Method for implementing all-optical nonlinear activation function based on silicon-based micro-ring resonator

Abstract

The application discloses a kind of based on silicon-based micro-ring resonator's all-optical nonlinear activation function implementation method, and the different resonant spectra are obtained based on different light wave output ports by the micro-ring resonator of uplink and downlink, and the different resonant spectra are respectively divided into regions by the resonant wavelength of activated nonlinear silicon-based micro-ring resonator and the resonant wavelength of unactivated nonlinear silicon-based micro-ring resonator, and the radial bias nonlinear function, Gauss nonlinear function, Asymptotic nonlinear function and Sigmoid nonlinear function are respectively constructed by input power and output power in different regions.The optical neural network with ultrafast time response and ultra-low energy consumption is realized.

Description

Technical Field

[0001] This invention belongs to the field of nonlinear optics, specifically relating to a method for implementing an all-optical nonlinear activation function based on a silicon-based microring resonator. Background Technology

[0002] In emerging Internet of Things (IoT) systems, big data analytics requires powerful computing capabilities and higher energy efficiency. Due to limitations imposed by Dennard scaling technology, device integration in electronic circuits, such as microprocessor chips, progressed steadily at the pace of Moore's Law. As Moore's Law neared its end, processor clock frequencies plateaued after 2004. Multiple processors then helped maintain steady throughput growth through parallel computing. However, according to Amdahl's Law, the speed of parallel computing is ultimately limited, and parallel computing cannot solve all problems.

[0003] Due to the aforementioned challenges faced by electronic devices, optical neural networks have developed rapidly. The high bandwidth and low latency processing capabilities of photonics, combined with the distributed processing of artificial neural networks, are uniquely suited for real-time processing previously impossible with electronic devices. The basic unit of a photonic neural network consists of linear matrix-vector multiplication and nonlinear activation. Nonlinear activation in neuromorphic photonic hardware can be implemented through photoelectric conversion or all-optical methods. Given that even state-of-the-art optoelectronic hardware still faces bandwidth and speed limitations in photoelectric (OEO) conversion, while optical nonlinear activation offers advantages such as high speed and high integrability, optical nonlinearity holds significant value.

[0004] The main structure in most existing on-chip integrated waveguide optical neural network and other on-chip optical computing platform designs consists of beam-splitter waveguide units and cascaded Mach-Zehnder interferometers (MZI), which are only suitable for linear calculations. Since the superposition of linear calculations is still linear, the above design, regardless of the total number of layers, yields a result equivalent to a single matrix multiplication operation, limiting the parameter range and failing to meet the data fitting requirements of neural networks. Therefore, such optical neural networks rely on further nonlinear calculations in electronic circuits during subsequent information processing, and cannot integrate the complete neural network functionality onto the optical platform. Currently, optical nonlinear materials have weak nonlinear responses and insufficient signal modulation depth; furthermore, there is often an inherent trade-off between ultrafast response time and large nonlinearity, meaning that large nonlinear coefficients usually come at the cost of slower response times. Therefore, it is difficult to achieve efficient and feasible nonlinear calculations during the porting of large-scale optical hardware platforms, thus optical nonlinear activation layers remain only theoretical concepts and cannot be implemented in real-world applications.

[0005] The difficulty in performing nonlinear calculations on optical platforms stems from the weak all-optical nonlinearity of materials. There is a lack of sufficiently strong nonlinear materials, making it difficult to achieve strong nonlinear effects in on-chip integrated devices. Furthermore, while GST materials, which can be used for nonlinear calculations, possess non-volatile properties, they are also unsuitable for high-efficiency computations with fast response times. Therefore, the actual introduction of all-optical nonlinear activation layers remains a pressing issue for achieving ultrafast time response and ultra-low power consumption in optical neural networks on hardware platforms. Summary of the Invention

[0006] This invention provides a method for implementing an all-optical nonlinear activation function based on a silicon-based microring resonator. This method can easily achieve the all-optical nonlinear activation function by adjusting the input light wavelength and optical power range.

[0007] A method for implementing an all-optical nonlinear activation function based on a silicon-based microring resonator, wherein the silicon-based microring resonator is a microring resonator with up and down channels, including an input terminal, a drop terminal, and a through terminal; when the input power is W1, the nonlinearity of the silicon-based microring resonator is not activated; when the input power is increased from W1 to W2, the nonlinearity of the silicon-based microring resonator is activated; when the light wave is output from the through terminal and the input power increases from W1 to W2, fitting the input power with the corresponding output power can obtain the first nonlinear function;

[0008] When the input power increases from W1 to W3, the nonlinearity of the silicon-based microring resonator is further excited. When the light wave is output from the Drop end, the input power increases from W1 to W3. By fitting the input power with the corresponding output power, the second nonlinear function can be obtained.

[0009] When the input power is W1, the resonant wavelength of the silicon-based microring resonator is λ1. When the input power is increased to W2, the microring resonator is nonlinearly excited, the spectrum shifts, and the resonant wavelength of the silicon-based microring resonator changes from λ1 to λ2. When the wavelength of the light wave is in the first interval near λ1, the input power and output power of the light wave exhibit a GeLU nonlinear function.

[0010] When the wavelength of the light wave is located in the second interval between λ1 and λ2, the input power and output power of the light wave exhibit a radial bias nonlinear function.

[0011] When the wavelength of the light wave is located in the third interval near λ2, the input power and output power of the light wave exhibit Gaussian nonlinear functions.

[0012] When the input power is increased to W3, the nonlinearity of the microring resonator is excited, the spectrum shifts, and the resonant wavelength of the silicon-based microring resonator changes from λ1 to λ3. When the wavelength of the light wave is in the fourth wavelength range near λ1, the input power and output power of the light wave exhibit an Asymptotic nonlinear function.

[0013] When the wavelength of the light wave is within the fifth wavelength range between λ1 and λ3, the input power and output power of the light wave exhibit a Sigmoid nonlinear function.

[0014] When the wavelength of light is greater than λ3, the input power and output power of the light wave exhibit a power function.

[0015] The silicon-based material in a silicon-based microring resonator is silicon. Silicon material exhibits nonlinear losses through the Kerr effect, two-photon absorption effect, free carrier dispersion, or free carrier absorption effect.

[0016] When W1=0.1mW, W2=1mW, W3=2mW;

[0017] When λ1 = 1547.7623nm, λ2 = 1547.7806nm, and λ3 = 1547.7968nm;

[0018] The first interval of the Through end is λ1-0.0023nm, λ1+0.0047nm; the second interval is λ1+0.0077nm, λ2-0.0026nm; the third interval is λ2, λ2+0.0014nm.

[0019] The fourth interval of the drop end is λ1-0.0423nm, λ1+0.0047nm; the fifth interval is λ1+0.0177nm, λ3-0.0092nm.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] This invention utilizes microring resonators with uplink and downlink channels to obtain different resonance spectra based on different optical output ports. By activating and deactivating the resonant wavelengths of the nonlinear silicon-based microring resonators, different regions are defined for each resonance spectrum. Within each region, radial bias, Gaussian, asymptotic, and sigmoid nonlinear functions are constructed using input and output power, respectively. This enables an optical neural network with ultrafast time response and ultra-low power consumption. Attached Figure Description

[0022] To more clearly illustrate the prior art and the present invention, the accompanying drawings used in the description of the prior art and the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other drawings from the provided drawings without any creative effort.

[0023] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed herein.

[0024] Figure 1 The diagram shows the structure of the micro-ring for uplink and downlink communication in a specific implementation.

[0025] Figure 2 The spectrum of the through end of the silicon-based microring is provided for a specific implementation at 0.1mW (unexcited nonlinearity) and 1mW (excited nonlinearity).

[0026] Figure 3 The output port provided for the specific implementation method is the Through port, which yields three types of nonlinear curves.

[0027] Figure 4 The output port provided for the specific implementation is the Through port, and the nonlinear function curve is obtained by fitting three types of nonlinear curves.

[0028] Figure 5 The spectrum of the drop end of the silicon-based microring is provided for specific implementation embodiments at 0.1mW (unexcited nonlinearity) and 2mW (excited nonlinearity).

[0029] Figure 6 The output port provided for the specific implementation method is the Drop terminal, which yields three types of nonlinear curves.

[0030] Figure 7 The output port provided for the specific implementation is the Drop terminal, and the nonlinear function curve is obtained by fitting three types of nonlinear curves. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0032] This embodiment extracts the nonlinear function by... Figure 1 The diagram shows the micro-loop structure of the up and down channels.

[0033] The silicon-based material of the silicon-based microring resonator is silicon, according to... Figure 2 The microring transmission spectra obtained when light is output from the through end at 0.1mW (nonlinearity not excited) and 1mW (nonlinearity excited) have resonant wavelengths of λ1 = 1547.7623nm and λ2 = 1547.7806nm, respectively. Due to the presence of a third-order nonlinear effect, a redshift occurs in the visible spectrum. The spectrum is divided into three regions: Region I, with the linear resonant wavelength near λ1 (1547.760nm-1547.767nm); Region II, between λ1 and λ2 (1547.770nm-1547.778nm); and Region III, to the right of λ2 (1547.7806nm-1547.782nm). Nonlinear functions can also be obtained outside these regions.

[0034] according to Figure 3 Therefore, when the port is selected at the Through end, with a fixed wavelength, the input power is gradually increased from 0.1mW to 1mW, and the output power at different input powers is recorded. Connecting the points yields the input-output nonlinear curve. When three wavelengths are selected from the three regions respectively, three categories of nonlinear functions for GELU, radial bias, and Gauss can be obtained.

[0035] according to Figure 4 Therefore, the curves obtained from the through end are subjected to nonlinear fitting, and the formulas for GELU, radial bias, and Gauss are as follows:

[0036] y=(0.9991x-0.0409) / (1.16687+e^(-4.08249x+1.38812)),R 2 =0.9999;

[0037] y=-0.02862+1.3292x-6.11838x^2+9.78334x^3-4.43319x^4,R 2 =0.99504;

[0038] y=-0.26616+0.64275 / (w(0.5π)^0.5)*e^(-2(x-0.55709)^2 / w^2),w=0.93872,R 2 =0.99766.

[0039] Where R 2The coefficients are linear regression coefficients; the closer they are to 1, the higher the curve fit. It can be seen that the nonlinear function obtained from the microloop has a very high degree of matching with GELU, radial bias, and Gauss.

[0040] according to Figure 5 The microring transmission spectra at 0.1mW (nonlinearity not excited) and 2mW (nonlinearity excited) when light is output from the Drop end are obtained, with resonant wavelengths of λ1 = 1547.7623nm and λ3 = 1547.7968nm, respectively. The spectrum is divided into two regions: Region I is near λ1, with a wavelength range of 1547.72nm-1547.770nm; Region II is between λ1 and λ3, with a wavelength range of [missing information]. Nonlinear functions can also be obtained outside these regions, but their shapes do not match existing function types. This range was mainly obtained through debugging.

[0041] according to Figure 6 Therefore, when the port is selected at the Drop end, the input power is gradually increased from 0.1mW to 2mW, and the output power under different input powers is recorded. Connecting the points yields the input-output nonlinear curve. By selecting a wavelength from each of the two regions, the Sigmoid and Asymptotic functions can be obtained. The right side of λ3 yields a power function, but it cannot be used for nonlinear activation and is therefore not considered.

[0042] according to Figure 7 Therefore, the curve fitting formulas obtained from the Drop end, namely the Sigmoid and Asymptotic formulas, are as follows:

[0043] y=0.75769-0.759*0.26093x, R 2 =0.99968;

[0044] y=1.57944 / (1+e^(-3.41955(x-0.90652))), R 2 =0.99751.

[0045] Among them, R 2 The coefficients represent linear regression; the closer they are to 1, the better the curve fit. It is evident that the functions obtained from dropping different regions show a high degree of similarity to Sigmoid and Asymptotic regression, providing significant potential for their application in neural networks.

Claims

1. A method for implementing an all-optical nonlinear activation function based on a silicon-based microring resonator, characterized in that, The silicon-based microring resonator is a microring resonator with up and down channels, including an input terminal, a drop terminal, and a through terminal; When the output port is selected at the Through end and the wavelength is fixed, the input power is gradually increased from 0.1mW to 1mW, and the output power under different input powers is recorded. Connecting the points yields the input-output nonlinear curve. When three wavelengths are selected from the three regions respectively, three categories of nonlinear functions of GELU, radial bias, and Gauss can be obtained. Micro-ring transmission spectra were obtained at 0.1 mW (power when nonlinearity is not excited) and 1 mW (power when nonlinearity is excited) when light is output from the through end, with resonant wavelengths of λ1 = 1547.7623 nm and λ2 = 1547.7806 nm, respectively. Due to the presence of third-order nonlinear effects, the visible spectrum exhibits a redshift, dividing the spectrum into three regions: Region I is located near the linear resonant wavelength λ1, with a wavelength range of 1547.760 nm - 1547.767 nm; Region II is between λ1 and λ2, with a wavelength range of 1547.770 nm - 1547.778 nm; Region III is to the right of λ2, with a wavelength range of 1547.7806 nm - 1547.782 nm. When the output port is selected at the Through end and the wavelength is fixed, the input power is gradually increased from 0.1mW to 1mW, and the output power under different input powers is recorded. Connecting the points yields the input-output nonlinear curve. When three wavelengths are selected from the three regions respectively, three categories of nonlinear functions of GELU, radial bias, and Gauss can be obtained. Alternatively, when light is output from the Drop end, the micro-ring transmission spectra at 0.1mW (non-nonlinear excitation power) and 2mW (nonlinear excitation power) are obtained, with resonant wavelengths of λ1 = 1547.7623nm and λ3 = 1547.7968nm, respectively. The spectrum is divided into two regions: Region I is near λ1, with a wavelength range of 1547.72nm-1547.770nm; Region II is between λ1 and λ3. When the port is selected at the Drop end, the input power is gradually increased from 0.1mW to 2mW, and the output power at different input powers is recorded. Connecting the points yields the input-output nonlinear curve. When a wavelength is selected from each of the two regions, the Sigmoid and Asymptotic functions can be obtained. The power function can be obtained on the right side of λ3, but it cannot be used for nonlinear activation.

2. The method for implementing the all-optical nonlinear activation function based on a silicon-based microring resonator according to claim 1, characterized in that, The silicon-based material in a silicon-based microring resonator is silicon. Silicon material exhibits nonlinear losses through the Kerr effect, two-photon absorption effect, free carrier dispersion, or free carrier absorption effect.

Images