Reservoir calculation based on ferromagnetic film with point deformation

Abstract

The invention particularly relates to a physical reservoir for a magnetic reservoir computing device. The physical reservoir includes a ferromagnetic film including a two-dimensional arrangement of point deformations. The point deformations are dimensioned to function as pinning sites of magnetic domains of the ferromagnetic film. The invention also relates to a magnetic reservoir computing device comprising such a physical reservoir, and methods of operating and manufacturing such a reservoir computing device. The proposed method results in a low power consumption physical reserve that is easy to manufacture.

Description

[0001] Statement regarding federally funded research or development

[0002] The project that led to this application has received funding from the EU Horizon 2020 Research and Innovation Program, which is part of the FET-Open SpinENGINE project, grant number 861618. Technical Field

[0003] This invention generally relates to the field of reservoir computing. In particular, the invention relates to a physical reservoir for a magnetic reservoir computing device, a magnetic reservoir computing device including a physical reservoir, a method of operating a reservoir computing device, and a method of manufacturing a reservoir computing device. Background Technology

[0004] Reservoir computing involves machine learning computation techniques where input signals are fed into a fixed, stochastic, dynamic system (a reservoir). The dynamic properties of the reservoir map the input to a higher-dimensional space. A simple readout mechanism (e.g., a single output layer) reads the state of the reservoir and maps that state to the output. Training is performed only during the readout phase, and the reservoir remains fixed. This method leverages the computational power enabled by the dynamic properties of a fixed reservoir while reducing the computational cost of training.

[0005] The reservoir can be virtual (e.g., a randomly generated network) or physical. Physical reservoirs often utilize the nonlinearity of certain natural systems. Various types of physical reservoirs have been proposed, including electronic, photonic, mechanical, chemical, and liquid reservoirs.

[0006] Recently, magnetic reservoirs have been proposed, which rely on nanomagnets / spintronic oscillators. The challenge of magnetic reservoirs is to achieve low-power physical reservoirs that are still easy to fabricate. Summary of the Invention

[0007] According to a first aspect, the present invention is embodied as a physical reservoir for a magnetic reservoir computing device. The physical reservoir includes a ferromagnetic film comprising a two-dimensional arrangement of point deformations. The dimensions of the point deformations are determined as pinning sites for magnetic domains of the ferromagnetic film.

[0008] For example, point deformation can be perforations or point protrusions. Perforations are preferred because they are easier to obtain than point protrusions. Point deformation in patterned ferromagnetic films allows for robust, highly nonlinear, and stochastic responses of the ferromagnetic film to external stimuli. Due to such point deformation, the physical reservoir is nonlinear, stochastic, and exhibits decaying memory. Furthermore, single-point deformation is easier to obtain than previously proposed nanomagnets / spintronic oscillators. Additionally, point deformation requires less power to set the reservoir state. Therefore, this method can be adequately used for reservoir computation.

[0009] The structural properties of the membrane can be optimized to reduce the power consumption of the physical reservoir. In this regard, in embodiments, the average first nearest neighbor distance between perforations is between 0.2 μm and 5.0 μm, preferably between 0.5 μm and 3.0 μm. Any first nearest neighbor distance between perforations is measured parallel to the main surface of the ferromagnetic membrane. The average diameter of the perforations is between 50 nm and 500 nm, preferably between 70 nm and 380 nm, noting that the average diameter of the perforations is smaller than the average first nearest neighbor distance. Any diameter of the perforations is again measured parallel to the main surface of the ferromagnetic membrane.

[0010] In this embodiment, the thickness of the ferromagnetic film is between 2 nm and 50 nm. Preferably, the thickness is between 5 nm and 20 nm. In this embodiment, the length of each side of the ferromagnetic film is between 20 μm and 200 μm. The physical reservoir typically includes a substrate supporting the ferromagnetic film.

[0011] In embodiments, the two-dimensional arrangement forms at least one lattice of dot deformation, such as a square lattice. Dot-deformed lattices can be readily obtained, for example, using nanoimprinting or nanotemplating techniques. Furthermore, any or every lattice in the lattice can be formed as an inverse dot lattice. Moreover, by introducing “defects” into the dot-deformed two-dimensional arrangement, the nonlinearity and randomness of the reservoir response can be enhanced. For example, the two-dimensional arrangement can form at least two dot lattices with different lattice parameters of dot deformation in different regions of the ferromagnetic film.

[0012] In some embodiments, and preferably, the ferromagnetic film comprises one or more elements selected from the group consisting of transition metal elements. Preferably, the ferromagnetic film comprises Fe. x Ni 100-x Where 20 ≤ x ≤ 60. This composition can be adjusted to reduce saturation magnetization and thus reduce power consumption in the reservoir. For example, the membrane could contain Fe. 55 Ni 45 .

[0013] According to another aspect, the present invention is embodied as a magnetic reservoir computing device. The device includes a physical reservoir, a drive system, and a readout unit. Consistent with the foregoing aspect, the physical reservoir includes a two-dimensionally arranged ferromagnetic film with point deformations, the dimensions of which are determined to serve as pinning sites for the magnetic domains of the ferromagnetic film. Furthermore, the drive system is configured to saturate the ferromagnetic film and couple an input signal into the ferromagnetic film to set the magnetic state of the ferromagnetic film. The readout unit is operatively connected to the physical reservoir to read an output signal from the ferromagnetic film.

[0014] In some embodiments, and preferably, the drive system includes two or more coils configured to apply a rotating magnetic field coplanar with the ferromagnetic film, thereby adapting the drive system to couple an input signal into a physical reservoir.

[0015] In some embodiments, the drive system is operatively connected to the ferromagnetic membrane to apply one or more current signals to the ferromagnetic membrane. For example, the physical reservoir may include an arrangement of electrical conductors connecting the drive system to multiple locations within the ferromagnetic membrane. In this case, the drive system is configured to apply current signals to the ferromagnetic membrane via the electrical conductors to locally generate a magnetic field coplanar with the ferromagnetic membrane.

[0016] For example, the drive system can be operatively connected to apply the current signal to the ferromagnetic film to move the domain walls in the ferromagnetic film according to one of the spin-transfer torque mechanism and the spin-orbit torque mechanism.

[0017] In some embodiments, and preferably, the apparatus further includes a processing unit connected to the readout unit and configured to further process the signal from the storage pool.

[0018] According to another aspect, the present invention is embodied as a method of operating a reservoir computing device. The method includes setting the magnetic state of a physical reservoir. Similarly, the physical reservoir includes a ferromagnetic membrane arranged in a two-dimensional pattern with point deformations, the dimensions of which are determined to serve as pinning sites for magnetic domains of the ferromagnetic membrane. The magnetic state is set by saturating the ferromagnetic membrane and coupling an input signal into the physical reservoir to generate one or more magnetic fields coplanar with the ferromagnetic membrane. In an embodiment, the method further includes reading an output signal from the physical reservoir and processing the output signal to obtain one or more inference results, such as classification results.

[0019] According to the last aspect, the present invention is embodied as a method for manufacturing a reservoir computing device. The method revolves around patterning a ferromagnetic film of a physical reservoir. Furthermore, the method includes connecting the physical reservoir (which ultimately comprises the ferromagnetic film) to one or more additional components of the reservoir computing device. Consistent with the previous aspects of the invention, the ferromagnetic film is patterned to obtain a two-dimensional arrangement of point deformations in the ferromagnetic film, wherein the dimensions of the point deformations are determined as pinning sites for magnetic domains of the ferromagnetic film to obtain the physical reservoir.

[0020] In some embodiments, the ferromagnetic film is patterned using nanoimprint lithography. In other embodiments, nanotemplate technology is used to pattern the ferromagnetic film. Attached Figure Description

[0021] These and other objects, features, and advantages of the present invention will become apparent from the following detailed description of illustrative embodiments thereof, which is taken in conjunction with the accompanying drawings. These descriptions are intended to clearly assist those skilled in the art in understanding the invention in conjunction with the detailed description. In the drawings:

[0022] Figure 1A This is a schematic diagram of selected components of a reservoir computing device according to an embodiment of the present invention, illustrating how input data is mapped onto a physical reservoir and read out to generate classification results. The physical reservoir includes a ferromagnetic membrane, such as... Figure 3A – Figure 4E As shown;

[0023] Figure 1B yes Figure 1A A high-level schematic diagram of the reservoir computing device, illustrating the operational connections between the device's components;

[0024] Figure 2 This is a diagram of an in-memory computing device with a cross-array structure according to an embodiment of the present invention, which can be used for, for example Figure 1A and Figure 2 The output of the storage pool calculation device shown in Figure A;

[0025] Figure 3A This is an exploded view of the layer structure of a physical reservoir according to an embodiment of the present invention, the physical reservoir involving a ferromagnetic film patterned to include two-dimensionally arranged point deformations (e.g., perforations).

[0026] Figure 3B yes Figure 3A A top view of the ferromagnetic membrane in the physical storage tank;

[0027] Figure 3C According to an embodiment of the present invention Figure 3A A 3D view of a portion of the ferromagnetic membrane in a physical reservoir, showing patterned electrical conductors on opposite sides of the membrane (in... Figure 3A (Also visible in the text);

[0028] Figure 4A This is a first 2D cross-sectional view of a portion of a ferromagnetic film, illustrating how the ferromagnetic film can be locally processed to form point deformations according to an embodiment of the invention. The size of the point deformations is determined as pinning sites for magnetic domains of the ferromagnetic film.

[0029] Figure 4B This is a second 2D cross-sectional view of a portion of the ferromagnetic film, illustrating how the ferromagnetic film can be locally processed to form point deformations according to an embodiment of the invention. The size of the point deformations is determined as pinning sites for the magnetic domains of the ferromagnetic film.

[0030] Figure 4C This is a third 2D cross-sectional view of a portion of the ferromagnetic film, illustrating how the ferromagnetic film can be locally processed to form point deformations according to an embodiment of the invention. The size of the point deformations is determined as pinning sites for the magnetic domains of the ferromagnetic film.

[0031] Figure 4D This is a fourth 2D cross-sectional view of a portion of the ferromagnetic film, illustrating how the ferromagnetic film can be locally processed to form point deformations according to an embodiment of the invention. The size of the point deformations is determined as pinning sites for the magnetic domains of the ferromagnetic film.

[0032] Figure 4E This is a fifth 2D cross-sectional view of a portion of the ferromagnetic film, illustrating how the ferromagnetic film can be locally processed to form point deformations according to an embodiment of the invention. The dimensions of the point deformations are determined as pinning sites for the magnetic domains of the ferromagnetic film. Figure 4A – Figure 4D In the example, point deformation is formed as perforation, while Figure 4E In the examples, they are formed as protrusions;

[0033] Figure 5A The permalloy Fe according to an embodiment of the present invention 55 Ni 45 The secondary electron map of the inverse dot lattice, obtained by "drilling" holes using a focused ion beam tool. For description, a threshold function is applied. Figure 5A Topographic map;

[0034] Figure 5B It shows relative to Figure 5A The corresponding in-plane magnetization distribution, where the in-plane magnetization direction is indicated by an arrow.

[0035] Figure 6 It represents the anisotropic magnetoresistance (AMR) of ferromagnetic films (such as... Figure 5A The graph (shown) is a function of the applied magnetic field, illustrating the magnetic response of the ferromagnetic film to a cyclic magnetic field for various lattice constants.

[0036] Figure 7 This is a top view of the ferromagnetic membrane of the physical storage tank according to an embodiment of the present invention, which is similar to... Figure 3B ,Apart from Figure 7 The membrane now includes different lattices of point deformation (e.g., perforation) in different regions of the membrane, where the lattices have different lattice parameters;

[0037] Figure 8 This is a flowchart illustrating the advanced steps of a method for setting the magnetic state of a ferromagnetic film according to an embodiment of the present invention.

[0038] Figure 9 This is a flowchart illustrating the advanced steps of a method for operating a storage pool computing device according to an embodiment of the present invention.

[0039] Figure 10 This is a flowchart illustrating the advanced steps of a method for manufacturing a ferromagnetic film for a physical storage pool of a computing device according to an embodiment of the present invention.

[0040] The accompanying drawings illustrate simplified representations of apparatus, devices, or portions thereof as described in the embodiments. Technical features depicted in the drawings are not necessarily drawn to scale. Unless otherwise indicated, similar or functionally similar elements in the drawings are assigned the same reference numerals.

[0041] The physical reservoir, magnetic reservoir computing apparatus, and method embodying the present invention will now be described by way of non-limiting examples. Detailed Implementation

[0042] The structure described below is as follows. Section 1 describes general embodiments and advanced variations. Section 2 includes a detailed description of the accompanying drawings. All reference numerals Sn indicate... Figure 8 – Figure 10 The flowchart describes the method steps, while the digital labels refer to the devices, components, and concepts involved in the embodiments of the present invention.

[0043] 1. General embodiments and advanced variations

[0044] This section presents a physical reservoir (Section 1.1), a magnetic reservoir computing device (Section 1.2), a method for operating the reservoir computing device (Section 1.3), and a manufacturing method (Section 1.4), all of which are based on embodiments.

[0045] 1.1 Physical Storage Pool

[0046] 1.1.1 Main Features

[0047] Now for reference Figure 1A , Figure 1B , Figure 3A , Figure 3B and Figures 4A to 5BA first aspect of the invention is described in detail. This aspect relates to a physical reservoir 22 for a magnetic reservoir computing device 20. According to an embodiment of the invention, the physical reservoir 22 includes a ferromagnetic membrane 222, wherein the membrane includes a two-dimensional (2D) arrangement of dot deformities 225, such as perforations. The dimensions of the dot deformities 225 are determined to serve as pinning sites for the magnetic domains of the ferromagnetic membrane 222.

[0048] 1.1.2 Notes and Definitions

[0049] As described in the background section, reservoir computing (RC) is a computational framework that utilizes the nonlinear dynamic characteristics of reservoirs. In this case, the reservoir is a physical reservoir comprising a ferromagnetic film (also called a "film") containing ferromagnetic material. As is common to RC, the basic idea is to map the input signal into a high-dimensional computational space embodied by the reservoir. The reservoir is a fixed nonlinear system containing, for example,... Figure 1A The diagram illustrates a network of storage pools. The physical storage pool 22 is intended for use in a device 20 employing a simple readout mechanism. This device 20 relates to another aspect of the invention, which will be described in detail later.

[0050] The readout unit 23 includes or is connected to the processing unit 10. The processing unit 10 implements a computational model that can be trained to read the state of the reservoir and learn how to map that state to a desired output (e.g., a label), as in a typical supervised setup. For inference purposes, after an input signal is fed into the reservoir (which remains fixed), the readout unit 23 reads the state of the reservoir, and the processing unit 10 maps it to an output based on the learned parameters. A key advantage is that training is performed only at the readout level, while the dynamic characteristics of the reservoir remain fixed. That is, as is common to RC, this method utilizes the inherent computational power of the reservoir. The main difference from existing RC methods is that, in this case, the reservoir comprises a patterned ferromagnetic membrane 222, as described in detail below.

[0051] The ferromagnetic film 222 comprises 2D arrangements of dot deformations 225, such as perforations. The 2D arrangement of the dot deformations 225 can advantageously form a lattice, i.e., a deformed ordered arrangement (2D array). In some embodiments, the 2D arrangement forms a disordered (or non-constant) arrangement of the dot deformations 225. The deformed 2D arrangement can, for example, form a polycrystalline arrangement of the dot deformations 225. That is, the 2D arrangement can combine several lattices with different lattice step sizes and / or different symmetries. One or more lattices may also intentionally include localized lattice defects to create specific nodes in the reservoir network embodied by the film 222. Using deformed lattices makes it easier to pattern the ferromagnetic film 222.

[0052] In an embodiment, such as Figure 4EAs shown, the dot deformation 225 is obtained as a dot protrusion 225e. In some embodiments, as Figure 4A – Figure 4D As shown, preferably, the point deformation 225 is obtained as perforations 225a, 225b, 225c, and 225d. In all cases, the point deformation 225 is a single-point, localized, and compact deformation of the ferromagnetic film, whereby a small amount of material is locally removed, displaced, or added to create depressions, pits, holes, or protrusions. In practice, perforations 225a–225d are easier to obtain than point protrusions 225e. Perforations are small holes and can be, for example, as shown in the diagram. Figure 4A Through hole 225a, such as Figure 4B – Figure 4D Blind holes or shallow pits 225b–225d. See also Figure 4D The perforations on one side of membrane 222d can further cause corrugation deformation.

[0053] Generally, as illustrated below, except where the perforation size is much smaller, the shape of the perforation can resemble the deformation obtained by drilling, boring, punching, or piercing (at least partially) or otherwise deforming a plate-like workpiece (e.g., a metal sheet) at a single point. That is, in this case, nano- or micro-deformation is obtained, for example, by nanoimprinting or nanostamping techniques. As used herein, except in cases where the perforation size is much smaller, the shape of the perforation can resemble the deformation obtained by drilling, boring, punching, or piercing (at least partially) or otherwise deforming a plate-like workpiece (e.g., a metal sheet). That is, in this case, nano- or micro-deformation is obtained, for example, by nanoimprinting or nanostamping techniques. Figure 4A – Figure 4E In addition to the different types of deformations shown in the supporting description being indicated by reference numerals 225a–225e, point deformation 225 is generally indicated by reference numeral 225.

[0054] Unlike nanorings or other similar processed shapes, the dot deformities 225 considered herein do not require complex photolithographic patterning steps. They can be obtained through simple processing steps. In particular, perforations can be obtained through a single processing step, for example, due to nanotemplates or nanoimprint lithography. The final shape of the deformed part differs from the ring deformities proposed in the prior art in terms of shape and compactness. That is, as their name suggests, the dot deformities 225 form more compact recesses and / or protrusions.

[0055] Ideally, any in-plane portion of a point deformation (i.e., the portion parallel to the main surface of the membrane) is essentially formed in a compact region at the periphery defined by a simple closed curve. This ideal portion can be drawn, for example, enclosing a region such as... Figure 4A , Figure 4BThe hollow region is a circle or a simple polygon, i.e., a polygon that does not intersect and has no holes. The surface of this part is closed at the outer periphery by the boundary formed by the outer periphery of the ferromagnetic films 222a and 222b. The in-plane portion of the point deformation 225 does not need to be constant along the z-axis (i.e., the out-of-plane axis) perpendicular to the main surface (x, y), depending on the process used to obtain the deformation. Therefore, ideally, the shape of the perforations 225a–225d will be a substantially convex object, for example, as Figure 4A cylindrical, such as Figure 4B Partial spherical, or like Figure 4C A pyramidal shape. For example... Figure 4D As shown, localized deformation can cause membrane 222d to bulge (or otherwise protrude) on the side of the membrane opposite to the perforation 225d. For example... Figure 4E The protrusions 225e shown are generally more difficult to obtain than the perforations 225a–225d because they require material to be deposited on top of the membrane.

[0056] certainly, Figure 4A – Figure 4E The deformations shown are illustrative and for purposes of explanation only; they are not necessarily drawn to scale. Furthermore, as... Figure 5A As shown, the required process steps will typically result in an "imperfect" shape. However, the obtained deformation essentially retains the point deformation.

[0057] Depending on the manufacturing process used, the extent or degree of deformation of membranes 222b–222e at the deformation center (or interior) can be more important than at the boundaries. That is, the degree of deformation decreases from the center of deformation towards its periphery. For example, the extent of the depression at the center of blind vias 225b–225d is more important than the extent of the depression at its periphery. (See also: through-holes) Figure 4A This is an extreme case where the degree of deformation varies from the maximum level (inside the hole) to the minimum (i.e., zero) deformation outside the hole (i.e. at the boundary), which is the level of the undeformed membrane at the outer periphery of the hole.

[0058] 1.1.3 Advantages

[0059] Notably, the point deformations 225 serve as pinning sites for the magnetic domains of the ferromagnetic film 222. These induce emergent behavior, whereby dipole / exchange coupling between adjacent nanomagnets leads to a variety of complex, collective phenomena that can be used for reservoir calculations. More precisely, the ferromagnetic film 222 allows emergent behavior to be directly driven by applying (or otherwise generating) a magnetic field or spin torque. The drive system 21 can be used to utilize a strength of H... satThe 2D array is initialized by an in-plane magnetic field pulse. Emergent behavior is driven and tuned by the applied field strength. For example, by applying a rotating magnetic field of appropriate amplitude, the magnetic state of the physical reservoir 22 is set, which demonstrates emergent behavior. This is caused by point deformation (pinning sites); the interaction of domain walls on the deformed lattice produces an emergent nonlinear change in array magnetization and the number of domain walls with respect to the rotating field strength.

[0060] The point deformation 225 allows for a robust and highly nonlinear response of the ferromagnetic film 222 to external stimuli. Furthermore, the 2D arrangement of the point deformations induces a decaying memory of previous magnetization states. Therefore, the ferromagnetic film 222 with the 2D arrangement of point deformations 225, as proposed herein, possesses all the properties required to realize a physical reservoir for magnetic reservoir calculations. Thus, such a film can be used as a physical reservoir. For example, data can be input by modulating the amplitude of a cyclic magnetic field, and the reservoir response can be measured as the normalized number of domain walls. In some embodiments, the reservoir response can also be measured as the total magnetization component along a selected axis.

[0061] In summary, the point deformities 225 serve as pinning sites for magnetic domains, making the physical reservoir 22 nonlinear and random (i.e., it responds randomly to external stimuli) and exhibiting decaying memory. Furthermore, as previously mentioned, single-point deformities are much easier to obtain than nanorings. Moreover, they require less power to set the reservoir state. For example, the write field in an inverse point array is typically in the range of 0.9 kA / m to 1.8 kA / m, which is a factor of 4 to 5 times smaller than that of a corresponding ring array.

[0062] All these advantages will now be described in more detail with reference to specific embodiments of the invention.

[0063] 1.1.4 Preferred Embodiment

[0064] Perforation. As described above, such as Figure 4A – Figure 4D As shown, the point deformation 225 is preferably implemented as perforations 225a–225d. Perforations 225a–225d may include through holes and / or blind holes. Figure 4A – Figure 4E In this context, d represents the average diameter of the point deformation, while D represents the first nearest neighbor distance.

[0065] Structural Characteristics. In embodiments, the average first nearest neighbor distance D between the perforations 225 is between 0.2 μm and 5.0 μm. In some embodiments, and preferably, the average first nearest neighbor distance D between the perforations 225 is between 0.5 μm and 3.0 μm. In embodiments, the average diameter d of the perforations 225 is between 50 nm and 500 nm. In some embodiments, and preferably, the average diameter d of the perforations 225 is between 70 nm and 380 nm. Using this size has advantages in terms of dip field, as discussed in detail later. The first nearest neighbor distance D between the perforations is measured parallel to the plane (x, y), i.e., parallel to the main surface of the ferromagnetic film 222 (i.e., within the plane of the ferromagnetic film 222). The diameter d of the perforations 225 is also measured parallel to the main surface of the ferromagnetic film 222. Note that the average diameter d of the perforations must be less than the average first nearest neighbor distance D, preferably at least twice as small. When a 2D arrangement forms a square or hexagonal lattice, the average first nearest neighbor distance D corresponds to the lattice constant (also known as the lattice step size in the literature).

[0066] Dot matrix. As described above, such as Figure 3A , Figure 3B , Figure 5A and Figure 5B As assumed, the 2D arrangement of the dot deformors 225 can form a lattice, such as a square lattice. In some embodiments, the lattice can be a tilted, rectangular, or hexagonal lattice. For example, the effectiveness of a hexagonal lattice has been experimentally demonstrated. The lattice of the dot deformors 225 can, for example, be formed as an inverted dot lattice. Inverted dots are typically formed as through-holes. The ferromagnetic film 222 can therefore include a periodic array of through-holes, which correspondingly form a periodic array of pinning sites.

[0067] Lattice defects. In principle, the 2D arrangement of dot deformations 225 can form at least two lattices of dot deformations 225 in different regions of the ferromagnetic film 222, wherein the lattices have different lattice parameters. Figure 7 In the example, the 2D arrangement of point deformations 225 forms three different square lattices P1, P2, and P3 in different regions of the ferromagnetic film 222. Changing the lattice constant on the film enhances the dynamic characteristics of the nonlinear system formed by the ferromagnetic film 222. For the same reason, point deformations can form “microcrystals” with different orientations (having the same or different lattice constants). Furthermore, point defects can be added. In some embodiments, the lattices may partially overlap at their boundaries.

[0068] Ferromagnetic materials. In embodiments, the ferromagnetic film 222 comprises one or more elements selected from the group consisting of transition metal elements. The ferromagnetic film 222 may be, in particular, a pure metal, alloy, or compound, such as pure metal, alloy, or compound of iron, cobalt, nickel, and / or certain rare earth metals. In some embodiments, and preferably, the ferromagnetic film 222 is composed of ferromagnets from the 3d series of the periodic table (e.g., Fe, Co, Ni) in any composition. For example, the ferromagnetic film 222 may include Fe. x Ni 100-x Where 20 ≤ x ≤ 60. An example of a suitable composition is Fe. 55 Ni 45 (i.e., x = 55) and Fe 20 Ni 80 (i.e., x = 20). For example, Fe 20 Ni 80 The saturation magnetization M of permalloy s It is 800 kA / M; it varies with the membrane composition. The composition can be adjusted to reduce the saturation magnetization; a lower M... s The value is preferred to achieve low-power operation.

[0069] If needed, nonmagnetic elements can be alloyed with these ferromagnets to tune their magnetic properties. For low-power devices, the material should be relatively soft (i.e., with a low coercive field H). c Typical coercive fields measured on planar films are on the order of approximately 5 kA / m or less, although they are typically slightly larger on patterned films due to domain wall pinning at the deformed array. In magnetic reservoir calculations, complex emergent states are formed due to wall pinning. These are characterized by a decrease in magnetic response relative to the applied cyclic magnetic field. The lower this "decline" field, the lower the energy required to set the state, which is critical for designing low-power devices. In this respect, and as the inventors have observed, the decline field is reduced using the deformed dimensions and lattice step size as described above. For example, for inverse dot arrays with an average diameter of 100 nm to 110 nm, a decline field with a lattice constant of 2 μm to 3 μm is stable.

[0070] Film Size. The thickness of the ferromagnetic film 222 is typically between 2 nm and 50 nm. In some embodiments, and preferably, the thickness of the ferromagnetic film 222 is between 5 nm and 20 nm. Such thickness values ​​are generally sufficient to allow emergent behavior. Furthermore, whether by focused ion beam (FIB), nanoimprinting, or nanotemplating techniques, these methods facilitate the acquisition of perforations. The ferromagnetic film 222 can, for example, be deposited as a square or rectangular block, wherein the length of each side of the block is between 20 μm and 200 μm. For example, this length could be 80 μm.

[0071] like Figure 3AAs shown, the physical reservoir 22 may further include a substrate 224 supporting the ferromagnetic film 222. This substrate provides a foundation on which the ferromagnetic film 222 can be grown and patterned, thereby promoting mechanical stability. The resulting layer structure can then be appropriately arranged and connected in a magnetic reservoir computing device, as described in detail below.

[0072] 1.2 Magnetic Storage Pool Calculation Device

[0073] refer to Figure 1A , Figure 1B and Figure 9 Now, another aspect of the invention will be described in detail, which relates to a magnetic reservoir computing device 20 (or simply "device").

[0074] The device 20 includes a physical reservoir 22, a drive system 21, and a readout unit 23. The physical reservoir 22 is the reservoir described in Section 1.1. That is, it essentially includes a ferromagnetic film 222 having dot deformities 225 arranged in a 2D pattern, the size of which is determined to serve as pinning sites for the magnetic domains of the ferromagnetic film 222.

[0075] The drive system 21 is configured to magnetically saturate the ferromagnetic membrane 222 and couple the input signal into the ferromagnetic membrane 222. That is, in operation, the drive system 21 is used to set the magnetic state of the ferromagnetic membrane 222. Note that setting the magnetic state of the reservoir means saturating the ferromagnetic membrane 222 and coupling the signal into the reservoir. Therefore, the drive system 21 includes all the components required to both saturate the ferromagnetic membrane 222 and couple the input signal into the reservoir.

[0076] The driving stimulus is a magnetic field. For example, such as... Figure 1B Symbolically illustrated, a coil can be used to saturate the ferromagnetic membrane 222, and then the magnetic field is rotated to couple an input signal (capture input data) into the ferromagnetic membrane 222. Applying a magnetic field to a ferromagnetic membrane using a coil is known to those skilled in the art. In this embodiment, an electronic signal (e.g., a voltage amplitude level) is mapped to a magnetic field amplitude and offset, such that the magnetic field conforms to a field range in which the magnetic array responds optimally to produce a highly complex magnetic state.

[0077] The readout unit 23 is operatively connected to the physical reservoir 22 to read out the output signal from the ferromagnetic film 222. For example, readout can be performed in three different ways. The first possibility is to perform optical readout by utilizing the magneto-optical Kerr effect, which results in the measurement of the total magnetization component along the in-plane direction. The second possibility is to perform electrical readout, thereby measuring the anisotropic magnetoresistance (AMR), i.e.,, in a general sense, the signal counting the domain walls. The third possibility is to use spin-polarized scanning electron microscopy (spin-SEM), a method for observing magnetic domains that images the localized magnetization distribution. Note that for most practical applications, the last possibility is usually not a practical solution. However, it is very useful for verifying that the domain walls are indeed pinned at the point deformation 225.

[0078] like Figure 1A As shown, the readout unit 23 includes at least one readout circuit that collects signals from the physical storage pool 22, converts them into digital signals (if needed), and forwards the resulting signals to the external processing unit 10. In some embodiments, the processing unit 10 forms part of the readout unit 23. In both cases, the processing unit 10 uses appropriately trained weights w i Perform the final operation, that is, output the network (e.g., a single layer), as described in detail later. Figure 1A As assumed in the text, the processing unit 10 itself can communicate with the server 2, and one or more users 4 can connect to the server 2, for example, via a personal computer (PC) 3.

[0079] As described above, coils can be specifically used to set the state of the physical storage pool 22. That is, as... Figure 1B As assumed in the description, the drive system 21 may include two or more coils configured to apply a rotating (macroscopic) magnetic field. Thus, such coils can be used to apply a rotating magnetic field coplanar with the ferromagnetic film 222. In this way, the drive system 21 is suitable for coupling an input signal into the physical reservoir 22. At least two coils are required to generate the rotating field. Using coils to apply a rotating magnetic field coplanar with a given material layer is known to those skilled in the art.

[0080] In principle, the drive system 21 can be configured to couple several parallel signals (corresponding to different input values) to various nodes of the reservoir network. For example, several input electrodes can be arranged on one side of the ferromagnetic membrane 222, thereby coupling several signals into the membrane through the input electrodes. Similarly, several output electrodes can be arranged on the other side of the ferromagnetic membrane 222 (e.g., opposite to said side) to read out the output signals, for example, by AMR measurement.

[0081] In some embodiments, a single input is preferred, and the input value can be converted into a time series. Accordingly, an input signal can be generated that captures the time series. For example, such as... Figure 1A As shown, the input time series can be converted into a time-dependent voltage signal V. input In other words, the input data can be applied to the physical reservoir 22 in the form of a time-dependent electrical signal. This can be achieved, for example, by modulating the amplitude of the rotating magnetic field, i.e., according to H(t) = H0 + ΔH×V input The modulation field is implemented using H(t), where H(t) is the modulation field, H0 is the field center, and ΔH is the modulation amplitude.

[0082] Ultimately, just like Figure 1A As shown, the signal extracted from the physical storage pool 22 can be read as a voltage signal. Therefore, it was achieved from... Captured input signal to output signal The nonlinear transformation. Note that a smaller field center H0 means that the operation requires less energy, which is preferred for low-power operation of the reservoir. Conversely, a larger value of ΔH produces more available states. The decay memory timescale is determined by the magnetic pinning strength at the point deformation in the membrane and the input rate. Similar to Output signal Multiple values ​​are encoded, and these values ​​can be interpreted to form multiple signals corresponding to the respective values. That is, if necessary, the formed multiple signals can be converted into digital values. These values ​​or signals are then passed through an output network executed by the processing unit 10, as discussed in detail below.

[0083] In a variant of the coil that applies a rotating magnetic field, the state of the ferromagnetic diaphragm 222 can be set by applying a current. In this case, the drive system 21 is operatively connected to the ferromagnetic diaphragm 222 to apply one or more current signals to the ferromagnetic diaphragm 222. The current signals can be applied locally or globally to the ferromagnetic diaphragm 222. In particular, as... Figure 3A and Figure 3C As assumed in the text, a magnetic field can be locally generated by a conductor (such as the current-carrying lines used in standard magnetoresistive random access memory (MRAM) devices). For example, as... Figure 3C As shown, the physical reservoir 22 may further include an arrangement of electrical conductors 221, 223 connecting the drive system 21 to multiple locations within the ferromagnetic membrane 222. In this case, the drive system 21 is configured to apply a current signal to the ferromagnetic membrane 222 via the electrical conductors 221, 223 to locally generate a magnetic field coplanar with the ferromagnetic membrane 222.

[0084] Another possibility is to use a spin torque mechanism to move the domain walls. As is known to those skilled in the art, such a spin torque mechanism can rely on spin-transfer torque or spin-orbit torque effects. Therefore, in an embodiment, the drive system 21 is operatively connected to apply a current signal to the ferromagnetic film 222 to move the domain walls in the ferromagnetic film 222 according to a spin-transfer torque mechanism or a spin-orbit torque mechanism, which rely on the spin-transfer torque effect and the spin-orbit torque effect, respectively, as is known to those skilled in the art.

[0085] As described above, the magnetic reservoir computing device 20 may further include a processing unit 10, which is connected to or forms part of the readout unit 23. The processing unit 10 may include one or more conventional processors. It may form part of a conventional computer or server 2, and itself communicates data with one or more user computers 3. It may also be a non-conventional processing unit, as illustrated below. In all cases, the processing unit 10 is configured to further process signals (or values ​​corresponding to such signals) obtained from the physical reservoir 22 via the readout unit 23. In particular, the processing unit 10 may be configured to execute an output network based on such values ​​or signals, wherein the output network includes at least one output layer. Once the parameters of the output network have been learned, the output network can be used to perform inference based on signals or values ​​obtained from the physical reservoir 22.

[0086] For simplicity, assume the output network consists of a single output layer. In this case, assuming the readout signal is converted into different signals (or corresponding values), linear operations, i.e., vector-matrix operations, are required. Therefore, as... Figure 2 As shown, in some embodiments, the processor unit may advantageously be an in-memory compute (IMC) unit 10 having a cross-array structure 15. Such a cross-array implements one neural layer at a time and can therefore be advantageously used at the output of the reservoir. The IMC unit 10 itself may include a digital processing unit 18, for example, configured as a near-memory compute unit. This processing unit 10 may be integrated with other components of the device 20. Figure 2 The IMC unit 10 is described in detail in Section 2. In some embodiments, if desired, several IMC units 10 are cascaded to implement several output neural layers.

[0087] 1.3 Method for operating the storage tank calculation device

[0088] Now for reference Figure 8 and Figure 9 Another aspect of the invention relates to a method of operating a reservoir computing device 20, such as that described in Section 1.2. Figure 8As shown, the key feature of this method is the appropriate setting of the magnetic state of the physical reservoir 22 in steps S1 and S2. The magnetic state is set by first saturating the ferromagnetic film 222 in step S1, and then coupling the input signal into the physical reservoir 22. This results in the generation of one or more magnetic fields coplanar with the ferromagnetic film 222 in step S2. That is, as... Figure 8 As shown, the ferromagnetic film 222 needs to be saturated (magnetically) before the magnetic field is modulated to set the state of the reservoir. As described in the previous section, the state of the reservoir can be set by the drive system 21.

[0089] As previously described, a signal can be coupled into physical reservoir 22 by converting an electrical signal into a properly regulated magnetic field and then moving the domain walls to induce complex magnetic modes. Alternatively, spin torque can be used instead of a magnetic field. In this case, the input electrical signal is converted into a properly regulated current that moves the domain walls. In all cases, coupling the signal into physical reservoir 22 results in the establishment of complex magnetic states in the ferromagnetic film 222.

[0090] Note that a perpendicularly magnetized ferromagnet can also be an option, as long as the perpendicular anisotropy remains relatively moderate. In this case, the rotating magnetic field will circulate from outside the plane to inside the plane and back. Note that even with this change in coordinate system, the concept of setting up complex magnetic states for reservoir calculations remains essentially unchanged.

[0091] like Figure 9 As further seen in Section 1.2, after coupling the input signal to the physical reservoir 22 at step S20, the method may include reading out the output signal from the physical reservoir 22 at step S40 and further processing such a signal at step S50 to obtain one or more inference results. Note that it is assumed here that the output signal is processed for inference purposes. For training purposes, the signal is similarly processed during the forward pass. Furthermore, weight gradients are computed during the backward pass to adjust the weights.

[0092] In summary: Injecting a signal into physical reservoir 22 causes it to generate a response signal at step S30. At step S40, the response signal generated by physical reservoir 22 is read out by readout unit 23. The readout signal can then be injected (if necessary, after being converted to a digital value) into processing unit 10, which runs the simple network in the output at step S50, whether for training or inference purposes.

[0093] 1.4 Manufacturing Method

[0094] Now for reference Figure 10The final aspect of the invention relates to a method of manufacturing a reservoir computing device 20. This manufacturing method revolves around patterning a ferromagnetic film 222 in step S140 to obtain a 2D arrangement of dot deformables 225 in the ferromagnetic film 222. Consistent with the foregoing aspects of the invention, the dot deformables 225 must be sized to serve as pinning sites for the magnetic domains of the ferromagnetic film 222. This allows for the acquisition of a physical reservoir 22 comprising the ferromagnetic film 222. Finally, the physical reservoir 22 is connected to one or more additional components to obtain the reservoir computing device 20.

[0095] In an embodiment, at step S140, the ferromagnetic film 222 is patterned using nanoimprint lithography. Nanoimprint lithography is known to those skilled in the art; it is a low-cost, high-throughput, and high-resolution nanolithography process that produces patterns through mechanical deformation. For example, a microscopic rubber stamp can be used. In some embodiments, at step S140, the ferromagnetic film 222 is patterned using nanotemplate technology (i.e., a photolithography-free process using evaporation through a mask). For example, nanotemplate technology can be used to fabricate a dot matrix. An inverse dot matrix can also be fabricated using a mask. For example, a dot matrix of shallow pits can be achieved using a mask that is adjustable by adjusting the distance between the mask and the substrate. Nanoimprint lithography and nanotemplate technology are low-cost technologies that result in sufficient resolution for this purpose. Other technologies, such as focused ion beam (FIB) technology, can still be considered. FIB technology allows flexibility in the selection of the first nearest neighbor distance between the arrangement, diameter, and dot deformation.

[0096] The above embodiments have been briefly described with reference to the accompanying drawings, and various modifications can be adapted to these embodiments. Several combinations of the above features can be considered. Examples are given in the next section.

[0097] 2. Detailed description of the attached figures

[0098] Figure 1A This is a schematic diagram of components 10, 21, 22, and 23 of the illustrated reservoir computing device 20 according to an embodiment. Figure 1BThis is a high-level schematic diagram of the same device, illustrating the connections between components 2, 3, 10, 21, 22, and 23. In this example, the device's drive system 21 includes two coils used to set the state of the physical reservoir 22. Specifically, user 4 interacts with server 2 via PC 3 to instruct the execution of inference based on the reservoir-based computational network. Input data (e.g., a vector x with vector components x1, x2, x3, x4) is loaded into drive system 21. Such input data may have been obtained from initial user data using feature extraction and dimensionality reduction, as is common in the art. Drive system 21 first converts such input data into a time series, which is captured as a time-dependent electrical signal. The electrical signal is coupled to the physical reservoir 22 (representing the reservoir network), causing the reservoir to respond non-linearly. For example... Figure 3A – Figure 4E The physical storage tank 22 shown includes a ferromagnetic membrane 222 (in Figure 1A (Not visible in the middle). The ferromagnetic film 222 is assumed to have been previously saturated ( Figure 8 (Step S1 in the process). The readout unit 23 reads the response signal, converts the response signal into an output signal, and feeds the signal to the processing unit 10. The processing unit 10 runs an output neural layer on the signal to generate a classification result (a binary classification in this example), which is returned to the server 2 and then to the user PC 3.

[0099] Figure 2 This is a schematic diagram of an IMC unit 10 with a cross array structure 15, which can be used in, for example... Figure 1A and Figure 2 In the output of the reservoir computing device 20 shown in Figure A, the IMC unit 10 includes N input lines 151 and M output lines 152, which are interconnected at intersections (i.e., nodes). The intersections correspondingly define N×M units 154, also called unit cells. The input and output lines are interconnected via a memory system 156. The IMC unit 10 contains at least two input lines and at least two output lines (i.e., N≥2 and M≥2). In practice, the number of effective input lines 151 and effective output lines 152 will depend on the dimension of the problem to be solved. That is, the number of effective input lines N' (N'≤N) corresponds to the number of signals (or values) obtained from the reservoir network, while the number of effective output lines M' (M'≤M) corresponds to the number of outputs required to perform the desired inference (i.e., classification or prediction). Each IMC unit 10 implements up to M neurons at a time.

[0100] The storage element can be a digital storage device. In this case, the signal read from the storage pool needs to be converted into a digital value first. In some embodiments, the storage element is an analog storage device, which can be, in particular, a phase-change memory (PCM) device, a resistive random access memory (RRAM) device, or a flash memory cell device. Such a device is used to map weight values ​​over the conductance range of a single storage element. Conversely, multiple binary devices representing different weight bits will be used for digital operations.

[0101] Using analog memory devices, the input vector (corresponding to the signal output from the reservoir) is encoded as a signal applied by input cell 11 to input line 151 of the cross array to perform a multiplication-accumulation (MAC) operation. The coefficients (“weights”) of the matrix are stored in the columns of the cell. Following each column of the cell is a column of arithmetic units (not shown), which multiplies the weights by the input vector values ​​(producing partial products) and finally accumulates all partial products to produce a full dot product. This architecture maps vector-matrix multiplication simply and efficiently. Thanks to programming unit 19, the weights can be updated by reprogramming storage element 156. This approach breaks down the “memory wall” when it integrates arithmetic and memory cells into a single IMC cell. Moreover, by leveraging the analog storage capabilities of the IMC device and Kirchhoff's circuit laws, the use of analog memory devices in the IMC cell allows for efficient execution of MVM operations.

[0102] The IMC unit 10 includes a readout circuit 16 connected to the output of the output line 152. The programming unit 19 can therefore be connected to the readout circuit 16 in its output to adjust the conductance of the memory elements according to, for example, a single-device programming method. Furthermore, the IMC unit 10 may also include a processing unit 18 connected to the output of the IMC unit 10, i.e., connected to the output of the readout circuit 16. This processing unit 18 is preferably arranged near the memory processing unit. In this case, as... Figure 2 As assumed in the text, the programming unit 19 can be advantageously connected to the output of the near-memory processing unit 18 to allow closed-loop programming of the cross array structure 15. In some embodiments, the processing unit 18 and the programming unit 19 are implemented as a single unit. Figure 1A As shown, the programming unit 19 may also include an input / output (I / O) controller and is configured to communicate with external devices or computers 2 and 3. Figure 1A As assumed in the text, the processing unit 18 can be specifically used to apply activation functions and perform classification, for example, binary classification (category #1, category #2).

[0103] Figure 3AThis is an exploded view of the layered structure of the physical reservoir 22, including the ferromagnetic membrane 222. The ferromagnetic membrane 222 is patterned to show the 2D arrangement of perforations. Figure 3B The corresponding top view is shown. Also, as... Figure 3C As seen in the 3D view, electrical conductors 221, 223 are further patterned on each side of the ferromagnetic film 222 in the form of current-carrying lines used in standard MRAM devices. The layers are stacked on a substrate 224 (e.g., copper).

[0104] Figure 4A – Figure 4E This is a 2D cross-sectional view of a portion of the ferromagnetic film 222a-222e in the physical reservoir, schematically illustrating how the ferromagnetic film can be locally processed to form point deformations 225a-225e. In this case, the dimensions of the point deformations 225a-225e are determined to serve as pinning sites for the magnetic domains of the ferromagnetic film 222. Figure 4A – Figure 4E In the example, point deformations 225a–225d form perforations, while Figure 4E In the example, they form protrusions of 225e. The quantity d represents the average diameter of the point deformation, while D represents the first nearest neighbor distance.

[0105] Figure 5A It is a permalloy Fe fabricated on a Si(001) substrate using FIB tools. 55 Ni 45 A diagram of the secondary electrons in the anti-dot lattice. Figure 5B The corresponding in-plane magnetization distribution is shown, with the magnetization direction indicated by arrows. For description, a threshold function is applied. Figure 5A Topographic map. Figure 5A The black dots in the image correspond to the inverted points (i.e., the missing parts of the permalloy in the 10 nm thick permalloy layer), with an actual diameter of about 100 nm and a distance of 2 μm between the inverted points. Figure 5B This illustrates how the magnetic domains are pinned at the inversion point. Note that in this example, ten cycles of a rotating magnetic field were applied to the membrane before image acquisition. (As shown from...) Figure 5A The black dot corresponds to the position of the inverted dot. The image size is 7.3μm × 7.3μm.

[0106] Figure 6 This indicates a ferromagnetic film (Fe) with a thickness of 10 nm. 55 Ni 45The graph illustrates the anisotropic magnetoresistance (AMR) of the magnetic field as a function of the applied magnetic field. This graph shows the total magnetic response for various lattice constants (1, 2, and 3 μm) and cyclic magnetic fields with dot diameters of 100 / 110 μm. For clarity, not all results are depicted. In actual experiments, the anti-dot lattice constant (i.e., the distance between the holes “drilled” into the magnetic film) varied from 0.5 μm to 3.0 μm, and the anti-dot diameter varied from 70 nm to 380 nm. The field that minimizes the total magnetic response to the cyclic magnetic field (the “falling” field) can be considered as the associated quality factor. At and near this field, complex emergent states are formed for reservoir calculation purposes; the lower the field, the lower the energy required to set this state. The inventors observed that for a constant anti-dot diameter, the falling field decreases for larger lattice constants but stabilizes between 2 μm and 3 μm.

[0107] Figure 7 Schematic representation similar to Figure 3B A top view of the ferromagnetic film 222, except Figure 7 The membrane comprises three distinct perforated lattices in different regions. These three lattices have different lattice constants.

[0108] Figure 8 and Figure 9 The diagram illustrates operations such as Figure 1A and Figure 1B A flowchart of the advanced steps of the method of the reservoir computing device 20 shown. See also Figure 8 The method involves setting the magnetic state of the ferromagnetic film 222 by first saturating the ferromagnetic film 222 at step S1 and then coupling an input signal to the physical reservoir 22 at step S2, which results in the generation of an in-plane magnetic field. Step S2 is repeated as needed to repeatedly couple various input signals to the physical reservoir 22. Figure 9 The complete sequence of operations is shown; the ferromagnetic film 222 is assumed to be saturated. In step S10, the input vector is forward-input to the drive system 21, which converts the vector into a time series and encodes the time series into a time-dependent electrical input signal. Then, in step S20, this signal is applied to the physical reservoir 22, i.e., coupled to the physical reservoir 22. In step S30, the physical reservoir 22 generates a response signal, which is read out by the readout unit 23 in step S40. In this example, the readout unit 23 converts the readout signal into a digital value and then feeds the digital value to conventional computing devices 2 and 3 so that in step S50, computing devices 2 and 3 run a neural network (e.g., output layer only) on these values ​​and obtain inference results. In some embodiments, such as Figure 2As shown and as described above, the readout unit 23 converts the readout signal into an input signal, which is fed to the IMC unit 10.

[0109] Figure 10 It is a diagram of manufacturing as follows Figure 3A A flowchart of the advanced steps of the method for constructing the ferromagnetic film 222 in the physical reservoir 22 is shown. A substrate is provided at step S110. Bit lines 223 are patterned at step S120, and then the ferromagnetic film 222 is deposited at step S130. Then, at step S140, the ferromagnetic film 222 is patterned using nanoimprint lithography to obtain dot deformations 225. Then, in step S150, word lines are patterned. The physical reservoir 22 is finally assembled in the device 20. In practical applications, the substrate typically comprises silicon, with some thin additional layers on top of the silicon layer.

[0110] According to one embodiment of the present invention, a physical reservoir for a magnetic reservoir computing device is provided. The physical reservoir includes a ferromagnetic film with a two-dimensional arrangement of point deformations, the dimensions of which are determined to serve as pinning sites for magnetic domains of the ferromagnetic film.

[0111] In this embodiment, the point deformation of the physical reservoir is perforation.

[0112] In the embodiment, the average first nearest neighbor distance between the perforations is between 0.2 μm and 5.0 μm, wherein any first nearest neighbor distance between the perforations is measured parallel to the main surface of the ferromagnetic film, and the average diameter of the perforations is between 50 nm and 500 nm, wherein any diameter of the perforations is measured parallel to the main surface of the ferromagnetic film, and further wherein the average diameter of the perforations is less than the average first nearest neighbor distance.

[0113] In the embodiments, the average first nearest neighbor distance between perforations is between 0.5 μm and 3.0 μm, and the average diameter of the perforations is between 70 nm and 380 nm.

[0114] In the embodiments, the thickness of the ferromagnetic film is between 2 nm and 50 nm.

[0115] In the embodiments, the thickness of the ferromagnetic film is between 5 nm and 20 nm.

[0116] In one embodiment, the two-dimensional arrangement forms at least one dot matrix with dot deformation.

[0117] In an embodiment, at least one dot matrix includes an inverted dot matrix.

[0118] In an embodiment, at least one dot matrix comprises a square dot matrix.

[0119] In one embodiment, at least two lattices with point deformation are formed in different regions of the ferromagnetic film by a two-dimensional arrangement, wherein the at least two lattices have different lattice parameters.

[0120] In an embodiment, the ferromagnetic film comprises one or more elements selected from the group consisting of transition metal elements.

[0121] In an embodiment, the ferromagnetic film comprises Fe x Ni 100-x , where 20≤x≤60.

[0122] In the embodiments, the length of each side of the ferromagnetic film is between 20 μm and 200 μm.

[0123] In one embodiment, the physical reservoir also includes a substrate supporting the ferromagnetic film.

[0124] According to another embodiment of the present invention, a magnetic reservoir device is provided. The magnetic reservoir computing device includes a physical reservoir comprising a ferromagnetic film arranged in a two-dimensional pattern with point deformations, the dimensions of which are determined to serve as pinning sites for magnetic domains of the ferromagnetic film. The magnetic reservoir computing device further includes a drive system configured to saturate the ferromagnetic film and couple an input signal into the ferromagnetic film to set the magnetic state of the ferromagnetic film. The magnetic reservoir computing device also includes a readout unit operatively connected to the physical reservoir to read an output signal from the ferromagnetic film.

[0125] In one embodiment, the drive system includes two or more coils configured to apply a rotating magnetic field coplanar with the ferromagnetic film, wherein the drive system is adapted to couple an input signal into a physical reservoir.

[0126] In one embodiment, the drive system is operatively connected to the ferromagnetic membrane to apply one or more current signals to the ferromagnetic membrane.

[0127] In one embodiment, the physical reservoir includes an arrangement of electrical conductors that connects a drive system to multiple locations within the ferromagnetic membrane. The drive system is configured to apply current signals to the ferromagnetic membrane via the electrical conductors to locally generate a magnetic field coplanar with the ferromagnetic membrane.

[0128] In one embodiment, the drive system is operatively connected to move domain walls in the ferromagnetic film by applying a current signal to the ferromagnetic film according to one of the spin-transfer torque mechanism and the spin-orbit torque mechanism.

[0129] In one embodiment, the magnetic reservoir computing device further includes a processing unit connected to the readout unit and configured to further process signals from the physical reservoir.

[0130] According to another embodiment of the present invention, a method for operating a reservoir computing device is provided. The method includes setting the magnetic state of a physical reservoir, wherein the physical reservoir includes a ferromagnetic membrane arranged in a two-dimensional pattern with point deformations, the dimensions of which are determined to serve as pinning sites for magnetic domains of the ferromagnetic membrane, and wherein the magnetic state is set by saturating the ferromagnetic membrane and coupling an input signal into the physical reservoir to generate one or more magnetic fields coplanar with the ferromagnetic membrane.

[0131] In an embodiment, the method further includes reading an output signal from a physical storage pool and processing the output signal to obtain one or more inference results.

[0132] According to another embodiment of the present invention, a method for manufacturing a reservoir computing device is provided. The method includes patterning a ferromagnetic film to obtain a two-dimensional arrangement of point deformations in the ferromagnetic film, wherein the dimensions of the point deformations are determined as pinning sites for magnetic domains of the ferromagnetic film, thereby obtaining a physical reservoir comprising the ferromagnetic film. The method further includes connecting the physical reservoir to one or more additional components of the reservoir computing device.

[0133] In this embodiment, the ferromagnetic film is patterned using nanoimprint lithography.

[0134] In this embodiment, nanotemplate technology is used to pattern the ferromagnetic film.

[0135] Although the invention has been described with reference to a limited number of embodiments, variations, and drawings, those skilled in the art will understand that various changes can be made and equivalents can be substituted without departing from the scope of the invention. In particular, features (of similar devices or methods) described in a given embodiment, variation, or shown in the drawings may be combined with or substituted for another feature in another embodiment, variation, or drawing without departing from the scope of the invention. Therefore, various combinations of features described with respect to any of the foregoing embodiments or variations are contemplated, which remain within the scope of the appended claims. Furthermore, many minor modifications can be made to adapt particular situations or materials to the teachings of the invention without departing from the scope of the invention. Therefore, the invention is not intended to be limited to the specific embodiments disclosed, but rather the invention will include all embodiments falling within the scope of the appended claims. In addition, many other variations besides those explicitly mentioned above are contemplated.

Claims

1. A physical reservoir for a magnetic reservoir computing device, wherein the physical reservoir comprises a ferromagnetic membrane with a two-dimensional arrangement of point deformations, the size of which is determined as a pinning point for magnetic domains of the ferromagnetic membrane.

2. The physical storage tank according to claim 1, wherein the point deformation is perforation.

3. The physical storage tank according to claim 2, wherein: The average first nearest neighbor distance between the perforations is between 0.2 μm and 5.0 μm, wherein any first nearest neighbor distance between the perforations is measured parallel to the main surface of the ferromagnetic film; and The average diameter of the perforation is between 50 nm and 500 nm, wherein any diameter of the perforation is measured parallel to the main surface of the ferromagnetic film, and further, wherein the average diameter of the perforation is less than the average first nearest neighbor distance.

4. The physical storage tank according to claim 3, wherein: The average first nearest neighbor distance between the perforations is between 0.5 μm and 3.0 μm; and The average diameter of the perforation is between 70 nm and 380 nm.

5. The physical storage tank according to claim 3, wherein the thickness of the ferromagnetic film is between 2 nm and 50 nm.

6. The physical storage tank according to claim 5, wherein the thickness of the ferromagnetic film is between 5 nm and 20 nm.

7. The physical storage pool according to claim 1, wherein the two-dimensional arrangement forms at least one lattice of the point deformation.

8. The physical storage pool according to claim 7, wherein the at least one dot matrix comprises an inverted dot matrix.

9. The physical storage pool according to claim 7, wherein the at least one dot matrix comprises a square dot matrix.

10. The physical reservoir of claim 1, wherein the two-dimensional arrangement forms at least two dot lattices with dot deformation, and wherein the at least two dot lattices with dot deformation have different lattice parameters in different regions of the ferromagnetic film.

11. The physical storage tank according to claim 1, wherein the ferromagnetic film comprises one or more elements selected from the group consisting of transition metal elements.

12. The physical storage tank according to claim 11, wherein the ferromagnetic film comprises Fe x Ni 100-x , where 20≤x≤60.

13. The physical storage tank according to claim 1, wherein the length of each side of the ferromagnetic film is between 20 μm and 200 μm.

14. The physical reservoir of claim 1, wherein the physical reservoir further comprises a substrate supporting the ferromagnetic film.

15. A magnetic reservoir computing device, comprising: A physical reservoir, wherein the physical reservoir comprises a ferromagnetic membrane with a two-dimensional arrangement of point deformations, the size of which is determined to serve as pinning sites for magnetic domains of the ferromagnetic membrane; A drive system configured to saturate the ferromagnetic film and couple an input signal into the ferromagnetic film to set the magnetic state of the ferromagnetic film; as well as A readout unit, operatively connected to the physical reservoir, is provided to read out an output signal from the ferromagnetic membrane.

16. The magnetic reservoir computing device of claim 15, wherein the drive system comprises two or more coils configured to apply a rotating magnetic field coplanar with the ferromagnetic film, and wherein the drive system is adapted to couple an input signal into the physical reservoir.

17. The magnetic reservoir computing device of claim 15, wherein the drive system is operatively connected to the ferromagnetic membrane to apply one or more current signals to the ferromagnetic membrane.

18. The magnetic reservoir computing device according to claim 17, wherein: The physical reservoir includes an arrangement of electrical conductors connecting the drive system to multiple locations within the ferromagnetic membrane; and The drive system is configured to apply a current signal to the ferromagnetic membrane through the electrical conductor to locally generate a magnetic field coplanar with the ferromagnetic membrane.

19. The magnetic reservoir computing device of claim 18, wherein the drive system is operatively connected to apply the current signal to the ferromagnetic membrane to move domain walls in the ferromagnetic membrane according to one of a spin-transfer torque mechanism and a spin-orbit torque mechanism.

20. The magnetic reservoir computing device of claim 15, further comprising a processing unit connected to the readout unit and configured to further process signals from the reservoir.

21. A method for operating a storage pool computing device, the method comprising: The magnetic state of a physical reservoir is set, wherein the physical reservoir comprises a ferromagnetic film with a two-dimensional arrangement of point deformations, the size of which is determined to serve as pinning sites for magnetic domains of the ferromagnetic film, and wherein the magnetic state is set by the following: Saturate the ferromagnetic film; as well as The input signal is coupled into the physical reservoir to generate one or more magnetic fields that are coplanar with the ferromagnetic film.

22. The method of claim 21, further comprising: Read the output signal from the physical storage pool; and The output signal is processed to obtain one or more inference results.

23. A method for manufacturing a reservoir computing device, the method comprising: A patterned ferromagnetic film is obtained to achieve a two-dimensional arrangement of point deformations in the ferromagnetic film, wherein the size of the point deformations is determined as pinning sites for magnetic domains of the ferromagnetic film, to obtain a physical reservoir containing the ferromagnetic film; and Connect the physical reservoir to one or more additional components in the reservoir computing device.

24. The method of claim 23, wherein the ferromagnetic film is patterned by nanoimprint lithography.

25. The method of claim 23, wherein nanotemplate technology is used to pattern the ferromagnetic film.

Images