An ultra-low insertion loss non-volatile photonic synapse device

Abstract

The application discloses an ultra-low insertion loss non-volatile photonic neural synapse device and belongs to the field of optical computing. The device comprises a substrate, a Bragg grating and a heating area. The Bragg grating is composed of a phase change material and a waveguide. The phase change material is periodically arranged in the waveguide. The heating area is located outside the Bragg grating and covers the phase change material and the waveguide. The fundamental mode in the waveguide is periodically reflected when passing through the Bragg grating. The fundamental mode satisfying the phase matching condition can realize low-loss transmission. The crystallization degree of the phase change material in the Bragg grating is adjusted to realize the modulation of the transmittance of the waveguide. Further, the periodicity, duty cycle and length of the modulation area of the phase change material in the Bragg grating are adjusted to realize that the device has the largest normalized intensity output range in different phase states (crystalline state and amorphous state).

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of optical computing, and more particularly, relates to a super-low insertion loss non-volatile photonic neuron synapse device. BACKGROUND

[0002] In recent years, artificial intelligence technology represented by neural networks has rapidly developed towards high speed and low power consumption. However, the von Neumann architecture adopted by electronic integrated chips (ICs) separates program space from data space, resulting in a large amount of tidal data load between the computing unit and the storage unit. Frequent data read-write operations not only reduce the computing speed but also increase the energy consumption of single computation. Therefore, neural networks based on electronic IC chips are difficult to further improve power efficiency and computing speed, which restricts the development and application of artificial intelligence technology. Non-volatile neural networks enable the computing unit to have data storage capability. The proposal of this in-memory computing architecture greatly improves the computing speed of neural networks and reduces data access energy consumption. At the same time, non-volatile photonic neural networks that take advantage of the high speed and low energy consumption of photonic computing have become an important research content of artificial neural network chips. Non-volatile neuron synapses are the core devices of non-volatile photonic neural networks. The working principle is to control the output optical intensity of the neuron synapse and store optical information based on the non-volatile optical state of phase change materials (PCMs). The existing non-volatile neuron synapses use PCMs including vanadium dioxide, liquid crystal, and GST, GSST, and other chalcogenide phase change materials. By covering a few tens of nanometers of GST film on the high refractive index waveguide, feldmann et al. successfully designed the first optical non-volatile neural network. However, due to the weak interaction between the phase change material and the high refractive index waveguide mode, the output intensity range of the non-volatile neuron synapse is small. In addition, the phase change material used has large inherent optical loss at the wavelength of the signal source, resulting in large insertion loss of the non-volatile neuron synapse. These two drawbacks result in a limited weight pool of the designed artificial neuron, which cannot achieve 0 / 1 modulation.

[0003] After that, a large number of scholars optimize the insertion loss and output intensity range of non-volatile neural synapses by optimizing the waveguide structure, trying to design a non-volatile neural synapse closer to the ideal state. In 2021, domestic researchers proposed that the groove ridge waveguide has a larger modulation range than the ordinary silicon-based ridge waveguide by analyzing the waveguide mode. On this basis, Qiang Zhiqiang et al. further analyzed the interaction between the waveguide mode and the phase change material, and proposed a non-volatile neural synapse based on the sub-wavelength groove ridge grating waveguide (SWGSR) structure, with an insertion loss of 1.1 dB and an output intensity range of 10 dB. At the same time, the extinction ability of non-volatile neural synapses in the crystalline state can be further improved using the principle of resonant coupling, including traditional micro-ring resonators and racetrack micro-ring resonators. The emergence of GSST material is a key step for low insertion loss non-volatile neural synapses, and its inherent optical loss is much smaller than that of GST and vanadium dioxide phase change materials. Researchers use the interaction between the micro-ring structure and the GSST material to control the resonant coupling state of the micro-ring by the GSST in different phase states, and then control the output optical response of the micro-ring, achieving a non-volatile neural synapse with an insertion loss of 0.5 dB and an output intensity range of 42 dB. However, it has extremely strict requirements for the working wavelength, and the 3dB bandwidth of this performance parameter is only 0.2nm. And limited by the size of the resonant structure, this structure is not conducive to large-scale integration applications. Although depositing GSST thin film directly on the sidewall of high refractive index waveguide can reduce the size of non-volatile neural synapses, it is limited by the smaller inherent optical contrast of GSST material, which cannot effectively modulate the high refractive index waveguide mode, and thus cannot achieve a larger optical modulation range. Qiang Zhiqiang et al. combined the characteristics of GSST phase change material and proposed a hybrid waveguide structure based on phase change material. The designed silicon-based hybrid waveguide has an insertion loss of 0.4 dB, an output intensity range of 17 dB, a bandwidth of 120 nm, and a very small device size under TM mode excitation. In addition, by using the differential idea, a differential non-volatile neural synapse based on a micro-ring can be designed, which has a larger output intensity range. Similarly, in 2021, foreign researchers designed a non-volatile neural synapse based on mode differential output by using the idea of mode conversion on the high refractive index waveguide side based on phase change material. However, this non-volatile neural synapse still cannot further reduce the insertion loss.

[0004] Sb2Se3 has smaller insertion loss and optical contrast compared to GSST. Further enhancing the interaction between the optical signal and the material is needed to construct a non-volatile neural synapse using Sb2Se3 material. Previous studies have used the principle of multimode interference (MMI) to realize a non-volatile two-port optical switching device, but the modulation region is too large and requires high-precision control of the phase change pixel array using femtosecond lasers, which cannot be applied to practical scenarios. Recent studies have shown that by using hybrid nanowaveguides, polarization and intensity-controlled non-volatile neural synapses can be achieved in a short distance. However, the output adjustable range of this non-volatile neural synapse is too small. SUMMARY

[0005] In view of the defects of the prior art, the purpose of the present application is to provide an ultra-low insertion loss non-volatile photonic neural synapse device, aiming to solve the problem that the insertion loss of the photonic neural synapse is too large and cannot be applied to deep neural networks.

[0006] To achieve the above-mentioned purpose, the present application provides an ultra-low insertion loss non-volatile photonic neural synapse device, comprising a substrate, a Bragg grating and a heating area, the Bragg grating is arranged above the substrate, the Bragg grating is composed of a phase change material and a waveguide, the phase change material is embedded in the waveguide and is periodically arranged, and the heating area is located outside the Bragg grating and covers the phase change material and the waveguide in the area. After the input light passes through the waveguide, it enters the photonic neural synapse device for transmission. Different cross-sectional geometric parameters of the waveguide will affect the distribution of the fundamental mode light field in the waveguide. The fundamental mode light field is periodically reflected when passing through the Bragg grating, and the fundamental mode that satisfies the phase matching condition can realize low-loss transmission. Adjusting the crystallization degree of the phase change material in the Bragg grating realizes the modulation of the transmittance of the waveguide.

[0007] Further, by optimizing the period, duty cycle and modulation region length of the phase change material in the Bragg grating to achieve the largest normalized intensity output range of the device in different phase states (crystalline state, amorphous state), an ultra-low insertion loss non-volatile photonic neural synapse device is realized. The actual optimization of the cross-sectional geometric parameters of the waveguide and the Bragg grating is related.

[0008] As preferred, the structure of the ultra-low insertion loss non-volatile photonic neural synapse device is a strip structure, and the initial cross-sectional geometric parameters are: the strip waveguide height is 220 nm, and the width is 500 nm. By changing these parameters, when the phase change material is in different phases, the effective mode refractive index imaginary part difference of the mode in the photonic neural synapse is larger, and the regulation range of the photonic neural synapse output intensity by the phase change material is larger. When the refractive index of the phase change material is equal to the refractive index of the waveguide material, the optical field energy can be transmitted in the waveguide with ultra-low insertion loss; when the phase change material is in a crystalline state, most of the optical field energy needs to be lost during transmission, so as to achieve the purpose of regulating the output intensity of the photonic neural synapse.

[0009] Further, in the working wavelength range, the phase change material has low optical loss characteristics in the amorphous state, such as GSST and Sb2Se3.

[0010] Further, when the phase change material is in different crystallization degrees, the dynamic change range of the real part of the refractive index of the phase change material includes the refractive index of the waveguide.

[0011] Further, the period and duty cycle of the Bragg grating satisfy the condition that when the phase change material is in a crystalline state, the Bragg grating satisfies the reflection-enhanced phase matching condition, and the optical transmittance of the device is minimum.

[0012] Further, the period of the Bragg grating can be a variable period or a fixed period grating.

[0013] Further, the structure of the waveguide includes a strip structure, a ridge waveguide, and a slot waveguide.

[0014] Further, the heating area can adopt an ITO electric heating method or ion implantation in the waveguide area, and the metal electrode adopts a gold electrode with good thermal conductivity. As preferred, the present application adopts the ITO electric heating method.

[0015] Further, the ultra-low insertion loss non-volatile photonic neural synapse device can be processed on a silicon-on-insulator (SOI) platform or a silicon nitride (SiN), indium phosphide (InP), and lithium niobate (LN) semiconductor platform. As preferred, the present application selects the SOI platform.

[0016] Further, at a wavelength of 1550 nm, the width and height of the waveguide satisfy that a low-loss fundamental mode can be formed in the waveguide.

[0017] The application provides a non-volatile neural synapse with ultra-low insertion loss, which is designed by using Sb2Se3 to design a silicon-based Bragg grating, and by adjusting the coupling effect of the grating structure and different phases (crystalline state, amorphous state) of the phase change material, excellent performance of an insertion loss of only 0.01 dB and an output adjustable range of 30.02 dB is achieved. Unlike the large size of the resonant coupling structure, the size of the ultra-low insertion loss non-volatile neural synapse provided by the application is only 0.5*12 mu 2 Therefore, the application has important significance for the design of high-performance and large-scale non-volatile neural networks. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a transmission cross-section schematic diagram of an ultra-low insertion loss non-volatile photonic neural synapse device.

[0019] Figure 2 is a simulation result diagram of an ultra-low insertion loss non-volatile photonic neural synapse device, (a) and (b) respectively represent the functional relationship between the normalized output intensity of the photonic neural synapse and the Bragg grating period and the duty cycle when the phase change material is in the amorphous state and the crystalline state, and (c) represents the functional relationship between the normalized output intensity difference between different phases and the Bragg grating period and the duty cycle.

[0020] Figure 3 is a relationship curve diagram of the modulation length (Bragg grating period number) and the normalized optical output intensity.

[0021] Figure 4 is a simulation result diagram of an ultra-low insertion loss non-volatile photonic neural synapse device, (a) and (b) respectively represent the normalized transmission electric field distribution when the phase change material is in the amorphous state and the crystalline state.

[0022] Figure 5 is a simulation result diagram of the multi-stage adjustable capability of an ultra-low insertion loss non-volatile photonic neural synapse device. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the application clearer and more understandable, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0024] The application provides an ultra-low insertion loss non-volatile photonic neural synapse device, comprising a substrate, a Bragg grating and a heating area, wherein the waveguide is arranged above the substrate, the Bragg grating is composed of a phase change material and the waveguide, the phase change material is embedded in the waveguide, and the heating area is located outside the Bragg grating and covers the phase change material and the waveguide in area.

[0025] The waveguide is a silicon waveguide with a width of 500 nm and a height of 220 nm, the period of the phase change material in the Bragg grating is 600 nm, the duty cycle is 0.5, and the number of periods is 5, which are initial geometric parameters of the ultra-low insertion loss non-volatile photonic neural synapse device.

[0026] Specifically, the phase change material is Sb2Se3.

[0027] Specifically, the ITO electrode of the heating area covers the Bragg grating, and the metal heating electrode is located on the ITO electrode on both sides of the transmission direction.

[0028] Specifically, the thickness of the ITO electrode and the metal electrode is 50 nm.

[0029] Embodiment

[0030] The application discloses an ultra-low insertion loss non-volatile photonic neural synapse device, comprising a substrate, a Bragg grating and a heating area, wherein the Bragg grating is arranged above the substrate, the Bragg grating is composed of a phase change material and a waveguide, the phase change material is embedded in the high-refractive waveguide, the heating area is located outside the Bragg grating and covers the phase change material and the high-refractive waveguide in area, and the device structure transmission direction section view is as shown in Figure 1 The fundamental mode in the high-refractive waveguide is periodically reflected when passing through the Bragg grating, wherein the fundamental mode satisfying the phase matching condition can realize low-loss transmission, the crystallization degree of the phase change material in the Bragg grating is adjusted to realize the modulation of the transmittance of the high-refractive waveguide, and the period, the duty cycle and the length of the modulation region of the phase change material in the Bragg grating are further adjusted to realize that the device has the largest normalized intensity output range in different phase states (crystalline state, amorphous state).

[0031] For the traditional modulation method of laying phase change material outside the high refractive index waveguide, the energy attenuation change caused by the phase change of the phase change material is not obvious, because the interaction between the optical field and the phase change material depends on the proportion of the leaked energy of the optical field, and if the proportion of the leaked energy of the optical field is increased, the insertion loss of the energy will be increased, therefore, the non-volatile photonic neural synapse device with ultra-low insertion loss has not been reported. The non-volatile photonic neural synapse device with ultra-low insertion loss provided by the present application can significantly reduce the insertion loss of the device, and still has a large dynamic output range when the phase change material is in different phase states. The geometry of the high refractive index waveguide cross section, such as the flat layer height, the ridge type region height and the ridge waveguide width, will affect the real part and the imaginary part of the effective mode refractive index of the fundamental mode, and then affect the periodic reflection of the high refractive index waveguide fundamental mode in the Bragg grating. Similarly, the non-volatile photonic neural synapse device with the lowest insertion loss and the largest output modulation range can be determined by scanning the period and the duty cycle in the Bragg grating. For the non-volatile photonic neural synapse provided by the present application, the structure optimization effect is analyzed and quantified by monitoring the normalized output intensity. The change of the normalized output intensity difference (ΔT) is obtained by using the three-dimensional finite difference time domain (FDTD) method, so as to quantify the multi-level adjustable ability of the non-volatile photonic neural synapse device.

[0032] Specifically, the period and the duty cycle of the phase change material in the Bragg grating are changed to observe the change of the optical output intensity difference when the phase change material is in different phase states. The functional relationship between the normalized output intensity and the period and the duty cycle of the Bragg grating is as shown in Figure 2 It can be found that when the Sb2Se3 is in amorphous state and crystalline state, the normalized output intensity of the non-volatile photonic neural synapse periodically appears maximum value with the change of the period and the duty cycle of the Bragg grating. And when the Bragg grating period is 600 nm and the duty cycle is 0.2, the normalized output intensity difference between different phase states (amorphous state, crystalline state) reaches the maximum value, which is 61.7%. At this time, the number of grating periods is only 5, therefore, the normalized output intensity difference between different phase states can be further improved by increasing the modulation length.

[0033] Specifically, the influence of the modulation distance on the normalized output intensity of the proposed non-volatile photonic neuron synapse device is observed by changing the length of the Bragg grating. In fact, the transmission modulation of the Bragg grating on the high refractive index waveguide fundamental mode satisfies an exponential decay relationship. When the modulation distance increases, the output intensity of the high refractive index waveguide corresponding to the case where the phase change material is in the amorphous state and the crystalline state will decrease, because when the refractive index of the phase change material is not equal to the refractive index of the high refractive index waveguide, the light field will be reflected once every time it passes through the interface between the high refractive index waveguide and the phase change material, thereby introducing additional transmission loss. The relationship between the number of periods of the Bragg grating and the output intensity of the high refractive index waveguide under the condition of the optimal Bragg grating period and duty cycle geometric parameters is shown in Figure 3 . When the number of periods of the Bragg grating is 20, the output intensity difference of the high refractive index waveguide under different phase states reaches a maximum value of about 99.1%. This is because when Sb2Se3 is in the amorphous state, the refractive index of the phase change material is similar to that of the high refractive index waveguide, so the attenuation of the light field along the propagation direction is not obvious; when Sb2Se3 is in the crystalline state, the refractive index of the phase change material is quite different from that of the waveguide, and due to the inherent optical loss of the material, the light field rapidly attenuates along the propagation direction, as shown in Figure 4 .

[0034] Specifically, in the multi-level adjustable test of the ultra-low insertion loss non-volatile photonic neuron synapse, the optical refractive index of the phase change material can be changed by changing the crystallization degree of the phase change material, and the relationship between the crystallization degree and the optical refractive index satisfies:

[0035]

[0036] where p represents the crystallization degree, e a and e c represent the dielectric constant when the phase change material is in the amorphous state and the crystalline state, respectively. As a preferred embodiment, the phase change material used in the present application is Sb2Se3, and the refractive index of the amorphous state and the crystalline state of Sb2Se3 at a wavelength of 1550 nm is 3.3684+0.0001i and 4.2816+0.0003i, respectively. Simulation results show that the multi-level adjustable optical switch can control the light output intensity at 256 different levels, which corresponds to an 8-bit programming resolution, as shown in Figure 5 . The normalized maximum output intensity is 99.7%, and the corresponding insertion loss is 0.01 dB, which is much smaller than the insertion loss of existing photonic non-volatile neuron synapses. And the minimum output transmittance is 9.93x10 -4 under a modulation length of 12 μm, achieving excellent performance with an output intensity range of 30.02 dB.

[0037] Those skilled in the art can easily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An ultra-low insertion loss non-volatile photonic neural synaptic device, characterized in that: The system includes a substrate, a Bragg grating, and a heating zone. The Bragg grating is arranged above the substrate. The Bragg grating is composed of a phase change material and a waveguide. The phase change material is GSST and Sb2Se3. The phase change material is periodically arranged inside the waveguide. The heating zone is located outside the Bragg grating, and its area covers the phase change material and the waveguide. The heating method of the heating zone includes ITO electric heating and ion implantation electric heating. The fundamental mode in the waveguide undergoes periodic reflection when passing through the Bragg grating, and the fundamental mode that meets the phase matching conditions realizes low-loss transmission.

2. The ultra-low insertion loss non-volatile photonic neural synaptic device according to claim 1, characterized in that: When the phase change material has different degrees of crystallinity, the dynamic variation range of the real part of its refractive index includes the refractive index of the waveguide.

3. The ultra-low insertion loss non-volatile photonic neural synaptic device according to claim 1, characterized in that: The period and duty cycle of the Bragg grating satisfy the reflection enhancement phase matching condition when the phase change material is in a crystalline state, and the optical transmittance of the device is minimum.

4. The ultra-low insertion loss non-volatile photonic neural synaptic device according to claim 1 or 3, characterized in that: The Bragg grating is a grating with a variable period or a fixed period.

5. The ultra-low insertion loss non-volatile photonic neural synaptic device according to claim 1, characterized in that: The structure of the waveguide includes a stripe structure, a ridge waveguide, and a slot waveguide.

6. The ultra-low insertion loss non-volatile photonic neural synaptic device according to claim 1, characterized in that: The ultra-low insertion loss non-volatile photonic neural synaptic device is processed on an SOI platform or a silicon nitride SiN, indium phosphide InP, or lithium niobate LN semiconductor platform.

7. The ultra-low insertion loss non-volatile photonic neural synaptic device according to claim 1, characterized in that: At a wavelength of 1550nm, the width and height of the waveguide are sufficient to form a low-loss fundamental mode in the waveguide.

Images