Ferromagnetic artificial neuron film and application thereof

Abstract

The invention provides a ferromagnetic artificial neuron film and application thereof. The ferromagnetic artificial neuron film comprises a transparent substrate and a ferromagnetic heterojunction constructed on the transparent substrate, the ferromagnetic heterojunction comprises a ferromagnetic alloy layer, a heavy metal layer and an anti-oxidation layer. Nonlinear spin oscillation of the ferromagnetic artificial neuron film is excited through femtosecond laser, terahertz radiation with intensity nonlinear response is directly generated, and the ferromagnetic artificial neuron film serves as an artificial neuron of a space optical neural network. The activation function response meeting the performance requirement of the artificial neuron can be effectively generated without introducing the regulation and control of an external magnetic field, and a better artificial neuron function facing the optical space neural network is realized.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of space optical neural network artificial neuron film design, in particular to a ferrimagnetic artificial neuron film and its application. Background Art

[0002] In recent years, the rapid development of artificial intelligence has put forward unprecedented demands on the parallelism, speed and energy efficiency of computing hardware. Traditional electronic neural networks are limited by the physical bottlenecks of electronic devices (such as resistance loss and clock frequency limitations), while optical neural networks, with the characteristics of high-speed photon propagation, high bandwidth and low energy consumption, have become a disruptive direction to break through the computing power dilemma of the "post-Moore era". In this field, spatial optical neural networks achieve parallel computing through the propagation and modulation of light fields in free space, and are particularly good at tasks such as image processing and pattern recognition.

[0003] In neural network calculations, neural networks introduce nonlinearity through activation functions, allowing neural networks to learn and simulate complex nonlinear relationships. Without activation functions, the neural network will become a linear model. No matter how many layers the network has, its output is a linear combination of the input, and it cannot handle complex nonlinear problems such as image recognition. For spatial optical neural networks, the core challenge is how to build optical artificial neurons with nonlinear activation functions - this component must simultaneously meet the requirements of high-speed response, low power consumption, and integration compatible with optical systems.

[0004] Terahertz (THz) waves (0.1–10 THz) have shown great potential in ultra-high-speed signal processing due to their unique spectral characteristics (between microwaves and infrared), injecting new impetus into spatial optical neural networks, but their efficient generation and dynamic modulation technology are still bottlenecks. Ferrimagnetic materials can theoretically support ultrafast optical-magnetic interactions due to their non-complete cancellation of spontaneous magnetic order and magnetic moment, and ultrafast spin dynamics (femtosecond to picosecond scale). However, traditional spin films as THz emission sources have limited nonlinear dynamic range or rely on complex external regulation, and usually rely on bulky laser systems or external magnetic field regulation, which makes it difficult to meet the needs of integrated optical neural networks. Summary of the invention

[0005] In view of the technical problems existing in the prior art, the present invention proposes a ferrimagnetic artificial neuron film and applications thereof.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows: In one aspect, the present invention provides a ferrimagnetic artificial neuron film, comprising: A transparent substrate and a ferrimagnetic heterojunction constructed on the transparent substrate; The ferrimagnetic heterojunction includes a ferrimagnetic alloy layer, a heavy metal layer, and an anti-oxidation layer.

[0007] Preferably, the heavy metal layer also serves as the anti-oxidation layer.

[0008] Preferably, the substrate is a double-sided polished substrate made of glass, quartz, sapphire, or mica, having high light transmittance and good thermal stability.

[0009] Preferably, the material of the ferrimagnetic alloy layer is an alloy of transition metals and rare earth metals.

[0010] Preferably, the material of the heavy metal layer is a heavy metal with a large spin Hall angle. Heavy metals with a large spin Hall angle include Pt, W, Ir, and Cu.

[0011] The above-mentioned application of the ferrimagnetic artificial neuron thin film in a spatial optical neural network. By exciting the non-linear spin oscillation of the ferrimagnetic artificial neuron thin film with femtosecond laser, directly generating terahertz radiation with an intensity non-linear response, and using the ferrimagnetic artificial neuron thin film as an artificial neuron of the spatial optical neural network.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: The ferrimagnetic thin film described in the present invention provides a terahertz non-linear artificial neuron for the spatial optical neural network.

[0013] The present invention can be independent of external magnetic field regulation, avoiding the additional energy consumption and the complexity of the spatial optical neural network structure caused by external excitation.

[0014] The present invention applies ferrimagnetic terahertz dynamics to neural network computing, making it have the characteristics of low loss and programmability. Description of the Drawings

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0016] Figure 1 It is a physical model diagram of the ferrimagnetic artificial neuron thin film; Figure 2 It is a comparison diagram of the change of terahertz radiation with pump power and pump duration of the ferrimagnetic artificial neuron thin film with and without external magnetic field regulation, where Figure 2 (a) is a diagram of the change of terahertz radiation of the ferrimagnetic artificial neuron thin film with the power and time of the femtosecond laser pump source under external magnetic field regulation;Figure 2 (b) is a graph showing the variation of terahertz radiation of a ferromagnetic artificial neuron thin film with the power and time of a femtosecond laser pump source without external magnetic field regulation; Figure 2 (c) is a graph showing the variation of the extreme value of terahertz radiation of a ferromagnetic artificial neuron thin film with the femtosecond laser power without external magnetic field regulation; Figure 3 For Figure 2 Matlab program graph for linear transformation of the experimental results in (c); Figure 4 For using Figure 2 Fitting result image of non - ideal ReLU activation function by non - linear fitting using the linearly transformed experimental results in (c); Figure 5 It is the working performance graph of the ferromagnetic artificial neuron thin film, where Figure 5 (a) is a conceptual graph of a two - layer artificial neural network built using a ferromagnetic artificial neuron thin film; Figure 5 (b) shows the graph of the relationship between the accuracy of the artificial neural network built using a ferromagnetic artificial neuron thin film for performing the MINIST handwritten digit recognition task and the number of iterations. Specific implementation manner

[0017] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0018] The present invention proposes a ferromagnetic artificial neuron thin film, including: A transparent substrate and a ferromagnetic heterojunction constructed on the transparent substrate; The ferromagnetic heterojunction includes a ferromagnetic alloy layer, a heavy metal layer, and an anti - oxidation layer.

[0019] The composition of the ferromagnetic alloy layer is an alloy of transition metals (such as iron Fe, cobalt Co) and rare earth metals (such as gadolinium Gd, terbium Tb), and the thickness is 1nm - 20nm. The doping of rare earth metals in the ferromagnetic alloy layer is to make the magnetic domain directions in the ferromagnetic layer have vertical anisotropy and a certain degree of in - plane anisotropy, ensuring the effective generation of spin electron flow without an external magnetic field.

[0020] The ferromagnetic alloy layer can be grown on the transparent substrate by means of vacuum magnetron co - sputtering to ensure the controllability of the rare earth metal doping concentration and the thickness of the ferromagnetic alloy layer.

[0021] is a heavy metal with a large Hall spin angle, such as Pt, W, Ir, Cu, etc., with a thickness of 1 nm to 10 nm. Such a selection can effectively ensure that the spin electron current can effectively utilize the inverse spin Hall effect to generate a transverse current when entering the non-ferromagnetic layer, ensuring the generation of terahertz radiation.

[0022] Regarding the selection of the transparent substrate, in principle, as long as good light transmittance is ensured to enable the femtosecond laser for excitation to be well incident, and at the same time, it should also be convenient for processing and manufacturing. Preferably, the selected material for the transparent substrate is a double-sided polished transparent material such as glass, quartz, sapphire, mica, etc., with a thickness of 0.2 - 5 nm.

[0023] Regarding the anti-oxidation layer, since the heavy metal layer used in the present invention can effectively prevent oxidation by itself, the heavy metal layer is used as the anti-oxidation layer. In principle, the presence or absence of the heavy metal layer and the selection of its composition should be based on whether its anti-oxidation function can be effectively achieved, and at the same time, it should not affect the use of the device itself.

[0024] Refer to Figure 1 , Figure 1 is the physical model diagram of the ferromagnetic artificial neuron thin film. Among them, a is the thickness of the ferromagnetic alloy layer, b is the thickness of the heavy metal layer, and c is the thickness of the transparent substrate.

[0025] To demonstrate the working effect of the present invention, in one embodiment, the transparent substrate material of the ferromagnetic artificial neuron thin film is quartz, the ferromagnetic alloy layer is FeTb, and the heavy metal layer is Pt. The following parameters are set for numerical simulation. The thickness of the transparent substrate is set to 5 nm. The ferromagnetic alloy layer is made of FeTb material with a thickness of a = 2 nm. The heavy metal layer is selected as Pt material with a thickness of b = 4 nm. To analyze the response effect of the ferromagnetic artificial neuron thin film without external magnetic field regulation, the present invention respectively conducts the response effects of the terahertz radiation of the artificial neuron thin film with and without external magnetic field regulation with respect to the femtosecond laser pump power. Among them, a sapphire laser oscillator (center wavelength 800 nm, pulse duration 35 fs, repetition frequency 10 KHz) is used as the pump source for terahertz emission measurement; a fiber-based terahertz time-domain spectrometer is used to measure the transmittance of the sample in the terahertz range. Refer to Figure 2 , Figure 2 is the comparison diagram of the changes of terahertz radiation with pump power and pump duration of the ferromagnetic artificial neuron thin film (transparent substrate material is quartz, ferromagnetic alloy layer is FeTb, heavy metal layer is Pt) with and without external magnetic field regulation, where Figure 2 (a) is the change diagram of the terahertz radiation of the ferromagnetic artificial neuron thin film with the power and time of the femtosecond laser pump source under external magnetic field regulation; Figure 2(b) Variation diagram of terahertz radiation of the ferrimagnetic artificial neuron thin film with the power and time of the femtosecond laser pump source without external magnetic field regulation; Figure 2 (c) Variation diagram of the extreme value of terahertz radiation of the ferrimagnetic artificial neuron thin film with the femtosecond laser power without external magnetic field regulation. It can be seen that the results of terahertz radiation are almost the same as those with external magnetic field regulation in terms of linear performance without external magnetic field, meeting the expected requirements for artificial neurons. Among them, the linear shape of the radiation extreme value is similar to that of the ReLU activation function in the artificial neural network, which is crucial for realizing the function of artificial neurons. It can be known that the response results of the present invention without external magnetic field regulation meet the expected requirements for the function of artificial neurons. The function of the artificial neuron device based on the terahertz source without external magnetic field regulation is well realized.

[0026] The above application of the ferrimagnetic artificial neuron thin film in the spatial optical neural network. By exciting the nonlinear spin oscillation of the ferrimagnetic artificial neuron thin film with a femtosecond laser, terahertz radiation with a nonlinear intensity response is directly generated, and the ferrimagnetic artificial neuron thin film is used as an artificial neuron in the spatial optical neural network.

[0027] To further verify the reliability of the experimental results of the present invention, the variation of terahertz radiation with the femtosecond pump laser power measured by experiments of the present invention is used as the original data of the activation function in the spatial optical neural network, and the non-ideal ReLU function ( ) is used to fit this data. Among them, the power of the femtosecond laser is used as the independent variable, and the terahertz radiation power is used as the dependent variable. To achieve a better fitting effect, the experimental data is first linearly transformed using the matlab code shown in Figure 3 : the abscissa is mirror-operated along the y-axis, and the abscissa and ordinate are translated; finally, to achieve a more appropriate learning rate, the abscissa and ordinate are appropriately scaled to obtain an appropriate value of a for the ReLU function. Figure 4 For the fitting result image of the non-ideal ReLU activation function by nonlinearly fitting the experimental results of Figure 2 (c) after linear transformation, this image is obtained by linearly transforming and fitting the experimental data of Figure 2 (c). It can be seen that the fitted activation function fits well with the distribution of the experimental data.

[0028] To verify the working performance of the present invention, the spatial optical neural network uses the activation function fitted by the present invention. Figure 5 For the working performance diagram of the ferrimagnetic artificial neuron thin film, where Figure 5(a) is a conceptual diagram of a two-layer artificial neural network built using a ferrimagnetic artificial neuron thin film. In this network model, the pixels of a handwritten digit image are divided into 28×28 as the input layer. 100 neurons are set as the hidden layer. 10 neurons are used as the output layer to represent the digits from 0 to 9. Figure 5 (b) shows a graph of the relationship between the accuracy of the artificial neural network built using a ferrimagnetic artificial neuron thin film in performing the MINIST handwritten digit recognition task and the number of iterations. By adjusting the learning rate and the number of iterations of the artificial neural network, the recognition accuracy can reach more than 97% at most, which is close to that of the neural network built using ideal neurons. Therefore, when the device processing accuracy is sufficient, the present invention is expected to achieve low-power consumption, high integration, and small-area spatial optical neuromorphic computing.

[0029] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0030] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

[0031] The above is only the preferred embodiment of the present invention and is not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A ferrimagnetic artificial neuron thin film, characterized in that, Comprising: A transparent substrate and a ferrimagnetic heterojunction constructed on the transparent substrate; The ferrimagnetic heterojunction includes a ferrimagnetic alloy layer, a heavy metal layer, and an anti-oxidation layer.

2. The ferrimagnetic artificial neuron thin film according to claim 1, wherein The heavy metal layer also serves as the anti-oxidation layer.

3. The ferrimagnetic artificial neuron thin film according to claim 1 or 2, characterized in that, The substrate is a double-sided polished substrate made of glass, quartz, sapphire, or mica, having high light transmittance and thermal stability.

4. The ferrimagnetic artificial neuron thin film according to claim 3, wherein The material of the ferrimagnetic alloy layer is an alloy of transition metals and rare earth metals.

5. The ferrimagnetic artificial neuron thin film according to claim 4, characterized in that, The magnetic domain direction in the ferrimagnetic alloy layer has vertical anisotropy and a certain degree of in-plane anisotropy, which can ensure the effective generation of spin electron flow without an external magnetic field.

6. The ferrimagnetic artificial neuron thin film according to claim 4, wherein The ferrimagnetic alloy layer is grown on the transparent substrate by means of vacuum magnetron co-sputtering.

7. The ferrimagnetic artificial neuron thin film according to any one of claims 3 to 6, characterized in that The material of the heavy metal layer is a heavy metal with a large spin Hall angle. Heavy metals with a large spin Hall angle include Pt, W, Ir, and Cu.

8. The ferrimagnetic artificial neuron thin film according to claim 7, characterized in that, The thickness of the transparent substrate is 0.2 nm to 5 nm, and the thickness of the ferrimagnetic alloy layer is 1 nm to 20 nm.

9. Application of the ferrimagnetic artificial neuron thin film according to claim 1 in a spatial optical neural network.

10. The application according to claim 9, wherein By femtosecond laser exciting the nonlinear spin oscillation of the ferrimagnetic artificial neuron thin film to directly generate terahertz radiation with an intensity nonlinear response, and using the ferrimagnetic artificial neuron thin film as an artificial neuron in a spatial optical neural network.

Images