphene saturable absorption effect in waveguide devices[105]; (b) nonlinear activation function achieved through graphene saturable absorption effect in plasmonic devices[106]; (c) nonlinear activation function implemented utilizing graphene saturable absorption effect in silicon-plasmonic hybrid devices[107]; (d) nonlinear activation function demonstrated by MXene saturable absorption effect in waveguide devices[108]; (e) nonlinear activation function established via MoTe₂ saturable absorption effect in glass waveguide devices[109]
Table 1. A summary of the on-chip optical activation function devices reviewed in this paper
View tableTable 1. A summary of the on-chip optical activation function devices reviewed in this paper
Implementation
method
Wavelength /
nm
Bias
voltage /
V
Threshold
power
Reconfigu-rability Response
time /
ns
Device
size /
μm2
NAF
curves
Task
demon-
stration
Experiment/
Simulation
Accuracy /%
（Architecture）
Electro-optic nonlinear activation function PD+MRR［ 34 ］1550 0.8 0.06 mW Yes 0.10 ‒ Multi-AFs Vowel classification Experiment 92.50（FC） PD+MRR/MZI［ 63 ］1550 1.5 0.60 mW Yes ‒ >100×
200
ReLU/Sigmoid MNIST Simulation 95.71（CNN） PD+MRR［ 64 ］1550 0.9 2.50 mW No 0.57 >90×30 ReLU Handwritten letters classification Experiment 93.80（CNN） PD+MRR［ 65 ］1550 ‒ ‒ Yes 1 50000 6AFs ‒ ‒ ‒ PD+EAM［ 66 ］1550 2 5 mW No ‒ 20×0.8 Inverse Sigmoid MNIST Simulation 98（FC） PD+EAM［ 67 ］1550 20 7 mW No ‒ 20×0.5 ReLU MNIST Simulation 95（FC） PD+MZI［ 70 ］1550 12.8 0.60 mW Yes ‒ ‒ ReLU/Sigmoid MNIST Simulation 93（FC） ECU+ORS+MZI［ 71 ］405/520 2 7 mW Yes ‒ ‒ Sigmoid/Softplus/Clamped ReLU MNIST Simulation 91.60（FC） MRR PD/modulator［ 73 ］2000 1 0.0080 mW Yes ‒ 80×80 Multi-AFs MNIST/ CIFAR-10 Simulation >95/>70（CNN） All-optic nonlinear activation function SOI TO FCD MRR+MZI［ 80 ］1550 ‒ 25 mW Yes 2.50 575×48 Sigmoid/Softplus/Clamped ReLU XOR classification/MNIST Simulation 100/94（FC） Si3N4 Kerr MRR+MZI［ 81 ］1550 ‒ ‒ Yes 0.10 ‒ ReLU/Inverse ReLU/Quadratic ‒ ‒ ‒ SOI Kerr inverse design［ 83 ］1550 0 2.90 mW No ‒ 5×5 ‒ MNIST Simulation 93.25（FC） SOI TO TPA Kerr FCA FCD MRR［ 84 ］1550 0 0.080‒2.95 mW Yes 1.20×104 20×20 4AFs MNIST Simulation 98（CNN） Ge-Si PD FCA［ 88 ］1550 4 1.10 mW No 0.050 30×8 Clamped ReLU MNIST Experiment 97.30（FC） Ge-Si WG FCA［ 90 ］1550 0 5.10 mW No 14.30 32×2 Clamped ReLU MNIST Simulation 96.80（FC） Ge-Si MRR FCA［ 89 ］1550 0 0.74 mW Yes 104 30×15 Radial basis/ReLU/ELU MNIST Simulation 94.80（FC） LiNbO3 WG SHG DOPA［ 91 ］I：1045
O：2090
0 16 fJ No 7.50×10-5 2500×1.7 ReLU MNIST Simulation 99.13（CNN） SOA-MZI+XGM-WC［ 95 ］1550 ‒ 8.09 mW No 0.40 ‒ Sigmoid ‒ ‒ ‒ Fano cavity laser［ 93 ］‒ 0 600 fJ No 0.30×10-3 100 Sigmoid ‒ ‒ ‒ Er∶Si3N4
amplifier［
96 ］1550 ‒ 0.070 mW No ‒ 1200×3600 Clamped ReLU ‒ ‒ ‒ PCM GST MRR［ 30 ］1550 0 5×105 fJ No 200 100×100 ReLU Language recognition Experiment 99.60（FC） PCM GST MRR［ 99 ］1550 0 5.586×
1012 fJ
No 1 ‒ Sigmoid MNIST Simulation 94.50（FC） PCM VO2
Si3N4 WG［
97 ］Broad 0 0.50 mW No 5×103 5 ELU CIFAR-10 Simulation 80‒85
（CNN）
Graphene+WG［ 105 ］1550 0 10 mW No 0.090 40×10 Clamped ReLU Eigenvalues of a matrix Experiment 78.8（FC） Graphene+plasmonic WG［ 106 ］1550 0 35 fJ No 2.60×10-4 10×10 ReLU ‒ ‒ ‒ Graphene+plasmonic/Si WG［ 107 ］1550 0 5.5 mW No 0.094 100 ReLU MNIST Simulation 92.58（CNN） 2D MXene+
Si WG［
108 ］1310/1550 0 0.05 mW No ‒ 4×50 ReLU MNIST Simulation 99.10（FC） 2D MoTe2+glass WG［ 109 ］Broad 0 9.4×
10-4 mW
No 4.80×10-4 50 Clamped ReLU MNIST/ CIFAR-10 Simulation 97.60/94.60（FC）
Tools
Get Citation
Copy Citation Text
Ruizhe Liu, Zijia Wang, Hongtao Lin. Recent Research Advances of On‐Chip Optical Nonlinear Activation Function Devices (Invited)[J]. Acta Optica Sinica, 2025, 45(14): 1420006
Paper Information
Category: Optics in Computing
Received: Apr. 15, 2025
Accepted: Jun. 30, 2025
Published Online: Jul. 22, 2025
The Author Email: Hongtao Lin (hometown@zju.edu.cn)
CSTR:32393.14.AOS250928
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20