CN-117872533-B - Reconfigurable all-optical nonlinear activation function on silicon integrated platform

Abstract

The application relates to a reconfigurable all-optical nonlinear activation function on a silicon integrated platform. The present application provides systems, devices, and methods for implementing all-optical reconfigurable activation devices for various activation functions using low input optical power. The devices and systems disclosed herein include a directional coupler including a first phase shift mechanism and an interferometer coupled to the directional coupler. The interferometer includes at least one microring resonator and a second phase shift mechanism coupled thereto. The interferometer and directional coupler include a waveguide formed of a first material, and the microring resonator includes a waveguide formed of a second material and a third phase shift mechanism. The second material is provided as a low loss material with a high kerr effect and a large band gap to produce various nonlinear activation functions. The first, second, and third phase shift mechanisms are configured to control the bias within the disclosed systems and devices to achieve a desired activation function.

Inventors

• PENG YIWEI
• YUAN YUAN
• S.Zhang

Assignees

• 慧与发展有限责任合伙企业

Dates

Publication Date

20260505

Application Date

20230714

Priority Date

20221010

Claims (17)

1. 1. An optical device, comprising: a directional coupler comprising a first phase shift mechanism; an interferometer coupled to the directional coupler, the interferometer comprising: A first branch including a first waveguide formed of a first material and a second waveguide formed of a second material different from the first material, A second branch including a third waveguide formed of the first material, and A second phase shift mechanism coupled to the first waveguide, and A microring resonator coupled to the second waveguide, the microring resonator formed from the second material and including a third phase shift mechanism, Wherein the first, second and third phase shift mechanisms are configured to control the bias of the optical device to achieve a desired activation function, Wherein the second material comprises at least one of a band gap exceeding 1.9eV, a nonlinear refractive index greater than 3x10 -20 m 2 /W, and a linear loss coefficient less than 200 dB/m.
2. 2. The optical device of claim 1, wherein the second material is aluminum gallium arsenide AlGaAs.
3. 3. The optical device of claim 1, wherein the second material is tantalum pentoxide Ta 2 O 5 .
4. 4. The optical device of claim 1, wherein a ratio of a nonlinear refractive index to a linear loss coefficient of the second material is greater than 3x10 -20 .
5. 5. The optical device of claim 1, wherein the first waveguide comprises a splicing end coupled to a splicing end of the second waveguide, the optical device further comprising: a first pair of reverse tapers comprising a first taper at a junction end of the first waveguide and a second taper at a junction end of the second waveguide, wherein the first taper is in an opposite direction relative to the second taper.
6. 6. The optical device of claim 1, wherein at least the third phase shift mechanism comprises at least one of a hetero-Metal Oxide Semiconductor (MOS) phase shifter and a heater.
7. 7. The optical device of claim 1, wherein at least the third phase shift mechanism comprises a heterogeneous metal oxide semiconductor MOS capacitor.
8. 8. The optical device of claim 1, wherein the directional coupler is a mach-zehnder coupler and the interferometer is a mach-zehnder interferometer.
9. 9. A nonlinear-enabled device comprising: a Mach-Zehnder coupler MZC, the MZC including a first phase shift mechanism; A Mach-Zehnder interferometer MZI coupled to the MZC, the MZI including a first branch including a first waveguide formed of a first material and a second waveguide formed of a second material different from the first material, and a second phase shift mechanism coupled to the first waveguide, and A microring resonator MRR coupled to the second waveguide, the microring resonator formed from the second material, Wherein the first phase shift mechanism and the second phase shift mechanism are configured to control the bias of the nonlinear-active device to switchably achieve a desired activation function, Wherein the second material comprises at least one of a band gap exceeding 1.9eV, a nonlinear refractive index greater than 3x10 -20 m 2 /W, and a linear loss coefficient less than 200 dB/m.
10. 10. The nonlinear-active device in accordance with claim 9, wherein a ratio of the nonlinear refractive index to the linear loss coefficient of the second material is greater than 5x10 -20 .
11. 11. The nonlinear activation device in accordance with claim 9, wherein the second material is aluminum gallium arsenide AlGaAs.
12. 12. The nonlinear activation device in accordance with claim 9, wherein the second material is tantalum pentoxide Ta 2 O 5 .
13. 13. The nonlinear-active device in accordance with claim 9, wherein the first waveguide comprises a first junction end coupled to a second junction end of the second waveguide, the nonlinear-active device further comprising a first reverse taper pair comprising a first taper at the first junction end of the first waveguide and a second taper at the second junction end of the second waveguide, wherein the first taper is in an opposite direction relative to the second taper.
14. 14. The nonlinear-active device in accordance with claim 13, wherein the first waveguide comprises a third junction end coupled to a fourth junction end of the second waveguide, the nonlinear-active device further comprising a second pair of inverted cones comprising a third cone at the third junction end of the first waveguide and a fourth cone at the fourth junction end of the second waveguide, wherein the third cone is opposite relative to the fourth cone and opposite relative to the first cone.
15. 15. A method for operating a nonlinear-enabled device, the method comprising: Tuning a first bias of a mach-zehnder interferometer MZI by controlling a first phase shift mechanism of the MZI to tune a phase difference between branches of the MZI, wherein the first phase shift mechanism is coupled to a first waveguide of a first branch of the MZI; Adjusting a second bias of a second waveguide coupled to the first branch of the MZI by controlling a second phase shift mechanism of the MRR of a microring resonator such that the nonlinear-active device operates at about a resonant frequency of the MRR, the MRR being formed of a second material including at least one of a band gap exceeding 1.9eV, a nonlinear refractive index greater than 3x10 -20 m 2 /W, and a linear loss coefficient less than 200dB/m, and Adjusting a third bias of the MZC coupled to the MZI by controlling a third phase shift mechanism of the Mach-Zehnder coupler MZC to tune the amplitude of the branches of the MZI relative to each other, Wherein the first bias, the second bias, and the third bias are controlled to achieve a desired activation function, an Wherein the first waveguide is formed of a first material and the second waveguide is formed of the second material different from the first material.
16. 16. The method of claim 15, wherein the first phase shift mechanism, the second phase shift mechanism, and the third phase shift mechanism each comprise one of a hetero-metal oxide semiconductor, MOS, phase shifter or a heater.
17. 17. The method of claim 15, wherein the second material of the MRR is one of aluminum gallium arsenide AlGaAs and tantalum pentoxide Ta 2 O 5 .

Description

Reconfigurable all-optical nonlinear activation function on silicon integrated platform Background The global artificial neural network market is expected to grow at a significant rate, driven by the increasing interest in Artificial Intelligence (AI). Artificial Neural Networks (ANNs) and learning algorithms have the ability to learn from large data sets, which can create a machine with human-like decision making capabilities with low latency and high energy efficiency. Neuromorphic photonics exhibits improved performance in multiplexing, energy dissipation, and cross-talk compared to electronic systems, which is advantageous for dense and high bandwidth interconnects. Thus, neuromorphic photonic systems are likely to provide several orders of magnitude faster operation speeds and higher efficiencies than neuromorphic electronic systems. ANNs are computing systems inspired by biological neural networks and consist of a collection of connected nodes or neurons. Neurons include linear weighting, summation, and nonlinear activation, which is a building block in an ANN and which enables complex mapping between inputs and outputs of learning tasks. Several nonlinear functions, such as sigmoid, radial basis, rectified linear unit (ReLU), and quadratic functions, are widely used in ANNs for different machine learning tasks. Drawings The present disclosure in accordance with one or more various embodiments is described in detail with reference to the following figures. The drawings are provided for purposes of illustration only and depict only typical or example embodiments. FIG. 1 illustrates a model of an example nonlinear neuron including synapses, weighted addition, and nonlinear activation functions in accordance with embodiments disclosed herein. Fig. 2 depicts a schematic diagram of an all-optical nonlinear-activation device in accordance with embodiments disclosed herein. Fig. 3A and 3B depict example resonant cavities included in the all-optical nonlinear-enabled device of fig. 2, according to example embodiments. Fig. 3C depicts an example mode transition within the example resonant cavity of fig. 3A and 3B. Fig. 4A-4C depict another example resonant cavity included in the all-optical nonlinear-active device of fig. 2 in accordance with an example embodiment. Fig. 4D depicts an example mode transition within the example resonant cavity of fig. 4A-4C. Fig. 5A-9C depict graphical representations of various normalized nonlinear activation functions as a function of input optical power in accordance with embodiments disclosed herein. Fig. 10A and 10B illustrate an example phase shift mechanism including a Metal Oxide Semiconductor Capacitor (MOSCAP) according to an embodiment of the present disclosure. FIG. 11 is an example computing component that may be used to implement various features of the all-optical nonlinear-enabled device in accordance with embodiments disclosed herein. FIG. 12 is an example computer system that may be used to implement various features of the all-optical nonlinear-enabled device of the present disclosure. The drawings are not intended to be exhaustive and do not limit the disclosure to the precise forms disclosed. Detailed Description As described above, ANN and machine learning algorithms have the ability to learn from large data sets to create human-like machines. The neurons of an ANN consist of linear weighting, summing, and nonlinear activation of inputs, where nonlinear activation implements a complex mapping between inputs and outputs for learning. Examples of nonlinear activation functions include, but are not limited to, logic functions, radial basis functions, linear rectification functions (such as ReLU, inverse ReLU, and leakage ReLU), and quadratic functions, each of which is used for signal processing of a different machine learning task. Various nonlinear activation functions are suitable for different tasks in neural networks and machine learning applications. For example, the ReLU function may provide a solution to the problem of nonlinear optimization with constraints and may be used in feed forward machine learning networks, such as multi-layer perceptrons and convolutional neural networks. Other examples include radial basis functions for support vector machine based multilayers and quadratic functions for modeling higher order polynomial neural networks. With the development of nonlinear optics, some all-optical methods have been proposed to implement the activation function. However, optical nonlinearities are relatively weak, so that all optically activated devices typically require high threshold power and large light injection (e.g., inputting an optical signal into the device). Another technical drawback of the conventional plenoptic method is that the activation device is typically stationary after manufacture and thus is not configurable to implement different activation functions. That is, conventional activation devices are typically manufactured for