A room temperature mode selection physical reservoir computing method and device
By switching the electrical drive mode in a skyminzi spring oscillator, rapid selection of nonlinearity and memory capabilities on the same device is achieved. This solves the problems of nonlinearity and memory trade-offs, task adaptability, and room temperature stability in physical reservoirs, improving the flexibility and integration of computing resources, and making it suitable for signal conversion and prediction tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINZHONG UNIV
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-02
AI Technical Summary
Existing physical memory pool solutions are difficult to optimize in the trade-off between nonlinearity and memory, have poor task adaptability, low system integration, insufficient room temperature stability, and are difficult to achieve rapid and flexible adjustment of computing resources and miniaturization of devices.
A skyminzi spring oscillator with an artificially synthesized antiferromagnetic structure achieves rapid selection of nonlinear characteristics and short-time memory capability on the same device by switching the electrical drive mode. Mode selection and stability improvement are achieved by using a multi-channel readout circuit and voltage-controlled magnetic anisotropy technology.
It enables rapid adjustment of computing resources at room temperature to adapt to different computing tasks, improves the stability and integration of devices, supports efficient processing of tasks such as signal conversion and time series prediction, and reduces hardware cost and complexity.
Smart Images
Figure CN122132175A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of neuromorphic computing and neuromorphic chip technology, and more specifically, to a method and device for calculating a room temperature mode-selective physical reservoir. Background Technology
[0002] Physical reservoir computation is an efficient computational paradigm that utilizes the transient dynamics of complex physical systems to process time-series information. Its core advantage lies in the fact that only a simple linear readout layer needs to be trained, while the complex nonlinear mappings are handled by the physical system itself. An ideal physical reservoir needs to possess sufficient nonlinearity, rich dynamic dimensions, moderate short-term memory, and good repeatability.
[0003] However, existing physical storage pool implementation solutions generally face the following key challenges:
[0004] Performance attribute coupling: In most physical systems, enhancing nonlinearity often leads to increased system dissipation and shortened effective memory time; while extending memory may cause the dynamics to tend towards linearity. This "nonlinearity-memory" trade-off limits the optimization space for specific tasks.
[0005] Poor task adaptability: To adapt to different task requirements, existing technologies usually require significant changes to external conditions, resulting in complex operation, high energy consumption, slow response, and difficulty in achieving rapid and flexible reconfiguration.
[0006] Low system integration: Many solutions rely on external optical systems, large electromagnets, temperature control devices or precision mechanical structures, making it difficult to achieve miniaturization, arraying of devices and integration with standard semiconductor processes.
[0007] Room temperature stability challenge: Room temperature effects such as thermal fluctuations and parameter drift can interfere with the dynamic behavior of physical systems, reduce the repeatability and accuracy of computational outputs, and limit their reliability in practical environments.
[0008] Therefore, there is a need for a new type of physical storage pool that can operate stably at room temperature, flexibly adjust its computing resources (nonlinearity and memory) in a simple manner, and is easy to integrate. Summary of the Invention
[0009] This invention aims to overcome the shortcomings of existing technologies and provide a room-temperature mode-selective physical reservoir calculation method and device. Its core objective is to achieve relatively independent and rapid electrical selection of the nonlinear characteristics and short-time memory capabilities of skyrmion spring oscillators within the same magnetic nanodevice, utilizing different intrinsic dynamic modes. This allows for rapid adjustment of computational resources on the same device at room temperature, efficiently adapting to different computational tasks such as signal conversion and time series prediction, and improving the stability and integration of the device.
[0010] Technical solution
[0011] To achieve the above objectives, the present invention provides the following technical solution:
[0012] A calculation method for a room temperature mode-selective physical reservoir
[0013] This method is based on a skyrmion spring oscillator device containing an artificially synthesized antiferromagnetic structure, characterized by comprising the following steps:
[0014] S1. Initialization: A room-temperature skyrmion is generated and stabilized as the initial state within the nano-confined magnetic region of the skyrmion spring oscillator.
[0015] S2, Mode Selection: Select the corresponding physical working mode by switching the electrical drive mode according to the type of computing task to be performed: configure the multi-channel readout circuit and set the sampling frequency and time step;
[0016] First operating mode (strong nonlinear mode): For signal conversion tasks, a low-frequency alternating current is applied along the restricted motion direction of the skyrmion, and the skyrmion is driven into a large-amplitude oscillation mode through spin-transfer torque; in this mode, the skyrmion performs large-amplitude nonlinear relaxation oscillations between the two ends of the confined potential well, providing a strong nonlinear mapping capability for the reservoir.
[0017] Second operating mode (strong memory mode): For signal prediction tasks, an alternating magnetic field is applied in a direction perpendicular to the device plane to excite the breathing mode of the skyrmion; in this mode, the topological size of the skyrmion changes periodically, and its dynamic response has a longer decay time constant, providing excellent short-term memory capability for the reservoir.
[0018] S3, Input Loading: The timing input signal to be processed is amplitude-modulated and modulated onto the selected drive signal;
[0019] S4. State Sampling and Matrix Construction: Under the action of the modulated drive signal, the physical state response signals of each channel are synchronously acquired through a multi-channel parallel readout circuit;
[0020] S5. Training and Execution: The linear readout layer weights are trained using multi-channel physical state signals and target output signals. During the inference phase, the weights are fixed, and the real-time acquired multi-channel physical state signals are used to perform calculations with the weight matrix to obtain the calculation results.
[0021] Preferably, when performing signal prediction tasks in the second operating mode, the voltage-controlled magnetic anisotropy modulation unit is enabled to apply a DC gate voltage to a specific region of the device, locally modulating the vertical magnetic anisotropy to form a shallow potential well, thereby suppressing the thermally induced slow drift of skyrmions at room temperature and improving the stability and accuracy of long-term prediction.
[0022] Preferably, the signal conversion tasks include waveform transformation, speech recognition, and image recognition; the signal prediction tasks include chaotic time series prediction, financial time series prediction, meteorological and environmental data prediction, and biomedical signal prediction.
[0023] A room temperature mode-selective physical reservoir computing device for implementing the above method
[0024] The device is a multilayer thin film stacked structure, characterized in that, from bottom to top, it includes: a substrate, a buffer layer formed on the substrate for electrical insulation and surface planarization, a bottom planar microcoil patterned on the buffer layer for generating an alternating magnetic field perpendicular to the device plane, and a bottom voltage-controlled magnetic anisotropy structure disposed on the bottom planar microcoil, including a bottom gate electrode and a gate dielectric layer.
[0025] An artificially synthesized antiferromagnetic nanostructure is deposited and patterned on the bottom voltage-controlled magnetic anisotropy structure to form the core magnetic region of the skyrmion spring oscillator. The structure comprises at least a sandwich stack of ferromagnetic layers / nonmagnetic spacer layers / ferromagnetic layers, wherein the thickness of the spacer layers is selected to generate antiferromagnetic interlayer exchange coupling. The nanostructure is patterned into nanostrips with a specific aspect ratio to form a geometric confinement of the skyrmion.
[0026] The segmented anomalous Hall effect readout array consists of multiple independent metal electrode pairs arranged on both sides along the long axis of the nanostructure, and is used to measure the local Hall resistance change caused by skyrmion motion in parallel.
[0027] A current-driven electrode is electrically connected to both ends of the artificially synthesized antiferromagnetic nanostructure to inject current along its long axis.
[0028] A top voltage-controlled magnetic anisotropy structure is disposed on the artificially synthesized antiferromagnetic nanostructure and the segmented readout array;
[0029] Top planar microcoils are patterned on the top voltage-controlled magnetic anisotropic structure;
[0030] The protective layer and interconnect structure cover the outermost layer and provide electrical connection channels.
[0031] Preferably, the segmented anomalous Hall effect readout array has more than or equal to thirty electrode pairs to achieve a state-space mapping of sufficient dimensions.
[0032] Preferably, the gate dielectric layer of the voltage-controlled magnetic anisotropy structure is a high dielectric constant dielectric.
[0033] Beneficial effects
[0034] Compared with the prior art, the present invention has the following significant advantages:
[0035] 1. Achieved rapid electrical reconfiguration of computational characteristics: This invention utilizes different intrinsic modes of the same skyrmion device, switching between operating modes in microseconds or less simply by switching between current-driven and magnetic-driven modes, thereby dynamically configuring the nonlinearity and memory properties of the reservoir. This solves the pain points of traditional solutions where the two are difficult to control independently, and task adaptation requires changes to hardware or large-scale scanning parameters.
[0036] 2. Multifunctionality of a single device: A single device can handle both signal conversion tasks that rely heavily on nonlinearity and timing prediction tasks that rely heavily on memory, which greatly improves the versatility and utilization of hardware resources and reduces the cost and complexity of customizing dedicated hardware for different tasks.
[0037] 3. Excellent room temperature operating stability:
[0038] In the large amplitude oscillation mode, the strong nonlinearity generated by the relaxation oscillation mode caused by low frequency and high current density, combined with an appropriate readout time, has good noise immunity.
[0039] In breathing mode, the innovative voltage-controlled magnetic anisotropy-assisted pinning technique effectively suppresses signal drift caused by thermal fluctuations, ensuring the long-term stability required for prediction tasks. Both mechanisms guarantee high-precision computational output at room temperature.
[0040] 4. High integration and system simplification: All key functional units (electric drive, electrical readout, and electrical control) are integrated on the same chip using planar technology, eliminating the need for complex external optical paths or strong magnetic field devices. This all-electric, on-chip integrated architecture is highly advantageous for large-scale array expansion and compatibility with CMOS circuits. Attached Figure Description
[0041] Figure 1 A system overview flowchart provided for embodiments of the present invention;
[0042] Figure 2 This is a flowchart illustrating the method execution provided in an embodiment of the present invention.
[0043] Figure 3 shows the frequency f = 0.2MHz and current density j = 1.5×10 in an embodiment of the present invention. 12Simulation results of sine wave to square wave conversion under large amplitude oscillation mode (A / m²);
[0044] Figure 4 shows the simulation results of the Mackey-Glass signal prediction under the respiratory mode driven by an alternating magnetic field with a frequency of f = 31.8 GHz in an embodiment of the present invention. Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0046] Example: A room temperature mode selective physical reservoir computing device and its operation method
[0047] 1. Device Structure and Fabrication
[0048] The device in this embodiment is fabricated on a semiconductor substrate using standard micro-nano fabrication processes. Its core is an artificially synthesized antiferromagnetic skyrmion spring oscillator, which integrates a mode selection drive unit, a multi-channel state readout unit, and a stability enhancement unit.
[0049] Specifically, the device, from bottom to top, includes: a substrate, such as a silicon wafer with an oxide layer.
[0050] Buffer layer: A deposited insulating film used for surface planarization and electrical isolation.
[0051] Bottom excitation unit: includes patterned planar microcoils (for generating an outward-facing alternating magnetic field) and a bottom voltage-controlled magnetic anisotropy structure (including bottom gate electrodes and a thin gate dielectric layer).
[0052] Core functional layer: Formed sequentially and graphically on the bottom VCMA structure:
[0053] Artificially synthesized antiferromagnetic (SAF) stacks employ a typical structure such as heavy metal / ferromagnetic layer / spacer layer / ferromagnetic layer / heavy metal. The heavy metal layer (e.g., Pt) provides strong spin-orbit coupling and perpendicular anisotropy at the interface, the ferromagnetic layer (e.g., Co, CoFeB) provides the magnetic moment, and the thickness of the non-magnetic spacer layer (e.g., Ru) is precisely controlled to achieve antiferromagnetic exchange coupling. This stack is defined as a nanoscale strip through photolithography and etching, its size ensuring that a skyrmion can be accommodated and stabilized at room temperature, forming a potential well that confines its motion.
[0054] Current-driven electrodes: fabricated at both ends of the SAF strip, used to inject in-plane current.
[0055] Segmented readout array: A series of spaced-apart pairs of metal electrodes are fabricated on both sides of the long side of the SAF strip. Each pair of electrodes is used to measure the anomalous Hall resistance of the corresponding magnetic region below, which is related to the presence and location of skyrmions in that region. Multiple such channels provide parallel, spatially distributed state information.
[0056] Top control and excitation unit: On the core functional layer, a voltage-controlled magnetic anisotropic structure (similar to the bottom structure) and a planar microcoil are fabricated sequentially at the top. The dual-gate and dual-coil design provides greater control flexibility and symmetry.
[0057] Packaging and interconnection: Finally, a passivation protective layer is applied, and contact holes are opened by photolithography and etching. Metal interconnects are then fabricated to lead out each electrode and coil port to the peripheral pads, completing the device fabrication.
[0058] 2. Calculation Methods and Workflow
[0059] The working steps are as follows:
[0060] (1) Initialization and mode selection: First, a skyrmion is initialized in the nano-confined structure by electrical or magnetic methods. The system controller selects the operating mode according to the received task instructions (such as "perform waveform transformation" or "perform sequence prediction").
[0061] If mode one (strong nonlinear mode) is selected, the current driving source is turned on, applying an alternating current with a frequency on the order of MHz or sub-MHz. The spin-transfer torque generated by this current drives the skyrmion to move along the long axis of the strip. Under appropriate current amplitude, the skyrmion enters a large-amplitude relaxation oscillation state, and its position-time trajectory exhibits strong nonlinear characteristics, making it suitable for handling tasks requiring complex nonlinear mappings.
[0062] If mode two (strong memory mode) is selected, the current drive is disconnected, and the radio frequency current source of the planar microcoil is connected, applying an alternating current of a specific frequency to generate a vertical alternating magnetic field. This magnetic field directly excites the breathing mode of the skyrmions, causing their radius to change periodically while the center remains relatively stable. This mode exhibits slower dynamic decay and contains richer historical information, making it suitable for prediction tasks.
[0063] (2) Signal injection: The timing input signal to be processed is applied to the drive signal of the selected mode by amplitude modulation. That is, the amplitude of the input signal is used to modulate the amplitude of the drive current or drive magnetic field in real time.
[0064] (3) State Acquisition: Under the excitation of the modulated driving signal, the skyrmion generates a corresponding dynamic response. The voltage signals of all channels are acquired synchronously and continuously through a segmented anomalous Hall effect readout array. These signals directly reflect the state of the skyrmion at different locations and times, constituting the multi-channel physical state signals of the physical reservoir.
[0065] (4) Calculation and output:
[0066] Training Phase: For supervised tasks, multi-channel physical state signals corresponding to the input signals and the desired target output sequence are collected over a period of time. An optimal readout layer weight matrix is trained using linear algorithms such as ridge regression. During this phase, for prediction tasks, VCMA stabilization is enabled: a DC gate voltage is applied in the central region of the device, locally altering the magnetic anisotropy and forming a shallow potential well to gently confine the skyrmion's center of mass, thereby suppressing the impact of room-temperature thermal drift on long-term prediction stability.
[0067] Inference phase: Fix the trained readout weights. For a new input signal, repeat steps (2) and (3) to obtain the real-time multi-channel physical state signal. Multiply it with the readout weight matrix to obtain the final calculation result (such as the transformed waveform, the predicted future value, etc.).
[0068] 3. Effects and advantages of mode selection
[0069] The core innovation of this invention lies in its mode selection mechanism, which endows a single physical device with unprecedented flexibility in task adaptation:
[0070] When handling tasks such as waveform transformation, speech recognition, and image recognition, the system switches to Mode 1. Here, the strong nonlinear characteristics of the skyrmion's large-amplitude oscillations are fully utilized, mapping the input to the high-dimensional nonlinear space of the skyrmion spring oscillator magnetic system, efficiently completing complex signal transformation functions.
[0071] When handling tasks such as chaotic time series forecasting, financial time series forecasting, meteorological and environmental data forecasting, and biomedical signal forecasting, the system switches to Mode 2. Skyrmion respiratory model dynamics exhibit a longer "memory effect," meaning the current response incorporates more information from past inputs. This characteristic allows the reservoir to learn and infer dependencies in the time series, thus achieving high-precision forecasts. VCMA-assisted stabilization further ensures the reliability of long-term forecasts at room temperature.
[0072] By switching between the two modes using simple electrical commands, the same device essentially possesses two different "computing personalities," enabling it to handle drastically different computing tasks with optimized performance. This avoids the hassle of preparing different hardware for different tasks or performing tedious parameter readjustments.
[0073] 4. Room temperature simulation results and analysis
[0074] (1) Simulation results of sine wave to square wave conversion under large amplitude oscillation mode:
[0075] At a frequency f = 0.2MHz and a current density j = 1.5×10 12 A simulation experiment was conducted to convert a sine wave to a square wave in a large-amplitude oscillation mode of A / m². The simulation results are shown in Figure 3, where the upper figure shows the conversion effect during the training phase (first 70% of the data). The mean square error (MSE) between the target square wave and the predicted square wave is 9.025 × 10⁻⁶. -4 The figure below shows the conversion effect during the testing phase (last 30% of the data). The mean square error (MSE) between the target square wave and the predicted square wave is 1.096 × 10⁻⁶. -3 As can be seen from the figure, the predicted square wave and the target square wave have a very high degree of similarity, which fully verifies that the strong nonlinear characteristics of this mode can efficiently complete complex signal conversion tasks.
[0076] (2) Simulation results of Mackey-Glass signal prediction under a respiratory model: A simulation experiment of Mackey-Glass signal prediction was conducted using a respiratory model driven by an alternating magnetic field with a frequency of f = 31.8 GHz. The simulation results are shown in Figure 4, where the upper figure shows the prediction effect at room temperature (300 K) during the training phase (first 70% of the data). The mean square error (MSE) between the target signal and the predicted signal is 8.027 × 10⁻⁶. -3 The figure below shows the prediction performance during the testing phase (last 30% of the data) at room temperature (300K). The mean square error (MSE) between the target signal and the predicted signal is 1.382 × 10⁻⁶. -2 Simulation results show that the strong memory characteristics of the respiratory model can effectively capture the temporal dependence of the Mackey-Glass signal, achieving high-precision signal prediction at room temperature.
[0077] 5. Extended Implementation Method
[0078] This invention is not limited to the specific embodiments described above. For example:
[0079] The core magnetic material can be selected from other thin film systems with strong DMI (Dzyaloshinskii-Moriya Interaction) and vertical anisotropy.
[0080] The exchange coupling strength of the SAF and the geometry of the nanoconfined domain can be optimized to adjust the potential energy distribution and the eigenmode frequencies of the skyrmions.
[0081] The readout method can be expanded, for example, by combining it with a magnetic tunnel junction to achieve higher sensitivity electrical readout.
[0082] Multiple such mode selection devices can be interconnected to form more complex reservoir networks or hierarchical computing systems.
[0083] The pattern selection logic can be combined with upper-layer application algorithms to achieve adaptive and intelligent task-pattern matching.
Claims
1. A method for calculating a room temperature mode-selective physical reservoir, characterized in that, The method, based on a skyminzi spring oscillator device incorporating an artificially synthesized antiferromagnetic structure, includes the following steps: S1. Initialization: A room-temperature skyrmion is generated and stabilized as the initial state within the nano-confined magnetic region of the skyrmion spring oscillator. S2. Mode Selection: Select the corresponding physical driving mode according to the type of computation task to be performed to excite the specific dynamic mode of the skyrmion: configure the multi-channel readout circuit and set the sampling parameters; For signal conversion tasks, the first driving mode is selected, and a low-frequency alternating current is applied along the restricted motion direction of the skyrmion. The skyrmion is driven into a large-amplitude oscillation mode through the spin-transfer torque. For signal prediction tasks, the second driving mode is selected, and an alternating magnetic field of a specific frequency is applied along the direction perpendicular to the device plane to excite the breathing mode of the skyrmion; S3, Input Loading: The timing input signal is amplitude-modulated and loaded onto the selected drive signal; S4. State Sampling and Matrix Construction: Driven by the input signal, the physical state signals of each channel are synchronously acquired through a multi-channel readout circuit. S5. Training: Train the linear readout layer weights using the multi-channel physical state signal and the target output. S6. Use the readout layer weights to perform computation tasks.
2. The method for calculating a room temperature mode-selective physical reservoir according to claim 1, characterized in that, In step S2, when the second driving mode is executed, voltage-controlled magnetic anisotropy technology is enabled, and a DC gate voltage is applied to a specific region of the skyrmion spring oscillator device to locally modulate the vertical magnetic anisotropy and form a potential well for suppressing skyrmion thermal drift.
3. The method for calculating a room temperature mode-selective physical reservoir according to claim 1 or 2, characterized in that, The signal conversion tasks include waveform transformation, speech recognition, and image recognition; the signal prediction tasks include chaotic time series prediction, financial time series prediction, meteorological and environmental data prediction, and biomedical signal prediction.
4. The method for calculating a room temperature mode-selective physical reservoir according to claim 1, characterized in that, In step S2, for the first driving mode, the alternating current frequency must be low enough and the amplitude must be large enough; for the second driving mode, the specific frequency is the intrinsic frequency of the breathing mode.
5. The method for calculating a room temperature mode-selective physical reservoir according to claim 1, characterized in that, In step S3, for the first driving mode, the amplitude of the alternating current is modulated by the input signal; for the second driving mode, the amplitude of the alternating magnetic field is modulated by the input signal.
6. The method for calculating a room temperature mode-selective physical reservoir according to claim 1, characterized in that, In step S4, the multi-channel readout circuit is a segmented anomalous Hall effect probe array arranged along the Skyrmion motion path, and the physical state signal is the Hall resistance signal measured by each probe.
7. A room-temperature mode-selective physical reservoir computing device for implementing the method of any one of claims 1-5, characterized in that, It is a multi-layer stacked structure, consisting of, from bottom to top: a substrate, a buffer layer, a bottom planar microcoil, a bottom voltage-controlled magnetic anisotropic structure, an artificially synthesized antiferromagnetic nanostructure, a top voltage-controlled magnetic anisotropic structure, a top planar microcoil, and a protective layer. It is patterned into submicron or nanometer confined geometry, forming the core of the skyrmion spring oscillator. A segmented anomalous Hall effect readout array is arranged on the side of the artificially synthesized antiferromagnetic nanostructure; current-driven electrodes are connected to both ends of the artificially synthesized antiferromagnetic nanostructure; VCMA voltage controls the magnetic anisotropy intensity; planar microcoils control the alternating magnetic field; a protective layer and interconnection structure cover the outermost layer and provide electrical connection channels.
8. The device according to claim 6, characterized in that, The segmented anomalous Hall effect readout array consists of multiple independent metal electrode pairs arranged on both sides along the long axis of the artificially synthesized antiferromagnetic nanostructure.
9. The device according to claim 6, characterized in that, Both the bottom and top voltage-controlled magnetic anisotropy structures include a gate electrode and a gate dielectric layer, wherein the gate dielectric layer is a high dielectric constant dielectric.
10. The device according to claim 6, characterized in that, The bottom and top planar microcoils are used to generate an alternating magnetic field perpendicular to the device plane.