An adjustable artificial neural network device based on topological insulator / two-dimensional ferromagnetic heterojunction

Abstract

The application discloses a kind of adjustable artificial neural network components based on topological insulator / 2D ferromagnetic heterojunction.The adjustable artificial neural network components of the application, from bottom to top, include substrate wafer, 2D ferromagnetic layer, topological insulator layer and metal layer;Among them, the 2D ferromagnetic layer uses chromium telluride CrTe2 or iron germanium telluride Fe3GeTe2 material;The topological insulator layer uses bismuth telluride Bi2Te3 or bismuth selenide Bi2Se3 material;The component generates self-spinning orbit moment of topological insulator layer and is regulated to 2D ferromagnetic magnetic moment, and corresponding multi-state magnetic domain flipping curve can realize the function of synapse;Meanwhile, the intrinsic magnetism of 2D ferromagnetic material enables adjustable activation function in data training, and then improves the training efficiency and accuracy of neural network.Compared with traditional RRAM, the component of the application realizes adjustable activation function, and has higher linearity and thermal stability.

Description

Technical Field

[0001] This invention relates to an adjustable artificial neural network device based on a topological insulator / two-dimensional ferromagnetic heterojunction, belonging to the fields of semiconductor and novel computing technologies. Background Technology

[0002] With the emergence of the concept of neuromorphic computing, researchers have been mimicking the working patterns of the brain to achieve the functions of neurons and synapses, enabling the rapid and accurate processing of large-scale information. In recent years, with continuous advancements in manufacturing processes, digital very-large-scale integrated circuits (VLSI) have become the primary method for implementing neural network hardware. While digital circuit implementations offer advantages such as stable device performance and accurate computational results, their massive circuit size exponentially increases the power consumption of the entire network.

[0003] Therefore, researchers have proposed implementing artificial neural networks using analog circuits to reduce power consumption. Among the many analog methods, the scheme based on resistive random access memory (RRAM) was the earliest. However, RRAM suffers from poor linearity and relatively poor thermal stability. In contrast, magnetic random access memory (MRAM) has become a major choice for the hardware implementation of artificial neural networks due to its advantages such as low power consumption, long lifespan, and fast flip-flop speed.

[0004] In MRAM, spin-orbit torque (SOT) based devices have become one of the candidates for next-generation MRAM due to their lower power consumption and faster switching speed. To obtain high-performance SOT devices, a non-magnetic layer capable of providing a large spin current and a ferromagnetic layer with polar anisotropy (PMA) are required. Topological insulators (TIs) have been widely used in spintronics in recent years due to their high spin-transfer efficiency and strong spin current. Currently reported TI-based SOT-MRAMs mostly require the introduction of an additional intercalation layer between the TI and the ferromagnetic layer (FM) to achieve PMA, which inevitably introduces more non-ideal conditions into the fabrication process.

[0005] Aside from the structural and manufacturing limitations of the devices themselves, functionally, memristors like RRAM can only act as synapses, requiring software to implement neuron functions. Secondly, from a neural network algorithm perspective, the concept of batch normalization (BN) dictates that the activation function in each layer of neurons should be adjusted according to the different input signals to achieve optimal training results. However, achieving adjustable activation functions in hardware is even more challenging. Therefore, currently, there is no neural network device that can simultaneously reduce power consumption and improve training performance with fully hardware-implemented adjustable neurons. Summary of the Invention

[0006] The technical problem to be solved by this invention is that the traditional simulation methods for implementing neural networks in hardware do not have the function of adjusting the activation function during training.

[0007] To address the aforementioned technical problems, this invention provides a tunable artificial neural network device based on a topological insulator / two-dimensional ferromagnetic heterojunction, characterized in that it comprises, from bottom to top, a substrate wafer, a two-dimensional ferromagnetic layer, a topological insulator layer, and a metal layer.

[0008] The two-dimensional ferromagnetic layer is made of chromium telluride CrTe2 or iron germanium telluride Fe3GeTe2 material.

[0009] The aforementioned topological insulating layer is made of bismuth telluride (Bi2Te3) or bismuth selenide (Bi2Se3).

[0010] The aforementioned components regulate the two-dimensional ferromagnetic moment through the spin-orbit torque generated by the topological insulator layer, and the corresponding multi-state magnetic domain flipping curve can realize the function of a synapse; at the same time, the intrinsic magnetism of the two-dimensional ferromagnetic material enables the adjustable activation function in data training, thereby improving the training efficiency and accuracy of the neural network.

[0011] Preferably, the substrate wafer is made of alumina (Al2O3), and the metal layer is a combination of titanium (Ti) and gold (Au).

[0012] Preferably, the thickness of both the two-dimensional ferromagnetic layer and the topological insulator layer is 1–99 nm; the metal layer includes a Ti layer of 1–99 nm and an Au layer of 100–999 nm.

[0013] This invention also provides a method for adjusting the activation function of the above-mentioned tunable artificial neural network device based on a topological insulator / two-dimensional ferromagnetic heterojunction, characterized by comprising the following steps:

[0014] Step 1: Apply current to both ends of the component. Due to the spin Hall effect of the topological insulator layer, the spin orbital torque will act on the magnetic moment of the two-dimensional ferromagnetic layer and change its magnetization direction.

[0015] Step 2: Measure the voltage across terminals 2-4. After adjusting the magnetization direction of the two-dimensional ferromagnetic layer in step 1, the electrons flowing through the layer will be scattered by the magnetic moment due to the anomalous Hall effect, resulting in a potential difference in the transverse direction.

[0016] Step 3: Based on the results of Step 1 and Step 2, according to R 24 =V 24 / I 13 Determine the value of the Hall resistance;

[0017] Step 4: Fix the pulse width of the current pulse to several hundred microseconds, start the pulse size from the minimum negative current of tens of milliamperes, gradually increase to the maximum positive current, and then decrease to the minimum negative current. Record the Hall resistance under different pulses, and obtain multiple RI loops according to the different maximum positive currents.

[0018] Step 5: Select curves ranging from negative milliamperes to negative tens of milliamperes, normalize the Hall resistances respectively, and then fit a Sigmoid function:

[0019] Step 6: Based on the different maximum positive currents in Step 4, Step 5 gives the fitted k values ​​of 0.37-1.28 for each case, representing the effect of the activation function;

[0020] Step 7: Build a complete neural network. Before the input signal of each stage is passed into the device, adjust the Hall resistance of the device with a positive current pulse according to the magnitude of the signal, thereby adjusting the k value to improve training efficiency and speed.

[0021] The principle behind adjusting the activation function is as follows:

[0022] After fabricating the device, a current is passed between the two ports of the cross-shaped device. The portion flowing through the Bi₂Te₃ layer undergoes a spin Hall effect (SHE), causing electrons with spin up (spin down) to move downwards (upwards). The downward-moving electrons transfer their angular momentum to the magnetic moment of the CrTe₂ layer, altering its magnetization direction. The current flowing through the CrTe₂ layer is then scattered by the magnetic moments in different directions (AHE), resulting in an unbalanced charge accumulation perpendicular to the current direction, leading to a potential difference. The Hall resistance is obtained by recording the voltage under different current pulses and dividing it by the applied current.

[0023] Due to the presence of multiple domains, a positive current causes the Hall resistor to tend towards a high-resistance state, and the resistance value is positively correlated with the magnitude of the positive current. Therefore, by fixing the low-resistance state and changing the magnitude of the high-resistance state, as the device transitions from a high-resistance state to a low-resistance state with a negative current, a linear shape similar to a sigmoid function will fit different coefficients k. According to the concept of Batch Normalization (BN), during the operation of a neural network, each neuron needs to adjust its activation function based on the magnitude of the input signal before receiving it. To this end, we will adjust the Hall resistor state of the device using a positive current pulse, thereby changing the value of the activation function k to improve training efficiency and speed.

[0024] This invention provides a memristor device based on a topological insulator / two-dimensional ferromagnetic heterojunction to realize the synaptic and neuron functions in an artificial neural network. By controlling the two-dimensional ferromagnetic magnetic moment through the spin-orbit torque generated by the topological insulator layer, the corresponding multi-state magnetic domain flipping curve can realize the synaptic function; simultaneously, the intrinsic magnetism of the two-dimensional ferromagnetic material can enable tunable activation functions during data training, thereby improving the training efficiency and accuracy of the neural network.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] This invention utilizes a multi-domain micron junction fabricated from a two-dimensional ferromagnetic combined topological insulator and the characteristic of the SOT flip curve of the sigmoid function. By applying positive current pulses of different magnitudes, the activation function can be changed during training, thereby improving the training speed and accuracy of the neural network. Compared with traditional RRAM, the components of this invention have higher linearity and thermal stability. Furthermore, compared with other hardware implementations that are not adjustable during neural network training, this invention enables the change of the activation function during training, improving the accuracy and speed of training. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the SOT cross-shaped device structure;

[0028] Figure 2 Schematic diagram of the normalized RI hysteresis loop under different maximum positive currents;

[0029] Figure 3 This is a schematic diagram of the Sigmoid function fitting of the SOT flip curve in the negative current region under different maximum positive current states. Detailed Implementation

[0030] To make the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings.

[0031] Example 1

[0032] Single SOT cross-shaped device such as Figure 1 As shown. From bottom to top, they are substrate layer a, two-dimensional ferromagnetic layer b, topological insulator layer c, and metal electrode layer d.

[0033] The substrate wafer is made of aluminum oxide (Al2O3).

[0034] The two-dimensional ferromagnetic layer is made of chromium telluride (CrTe2) and has a thickness of 3 nm. The topological insulator layer is made of bismuth telluride (Bi2Te3) and has a thickness of 25 nm. Both the two-dimensional ferromagnetic layer and the topological insulator layer are grown on the substrate using conventional magnetron sputtering or molecular beam epitaxy methods.

[0035] The metal electrode layer uses a combination of titanium (Ti) and gold (Au), with 15 nm of titanium and 150 nm of gold deposited using conventional methods in the art.

[0036] The device works as follows:

[0037] Step 1: Apply current in the direction of the arrow at both ends 1-3. Due to the spin Hall effect of the Bi2Te3 layer, the spin orbital torque will act on the magnetic moment of CrTe2 and change its magnetization direction.

[0038] Step 2: Measure the voltage across terminals 2-4. After adjusting the magnetization direction of CrTe2 in Step 1, electrons flowing through this layer will be scattered by the magnetic moment due to the anomalous Hall effect, resulting in a potential difference in the transverse direction.

[0039] Step 3: Based on the results of Step 1 and Step 2, according to R... 24 =V 24 / I 13 This allows us to determine the magnitude of the Hall resistance. Figure 1 )

[0040] Step 4: Set the pulse width of the fixed current pulse to 500 microseconds. Start with a minimum negative current of -32 mA, gradually increase the pulse size to the maximum positive current, and then decrease it back to the minimum negative current. Record the Hall resistance under different pulse conditions. Based on the different maximum positive currents (from 19 mA to 32 mA), obtain multiple RI loops (…). Figure 2 ).

[0041] Step 5: Select a curve from -5 mA to -32 mA, normalize the Hall resistance, and then fit a Sigmoid function. ( Figure 3 ).

[0042] Step Six: Based on the different maximum positive currents in Step Four, Step Five gives the fitted k values ​​of 0.37-1.28 for each case, representing the effect of the activation function.

[0043] Step 7: Build a complete neural network. Before the input signal is passed to the device at each stage, adjust the Hall resistance of the device with a positive current pulse according to the magnitude of the signal, thereby adjusting the k value to achieve the optimal training effect (>95%).

[0044] Figure 2The entire loop represents the process of the magnetic moment of CrTe2 flipping from downward to upward and back to downward; therefore, there are two magnetic moment flipping processes. The results show that during the positive current flipping process, the different loops highly overlap, but during the negative current flipping process, we can extract multiple sigmoid functions with different slopes. This provides a feasible solution for implementing an adjustable activation function during training. In this example, the adjustable range of k is 0.37-1.28. By adjusting the device size and thickness, the adjustable range of k will be further increased.

[0045] Based on our simulation results in multi-level neural networks, adjusting k during training can improve the training accuracy from 93.17% to over 95% compared to a fixed k.

Claims

1. A tunable artificial neural network device based on a topological insulator / two-dimensional ferromagnetic heterojunction, characterized in that, From bottom to top, it includes a substrate wafer, a two-dimensional ferromagnetic layer, a topological insulator layer, and a metal layer; The two-dimensional ferromagnetic layer is made of chromium telluride CrTe2 or iron germanium telluride Fe3GeTe2 material. The aforementioned topological insulating layer is made of bismuth telluride (Bi2Te3) or bismuth selenide (Bi2Se3). The aforementioned components regulate the two-dimensional ferromagnetic moment through the spin-orbit torque generated by the topological insulator layer, and the corresponding multi-state magnetic domain flipping curve can realize the function of a synapse; at the same time, the intrinsic magnetism of the two-dimensional ferromagnetic material enables the adjustable activation function in data training, thereby improving the training efficiency and accuracy of the neural network.

2. The tunable artificial neural network device based on a topological insulator / two-dimensional ferromagnetic heterojunction as described in claim 1, characterized in that, The substrate wafer is made of aluminum oxide (Al2O3), and the metal layer is a combination of titanium (Ti) and gold (Au).

3. The tunable artificial neural network device based on a topological insulator / two-dimensional ferromagnetic heterojunction as described in claim 2, characterized in that, The thickness of both the two-dimensional ferromagnetic layer and the topological insulator layer is 1–99 nm; the metal layer includes a Ti layer of 1–99 nm and an Au layer of 100–999 nm.

4. The method for adjusting the activation function of the tunable artificial neural network device based on a topological insulator / two-dimensional ferromagnetic heterojunction as described in any one of claims 1 to 3, characterized in that, Includes the following steps: Step 1: Apply current to both ends of the component. Due to the spin Hall effect of the topological insulator layer, the spin orbital torque will act on the magnetic moment of the two-dimensional ferromagnetic layer and change its magnetization direction. Step 2: Measure the voltage across terminals 2-4. After adjusting the magnetization direction of the two-dimensional ferromagnetic layer in step 1, the electrons flowing through the two-dimensional ferromagnetic layer will be scattered by the magnetic moment due to the anomalous Hall effect, resulting in a potential difference in the transverse direction. Step 3: Based on the results of Step 1 and Step 2, according to R 24 =V 24 / I 13 Determine the value of the Hall resistance; Step 4: Fix the pulse width of the current pulse to several hundred microseconds, start the pulse size from the minimum negative current of tens of milliamperes, gradually increase to the maximum positive current, and then decrease to the minimum negative current. Record the Hall resistance under different pulses, and obtain multiple RI loops according to the different maximum positive currents. Step 5: Select curves ranging from negative milliamperes to negative tens of milliamperes, normalize the Hall resistances respectively, and then fit a Sigmoid function: Step 6: Based on the different maximum positive currents in Step 4, Step 5 gives the fitted k values ​​of 0.37-1.28 for each case, representing the effect of the activation function; Step 7: Build a complete neural network. Before the input signal of each stage is passed into the device, adjust the Hall resistance of the device with a positive current pulse according to the magnitude of the signal, thereby adjusting the k value to improve training efficiency and speed.

Images