A scalable nonvolatile on-chip optical neural network

Abstract

This invention discloses a scalable non-volatile on-chip optical neural network. The network involves a weighted modulation unit array fabricated on a silicon platform on a standard insulator. Each tunable unit is designed with a corresponding integrated non-volatile / volatile phase shifter. The network training process is based on volatile modulation schemes such as thermo-optical modulation. Non-volatile modulation of the network weight coefficients is achieved using an electrical pulse modulation scheme based on phase change materials. The basic unit device of this invention exhibits only a single resonant mode in its spectral response, a characteristic that significantly increases the number of channels and parallel data throughput of the wavelength division multiplexing optical neural network. The design scheme combining fast volatile tuning with non-volatile phase change memory meets the high-frequency refresh requirements of applications such as neural networks while also addressing the critical issue of energy consumption in optical neural network modulation.

Description

Technical Field

[0001] This invention relates to the technical field of on-chip optical neural networks, specifically a scalable non-volatile on-chip optical neural network. Background Technology

[0002] In recent years, neural networks have been widely used in fields such as autonomous driving, graphics processing, and natural language processing. However, due to the approaching physical limits of transistor size and issues such as the memory wall, the increased computing power gained by simply expanding the scale of traditional hardware is significantly less than the accompanying energy efficiency costs. In contrast, optical neural networks, due to their high speed, high bandwidth, and low power consumption, are considered a new hardware architecture solution that can significantly improve computing speed, increase data throughput, and reduce system power consumption.

[0003] Existing on-chip optical neural network architectures are mainly divided into interferometric modulation networks and wavelength division multiplexing (WDM) networks. Compared to interferometric modulation networks, WDM networks have attracted widespread attention from researchers due to their high integration density, low network loss, and simple training methods. However, one of the key challenges of WDM optical neural networks is that the number of input channels is limited by the free spectral range (FSR) (the typical number of MRR channels on the SOI platform is about 30), which hinders the construction of large-scale optical neural networks and prevents them from performing complex recognition or classification tasks. At the same time, similar to integrated circuits, the expansion of network size leads to a sharp increase in energy consumption. Therefore, how to reduce the static energy consumption of optical networks and address the severe on-chip heat accumulation and thermal crosstalk has become a key research issue for integrated optical neural networks.

[0004] Non-volatile devices based on phase change materials (PCMs) are considered a promising approach for controlling large-scale optical networks due to the lack of a continuous external power supply and the size advantage resulting from the significant difference in refractive index between crystalline and amorphous states. However, the power consumption of non-volatile devices during a single switch is relatively high. Compared to volatile tuning methods such as thermo-optical tuning and carrier injection tuning, which both require energy for bidirectional switching, these devices are unsuitable for frequent and repeated parameter updates during neural network training. Therefore, the large-scale integrated PCM non-volatile optical neural networks face difficulties in practical applications.

[0005] For example, CN212569201U discloses a compact WDM wavelength division multiplexer that achieves low optical path loss, ultra-small module size, and stable function. This compact WDM wavelength division multiplexer includes a substrate assembly; a support layer is disposed on the substrate assembly; multiple layers of wavelength division multiplexing components are disposed on the substrate assembly; and the bottommost wavelength division multiplexing component is mounted on a base plate, while the remaining layers of wavelength division multiplexing components are disposed on the support layer.

[0006] The above technical solution uses a compact WDM wavelength division multiplexer, which greatly reduces the space required and is conducive to the development of network construction towards integration and miniaturization. However, the above technical solution is still limited by the free spectral range (FSR), which hinders the construction of large-scale optical neural networks and still cannot perform complex recognition or classification tasks.

[0007] For example, CN113655565A discloses a polarization beamsplitter based on a Y-branch waveguide structure modulated by phase change materials, belonging to the fields of optical device applications and information processing technology. The polarization beamsplitter includes a substrate, a Y-branch optical waveguide, and a phase change material thin film. The phase change material thin film is deposited on the Y-branch waveguide in a single-layer top-mount and two single-layer side-mount configurations, respectively. The state of the phase change material thin film can be changed through evanescent field coupling in the waveguide, and adjusting the dimensions of the hybrid optical waveguide allows for the separation and output of TM / TE polarization modes. The phase change material thin film exhibits significant differences in optical constants between its crystalline and amorphous states, which can alter the light propagation behavior in the optical waveguide. The difference in effective refractive index allows for the realization of polarization state output and cutoff.

[0008] The aforementioned technical solution, based on Si waveguides and phase-change materials, is a non-volatile ultrafast phase-change polarization beamsplitter that achieves low loss and miniaturization of the device. Its simple design structure facilitates integration, laying the foundation for the future development of all-optical devices. However, due to the relatively high power consumption of non-volatile devices during a single switch, compared to volatile tuning methods such as thermo-optical tuning and carrier injection tuning, which require energy for bidirectional switching, the above technical solution remains unsuitable for performing frequent and repeated parameter updates during the training process of neural networks. This makes the large-scale integration of phase-change non-volatile optical neural networks difficult to advance in practical applications.

[0009] To address the shortcomings of existing technologies, this invention proposes an optical neural network combining volatile and non-volatile tuning schemes. The single-resonant mode characteristic of the unit devices significantly increases the number of channels in the optical neural network, enabling large-scale computation. Simultaneously, this network employs conventional volatile modulation methods to set and modify the weight parameters in real time, while utilizing the non-volatility of phase change materials for a fixed weight parameter training scheme, forming a zero-static-power optical computing network. This greatly alleviates energy efficiency issues in large-scale computing applications. This invention is significant not only for neural networks themselves but also for fields such as autonomous driving, graphics processing, and natural language processing. Summary of the Invention

[0010] The purpose of this invention is to propose a scalable, non-volatile on-chip optical neural network to address the limitations imposed on data channels by the free spectral range in optical computing, as well as the excessive static power consumption and resulting on-chip heat accumulation and severe thermal crosstalk caused by the need for continuous power supply to maintain the optical network state. This architecture aims to significantly improve the integration density and data processing capabilities of the optical neural network. Simultaneously, its unique non-volatile control methods enable its energy efficiency to continuously improve with increasing computational load, thus providing an effective solution to address the massive computing power demands and hardware energy efficiency issues in applications such as neural networks.

[0011] To achieve the above objectives, this invention proposes a scalable, non-volatile on-chip optical neural network, comprising the following components:

[0012] SOI substrate;

[0013] Optical waveguide;

[0014] Phase change materials;

[0015] Drive electrode;

[0016] Dielectric cladding;

[0017] The optical waveguide is disposed on the SOI substrate, and the optical waveguide includes a ridge waveguide and a strip waveguide;

[0018] The phase change material and the driving electrode are disposed on the protrusion of the optical waveguide, and a dielectric cladding is provided between the phase change material and the driving electrode.

[0019] The driving electrode and phase change material can change the effective refractive index of the optical waveguide.

[0020] This application discloses a scalable, non-volatile on-chip optical neural network. The network is constructed by cascading unit devices whose spectral responses exhibit only a single resonant mode characteristic, thereby increasing the number of data channels. The single resonant mode characteristic can be achieved by matching the periodic response of the resonant device with the passband of a bandpass device, or by designing a resonant device with a feed-in reflected signal. The former extracts the target resonant mode by matching the periodic response of the resonant device with the passband of a bandpass device. When the target resonant mode is within the passband and its width is smaller than the free spectral range of the resonant device, other resonant modes are suppressed because they are located within the stopband. In this case, the device's spectral response exhibits only a single dominant resonant mode. The latter utilizes a feed-in reflected signal within the periodic resonant device, where the reflected signal spectrum is zero or close to zero only at the target resonant mode. This characteristic significantly improves the side-mode suppression ratio between the original resonant device and the target resonant mode by suppressing other resonant modes of the original device with the reflected signal. Similarly, the device's spectral response exhibits only a single dominant resonant mode. The spectral response of the unit device contains only a single resonant mode, which can significantly expand the data channels of the optical neural network and significantly improve the data throughput and parallel computing capability of the optical neural network.

[0021] This application discloses a scalable, non-volatile on-chip optical neural network. The unit devices used to construct the neural network feature integrated volatile and non-volatile control within the same device. Parameter updates during the training process of the optical on-chip neural network are achieved through volatile control (thermo-optic or carrier dispersion effects). After training, the parameters are fixed to the target values ​​using phase-change non-volatile control methods. This allows the pre-trained neural network model to perform zero-static-power calculations for the same recognition or classification task in practical applications, and the energy efficiency further improves as the scale of the optical neural network increases.

[0022] By implementing the training process of the neural network through a non-volatile control scheme and finally using the non-volatile nature of the phase change material to fix the training parameters, a static optical neural network can be realized, which can significantly improve the energy efficiency of the neural network.

[0023] Preferably, the driving electrode is a metal electrode, a transparent electrode (two-dimensional material, indium oxide), or doped silicon.

[0024] Preferably, the driving electrode is a doped silicon / doped waveguide, and the doping method is p++-i-p++ type doping, p++-i-n++ doping, n++-i-n++ doping, p++-pip-p++ doping, p++-pin-n++ doping, n++-nin-n++ doping or p++-pn-n++ doping.

[0025] Preferably, the phase change material is a sulfide-based phase change material, which is a compound composed of any two to four elements selected from Ge, Sb, Se, Te, or S, including Ge₂Sb₂Se. x Te 5-x (x=0,1,2,3,4,5), Sb2Se3, Sb2S3, GeTe.

[0026] Preferably, the phase change material may also be doped with non-metallic elements and / or metallic elements, wherein the non-metallic elements include oxygen (O), nitrogen (N), carbon (C), and silicon (Si); and the metallic elements include tin (Sn), silver (Ag), cadmium (Cd), bismuth (Bi), and indium (In).

[0027] Preferably, the dielectric cladding is an oxide dielectric material, including aluminum oxide (Al2O3), silicon oxide (SiO2), zirconium oxide (ZrO2), and hafnium oxide (HfO2).

[0028] Preferably, the unit devices constituting the optical neural network can be completed by matching the periodic response of the resonant device and the passband of the bandpass device, or by feeding a reflected signal into the periodic resonant device.

[0029] Preferably, the unit devices constituting the optical neural network include fast volatile modulation function and non-volatile storage function, and the function of the unit devices can be controlled by the characteristics of external control signals.

[0030] Preferably, the real-time control scheme for the fast volatile control function can be achieved by utilizing nonlinear effects, including thermo-optical effects, carrier depletion, carrier dispersion, as well as Pockel and Kerr effects.

[0031] Preferably, the induction of the phase change material can be achieved by sharing the same driving electrode with the volatility control function, or by using different driving electrodes for the phase change material and the volatility control function respectively.

[0032] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0033] 1. The design of the unit devices constituting the network architecture in this invention ensures that there is only a single resonant mode in the spectral response, which solves the problem of the limitation on the number of data channels in wavelength division multiplexing networks. It is suitable for building large-scale computing networks to realize high-dimensional matrix operations and related applications.

[0034] 2. This invention uses a volatile tuning scheme as a training method and combines it with a non-volatile phase transition control parameter fixing method to achieve a high-energy-efficiency static optical neural network.

[0035] 3. This invention will significantly improve the computing power and energy efficiency of optical networks, providing an effective solution to the two major pain points of insufficient computing power and excessive power consumption in the current high-performance computing fields such as neural networks. It can be widely applied in artificial intelligence fields such as autonomous driving, graphics processing, and natural language processing.

[0036] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0037] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The accompanying drawings are incorporated in and constitute a part of this specification, illustrating embodiments consistent with the present invention, and together with the specification are used to explain the principles of the present invention.

[0038] Figure 1 This is an exemplary architecture diagram of a scalable nonvolatile on-chip optical neural network according to the present invention.

[0039] Figure 2 This is a structural diagram of the phase change material and driving electrode (also a phase change induction electrode) in a scalable non-volatile on-chip optical neural network according to the present invention.

[0040] Figure 3 This is a schematic diagram illustrating the working principle of a unit device in a scalable, non-volatile on-chip optical neural network according to the present invention.

[0041] Figure 4 This is a spectral response diagram of a unit device in a scalable non-volatile on-chip optical neural network according to the present invention.

[0042] In the figure, 1-1 is a standard SOI substrate, 1-2 is an optical waveguide, 2-1 is a phase change material, 2-2 is a driving electrode, and 2-3 is a dielectric cladding. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0044] like Figure 1-4 As shown, a scalable non-volatile on-chip optical neural network includes the following components: an SOI substrate 1-1; an optical waveguide 1-2; a phase change material 2-1; a driving electrode 2-2; and a dielectric cladding 2-3. The optical waveguide 1-2 is disposed on the SOI substrate 1-1, and the optical waveguide 1-2 can be any optical waveguide structure, including ridge waveguides and strip waveguides. The phase change material 2-1 and the driving electrode 2-2 are disposed on the protrusions of the optical waveguide 1-2, and a dielectric cladding 2-3 is provided between the phase change material 2-1 and the driving electrode 2-2. The driving electrode 2-2 and the phase change material 2-1 can change the effective refractive index of the optical waveguide 1-2.

[0045] like Figure 1-3 As shown, the driving electrode 2-2 is a metal electrode, a transparent two-dimensional material, indium oxide, or doped silicon. The driving electrode 2-2 can be any one of titanium nitride (TiN), tungsten metal (W), indium tin oxide (ITO), graphene, or doped silicon.

[0046] The driving electrode 2-2 is used to apply voltage or as a heating electrode;

[0047] If a silicon-doped scheme is adopted, the optical waveguide 1-2 itself can be doped to become a heating electrode, and no additional heating electrode is needed. At this time, the driving electrode 2-2 is a doped waveguide (i.e., the doped optical waveguide 1-2), and the doping method can be any one of p++-i-p++ type doping, p++-i-n++ doping, n++-i-n++ doping, p++-pip-p++ doping, p++-pin-n++ doping, n++-nin-n++ doping, and p++-pn-n++ doping.

[0048] like Figure 1-3As shown, the phase change material 2-1 can be any one of the sulfide-based phase change materials. The phase change material 2-1 is a compound composed of any two to four elements selected from Ge, Sb, Se, Te, or S, including but not limited to Ge2Sb2Se. x Te 5-x Materials such as (x=0,1,2,3,4,5), Sb2Se3, Sb2S3, and GeTe.

[0049] The phase change material 2-1 is also doped with at least one non-metallic element and one metallic element. The non-metallic element includes oxygen (O), nitrogen (N), carbon (C), and silicon (Si); the metallic element includes tin (Sn), silver (Ag), cadmium (Cd), bismuth (Bi), and indium (In).

[0050] like Figure 2 As shown, the dielectric cladding 2-3 is an oxide dielectric material, including aluminum oxide (Al2O3), silicon oxide (SiO2), zirconium oxide (ZrO2), and hafnium oxide (HfO2).

[0051] The implementation scheme of the unit devices constituting the optical neural network in this application can be accomplished by matching the periodic response of the resonant device and the passband of the bandpass device, or by feeding a reflected signal into the periodic resonant device.

[0052] A schematic diagram of a scalable, non-volatile on-chip optical neural network architecture involved in this invention is shown below. Figure 1 As shown, a wavelength division multiplexing (WDM) optical network architecture is formed by matching the periodic response of resonant devices and the passband of bandpass devices. Each device has different structural parameters, ensuring that each device operates in a different signal channel.

[0053] The working principle of the unit device involved in this invention is as follows: Figure 3 As shown, this unit device includes a Fabry-Perot resonant cavity (FP) structure with Bragg gratings on both sides as reflective structures. When the passband width of the Bragg grating is smaller than the free spectral range of the FP cavity, only one resonant mode is located within the passband of the Bragg grating, while other resonant modes are suppressed by the stopband of the Bragg grating. Figure 4 As shown, the device exhibits single-resonance characteristics, which means that the number of channels in the wavelength division multiplexing optical neural network based on this device unit is no longer limited by the resonant period. Therefore, the number of input channels within the entire operating bandwidth is greatly increased, enabling large-scale parallel computing.

[0054] The unit device in this application has both fast volatile regulation function (the state reached by the device after being regulated by the external signal automatically returns to the state before the change after the external control signal stops) and non-volatile storage function (the device automatically maintains the current state after the external control signal stops), and the function under which the device operates is determined by the characteristics of the external control signal.

[0055] The fast volatility real-time control scheme in this application can utilize thermo-optical effects, carrier depletion, carrier dispersion, and nonlinear effects such as the Pockel effect and the Kerr effect; the induction of the phase change material 2-1 can be achieved by sharing the same driving electrode 2-2 with the volatility control scheme, or by using different driving electrodes 2-2 separately.

[0056] The structures of the driving electrode 2-2 and the phase change material 2-1 used for training optical neural networks and inducing phase transitions in this invention are as follows: Figure 2 As shown, the network is thermo-optically tuned by continuously injecting current at a low voltage into the driving electrode 2-2, thereby changing the effective refractive index of the waveguide 1-2, enabling the loading and refreshing of parameters in the neural network, completing real-time general matrix operations, and thus realizing the training process of the neural network. Meanwhile, using excitation signals with relatively high voltage and varying pulse energies to induce different degrees of crystallization in the phase change material 2-1 allows for continuous control of the phase change material's refractive index, which can also affect the effective refractive index of the mode field propagating in the waveguide. A single configuration can fix the pre-trained parameter model for a long time, thus achieving a near-zero static power consumption on-chip optical neural network and significantly improving its energy efficiency.

[0057] The working principle of this application is that the optical neural network completes training based on the volatile control function of each unit device, and stores the network weight coefficients based on the non-volatile control brought about by the phase change of the phase change material. After that, the optical neural network can perform data processing and calculation without external power supply until the network needs to be refreshed and its functions changed.

[0058] In summary, this invention innovatively combines conventional volatile tuning schemes for on-chip photonic devices with emerging non-volatile phase-change tuning schemes, significantly improving the energy efficiency of optical neural networks without affecting their reconstruction efficiency. Furthermore, the number of parallel channels in this invention is no longer limited by the device's free spectral range, which enables a significant increase in the scale of optical neural networks and broadens their application scenarios.

[0059] The above description is only a few specific embodiments of the present invention. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the principle of the present invention. For example, phase change films with other components can be used, other types of semiconductors, photonic crystal materials, two-dimensional materials can be used, or other coating layers can be added to the surface of two-dimensional materials. These improvements and substitutions should also be considered within the scope of protection of the present invention.

[0060] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0061] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

[0062] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A scalable, non-volatile on-chip optical neural network, characterized in that, It includes the following components: SOI substrate (1-1); Optical waveguide (1-2); Phase change materials (2-1); Drive electrode (2-2); Dielectric cladding (2-3); The optical waveguide (1-2) is disposed on the SOI substrate (1-1), and the optical waveguide (1-2) includes a ridge waveguide and a strip waveguide; The phase change material (2-1) and the driving electrode (2-2) are disposed on the protrusion of the optical waveguide (1-2), and a dielectric cladding (2-3) is provided between the phase change material (2-1) and the driving electrode (2-2); The driving electrode (2-2) and the phase change material (2-1) can change the effective refractive index of the optical waveguide (1-2).

2. The scalable, non-volatile on-chip optical neural network according to claim 1, characterized in that, The driving electrode (2-2) is a metal electrode, a transparent electrode, or a doped silicon electrode, and the driving electrode (2-2) is used to apply voltage or as a heating electrode.

3. The scalable, non-volatile on-chip optical neural network according to claim 2, characterized in that, The transparent electrode is a two-dimensional material or indium oxide.

4. The scalable, non-volatile on-chip optical neural network according to claim 2, characterized in that, The driving electrode (2-2) is doped silicon, the optical waveguide (1-2) can be doped, and the doped optical waveguide (1-2) serves as a heating electrode. The doping methods are p++-i-p++ type doping, p++-i-n++ doping, n++-i-n++ doping, p++-pip-p++ doping, p++-pin-n++ doping, n++-nin-n++ doping, or p++-pn-n++ doping.

5. The scalable, non-volatile on-chip optical neural network according to claim 1, characterized in that, The phase change material (2-1) is a sulfide-based phase change material, which is a compound composed of any two to four elements selected from Ge, Sb, Se, Te, or S, including Ge2Sb2SexTe. 5-x (x=0,1,2,3,4,5), Sb2Se3, Sb2S3, GeTe.

6. The scalable non-volatile on-chip optical neural network according to claim 5, characterized in that, The phase change material (2-1) may also be selectively doped with non-metallic elements and / or metallic elements, wherein the non-metallic elements include oxygen (O), nitrogen (N), carbon (C), and silicon (Si); and the metallic elements include tin (Sn), silver (Ag), cadmium (Cd), bismuth (Bi), and indium (In).

7. The scalable, non-volatile on-chip optical neural network according to claim 1, characterized in that, The dielectric cladding (2-3) is an oxide dielectric material, including aluminum oxide (Al2O3), silicon oxide (SiO2), zirconium oxide (ZrO2), and hafnium oxide (HfO2).

8. The scalable non-volatile on-chip optical neural network according to claim 1, characterized in that, The unit devices that make up an optical neural network can be completed by using the periodic response of a resonant device and the matching between the passbands of a bandpass device, or by feeding a reflected signal into a periodic resonant device.

9. A scalable, non-volatile on-chip optical neural network according to claim 8, characterized in that, The unit devices that make up an optical neural network include fast volatile modulation functions and non-volatile storage functions, and the functions of the unit devices can be controlled by the characteristics of external control signals.

10. A scalable, non-volatile on-chip optical neural network according to claim 9, characterized in that, The real-time control scheme for fast volatile control function can be achieved by utilizing nonlinear effects, including thermo-optical effects, carrier depletion, carrier dispersion, as well as Pockel and Kerr effects.

11. A scalable, non-volatile on-chip optical neural network according to claim 10, characterized in that, The induction of phase change material (2-1) can be achieved by sharing the same driving electrode (2-2) with the volatile control function, or by using different driving electrodes (2-2) with the volatile control function.

Images