Tensioning integrated coherent Isin machine
Through tensorization integrated Ising machine, the tensorization optical coupling matrix and multi-wavelength Ising node set is used to solve the scalability limitations of traditional Ising machine when dealing with complex NP difficulties, achieving more efficient, compact and low cost computing power.
Patent Information
- Application Number
- CN202410474381.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-06-02
- Filing Date
- 2024-04-19
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-04-19
AI Technical Summary
When traditional Ising machines deal with complex NP difficult problems due to scalability limitations, it is difficult to improve computing power without increasing space and cost. Especially when dealing with higher complex NP difficult problems, there is hardware occupation. and the issue of increased costs.
The tensorized integrated Ising machine is adopted to reduce hardware and reduce space by utilizing tensorized optical coupling matrix and multi-wavelength Ising node set. The coupling matrix is compressed by using tensor string decomposition algorithm to form a more compact and efficient Ising Xin Machine feedback loop.
It realizes solving complex NP difficult problems with less hardware and less space, reducing manufacturing costs and power consumption, reducing control complexity, and improving system scalability and compactness.
Smart Images

Figure CN119067227A_ABST
Abstract
Description
Background Art
[0001] The Ising model, as a model of ferromagnets, originated from solid-state physics. The Ising model maps a wide class of combinatorial optimization problems for which efficient and accurate classical algorithms do not yet exist. The Ising model can be described by the Hamiltonian (H) as follows, where J ij is the spin σ i With σ j The elements of the coupling matrix between .
[0002]
[0003] The Ising model can be used to formulate NP-hard combinatorial optimization problems (i.e., problems such as the "traveling salesman" problem where the number of possible solutions grows exponentially with the number of components in the system) with only polynomial overhead. Therefore, a machine / apparatus that can solve the Ising model efficiently and effectively would be very valuable. BRIEF DESCRIPTION OF THE DRAWINGS
[0004] The present disclosure is described in detail according to one or more different examples with reference to the following drawings. The drawings are provided for illustrative purposes only and depict only examples.
[0005] Figure 1 Depicted are example integrated coherent Ising machines (CIMs) according to various examples of the presently disclosed technology.
[0006] Figure 2 Depicted are example bistable Ising nodes according to various examples of the presently disclosed technology.
[0007] FIG. 3A to FIG. 3C Depicted are example tensorized integrated CIMs according to various examples of the presently disclosed technology.
[0008] FIG. 4A to FIG. 4D Another example tensorized integrated CIM according to various examples of the presently disclosed technology is depicted.
[0009] Figure 5 Depicted are example multi-wavelength Ising node sets according to various examples of the presently disclosed technology.
[0010] FIG. 6A to FIG. 6D Depicted are example tensorized optical coupling matrices according to various examples of the presently disclosed technology.
[0011] Figure 7 Depicted are example Mach-Zehnder interferometer (MZI) grids according to various examples of the presently disclosed technology.
[0012] Figure 8Depicted are example MZI grid elements according to various examples of the presently disclosed technology.
[0013] Fig. 9 Depicted are examples of weight matrix TT decomposition for parameter compression according to various examples of the presently disclosed technology.
[0014] The drawings are not exhaustive and do not limit the disclosure to the precise forms disclosed. DETAILED DESCRIPTION
[0015] Ising machines are devices built / designed for finding the absolute or approximate ground state of the Ising model. Since NP-hard combinatorial optimization problems (referred to herein as NP-hard problems) can be reformulated using the Ising model, these Ising machines can be used to solve NP-hard problems (i.e., find approximate solutions to NP-hard problems).
[0016] Conventional Ising machines typically use quantum annealing to find the absolute or approximate ground state of the Ising model based on quantum fluctuations. These conventional Ising machines may be difficult to scale due to the limitations of the Ising node interconnection characteristics of quantum annealing-based systems. Due to the limitations of scalability, these conventional Ising machines may have difficulty solving NP-hard problems of increasing complexity (i.e., higher) without incurring excessively high footprint and increased cost.
[0017] Recently, coherent Ising machines (CIMs) have been introduced, which find the absolute or approximate ground state of the Ising model by processing coherent optical signals (e.g., optical signals with the same / similar frequency and waveform). CIMs have greater scalability potential than quantum annealing-based Ising machines because CIMs do not have the Ising node interconnection limitation characteristic of quantum annealing-based systems.
[0018] Among various CIMs, integrated CIM (i.e., integrated photonic circuits implementing CIM) has attracted much attention due to its high energy efficiency, compact footprint, high speed, and low manufacturing cost. For example, PCT / US2015 / 048952 (which is incorporated herein by reference in its entirety) proposes a CIM architecture, which includes: (1) a bistable Ising node implemented using a tunable optical resonator / amplifier with self-feedback; (2) a coupling matrix; and (3) a waveguide connecting the bistable Ising node and the coupling matrix to form an Ising machine feedback loop.
[0019] While integrated CIMs are generally more scalable than conventional / quantum annealing-based Ising machines, it is still highly desirable to further improve the scalability of integrated CIMs.
[0020] In this context, examples of the present technology provide a "tensorized" integrated CIM that improves scalability by utilizing a tensorized optical coupling matrix. The tensorized optical coupling matrix can include layers of multi-wavelength photonic tensor train (TT) cores cascaded together via passive optical cross-connections. The multi-wavelength photonic TT core can include a grid of Mach-Zehnder interferometers (MZIs) (i.e., a lattice / array of interconnected MZIs) that modulate the phase and / or amplitude of optical signals at multiple wavelengths.
[0021] The tensorized optical coupling matrix can implement a tensor string (TT) decomposition algorithm that efficiently compresses the over-parameterized coupling matrix (especially the low-rank sparse coupling matrix) used to solve the Ising model. In addition, by cascading multi-wavelength photonic TT cores via passive optical cross-connections, further reductions in hardware (e.g., fewer MZIs) and further reductions in occupied space can be achieved. Therefore, the tensorized integrated CIM of the present technology can be expanded to solve complex NP-hard problems with less hardware and a smaller occupied space compared to existing integrated CIMs. Relatedly, compared to existing integrated CIMs, the manufacturing cost of the tensorized integrated CIM can be lower, the power consumed can be less, and the control complexity required can be less.
[0022] As described in more detail below, the tensorized integrated CIM can achieve further scalability optimization through the intelligent / specialized design of bistable Ising nodes. For example, the bistable Ising nodes of the tensorized integrated CIM can be implemented via one or more multi-wavelength Ising node sets. The multi-wavelength Ising node set may include multiple bistable Ising nodes, wherein each bistable Ising node in the multi-wavelength Ising node set is associated with a separate optical wavelength. A single multi-wavelength Ising node set to a tensorized optical coupling matrix waveguide can optically connect the output from the multi-wavelength Ising node set to the input multi-wavelength photonic TT core of the tensorized optical coupling matrix. Relatedly, a single tensorized optical coupling matrix to a multi-wavelength Ising node set waveguide can optically connect the output multi-wavelength photonic TT core of the tensorized optical coupling matrix to the input of the multi-wavelength Ising node set. Therefore, the multi-wavelength Ising node set to the tensorized optical coupling matrix waveguide and the tensorized optical coupling matrix to the multi-wavelength Ising node set waveguide can optically connect the multi-wavelength Ising node set and the tensorized optical coupling matrix to form an Ising machine feedback loop.
[0023] The multi-wavelength Ising node set can reduce the hardware and footprint required to implement the tensor integration CIM of the presently disclosed technology. That is, the tensor integration CIM utilizing the multi-wavelength Ising node set architecture can achieve greater scalability than potential alternative tensor integration CIM designs.
[0024] For example, the bistable Ising nodes in the existing integrated CIM are not grouped into sets of multi-wavelength Ising nodes. Instead, each bistable Ising node is associated with its own bistable Ising node to coupling matrix waveguide and its own coupling matrix to bistable Ising node waveguide. Therefore, using such an existing bistable Ising node arrangement to implement a tensorized integrated CIM typically requires a multiplexer to multiplex the optical signals received from the bistable Ising nodes to the waveguides associated with the input multi-wavelength photonic TT core of the tensorized optical coupling matrix. Relatedly, a demultiplexer will be required to demultiplex the signals received from the output multi-wavelength photonic TT core of the tensorized optical coupling matrix to the waveguides going to the bistable Ising nodes. Depending on the size of the tensorized integrated CIM (e.g., the number of multi-wavelength photonic TT cores composed of each layer), the number and / or size of these multiplexers and demultiplexers may also be increased. Furthermore, for a tensorized integrated CIM comprising N bistable Ising nodes, this simple bistable Ising node arrangement will require N bistable Ising node-to-multiplexer waveguides and N demultiplexer-to-bistable Ising node waveguides.
[0025] However, by utilizing a multi-wavelength Ising node set rather than a simple / conventional bistable Ising node arrangement, the tensor-integrated CIM of the presently disclosed technology can: (a) reduce / eliminate the aforementioned multiplexer(s) and demultiplexer(s); (b) reduce the number of waveguides required to connect the bistable Ising nodes and the tensor coupling matrix; and (c) improve the on-chip arrangement / merging of the bistable Ising node hardware. Thus, a tensor-integrated CIM implementing a multi-wavelength Ising node set can achieve greater compactness (i.e., smaller footprint) with less hardware than a potential alternative (and simpler) tensor-integrated CIM design. Through these footprint and hardware reductions, larger (and more powerful systems) can be implemented on smaller chips.
[0026] Examples of the present technology will be described in more detail below in conjunction with the following figures.
[0027] Figure 1 An example integrated CIM 100 is depicted. Integrated CIM 100 may represent a general integrated CIM upon which examples of the presently disclosed technology improve.
[0028] As depicted, the integrated CIM 100 includes: (1) a bistable Ising node 110; (2) a coupling matrix 130; (3) a bistable Ising node to coupling matrix waveguide 120; and a coupling matrix to bistable Ising node waveguide 125. The bistable Ising node to coupling matrix waveguide 120 and the coupling matrix to bistable Ising node waveguide 125 connect the bistable Ising node 110 to the coupling matrix 130 to form an Ising machine feedback loop.
[0029] As depicted, each of the bistable Ising nodes 110 may implement a i To its output v i The nonlinear input-output relationship v i =T(u i ). The function T can be a threshold-like function.
[0030] As depicted, coupling matrix 130 applies analog weights to the outputs of bistable Ising nodes 110. For example, coupling matrix 130 may calculate a weighted sum of the outputs of bistable Ising nodes 110 in the following manner:
[0031] u′ i =∑w ij v j ; Equation 2
[0032] Here, the output v from the bistable Ising node 110 is j Multiply by weight w ij To generate the signal u′ i In a circular feedback loop, the coupling matrix to bistable Ising node waveguide 125 connects the i-th output from the coupling matrix 130 to the input of the i-th bistable Ising node in the bistable Ising node 110 (eg, u i =u′ i ).
[0033] Through the above-mentioned Ising machine feedback loop, the integrated CIM 100 can find the absolute or approximate ground state of the Ising model. Since NP-hard problems can be reformulated using the Ising model, the integrated CIM 100 can be used to solve such NP-hard problems.
[0034] In combination Figure 2 Before describing the hardware implementation of the bistable Ising node 110 in more detail, it may be helpful to understand some of the advantages that the integrated CIM 100 provides over traditional quantum annealing-based Ising machines. For example, the integrated CIM 100 includes full-to-full connectivity between the bistable Ising nodes 110 and the coupling matrix 130. That is, the N outputs from the bistable Ising nodes 110 are connected to the N inputs of the coupling matrix 130 via the bistable Ising node-to-coupling matrix waveguides 120. Relatedly, the N outputs from the coupling matrix 130 are connected to the N input bistable Ising nodes 110 via the coupling matrix-to-bistable Ising node waveguides 125. Due to this full-to-full connectivity, the integrated CIM 100 can outperform traditional quantum annealing-based Ising machines based on the associated higher connectivity and reconfigurability of the integrated CIM 100. For example, the computational power of the Ising machine can be determined by the number of Ising nodes, but can also be determined by the independent non-zero weights w in Equation 2.ij The number of possible connections between Ising nodes is determined by the number of u corresponding to the number of u. As implied above, an integrated CIM such as the integrated CIM 100 processes coherent optical signals. Therefore, u i , v i , w ij , etc. are complex-valued. Although traditional applications may only use real-valued w ij weight, but the integrated CIM 100 can be used to act on u i This facilitates processing of spin variables stored in a particular quadrature of super modes of the anti-correlated internal states of the two resonators 210 (a) and 210 (b).
[0035] Figure 2 An example bistable Ising node 200 is depicted. The bistable Ising node 200 may represent a bistable Ising node that is not implemented in a set of multi-wavelength bistable Ising nodes. The bistable Ising node 200 may be from Figure 1 An example of a bistable Ising node in bistable Ising node 110 .
[0036] As depicted, the bistable Ising node 200 includes two optical resonators (i.e., optical resonators 210(a) and 210(b)) implemented on arms (i.e., first arm waveguide 214 and second arm waveguide 215) of a Mach-Zehnder interferometer (MZI). In some examples, the optical resonators 210(a) and 210(b) may include the same optical ring resonator with dispersive optical nonlinearity.
[0037] The MZI on which the optical resonators 210(a) and 210(b) are implemented includes: (a) a first arm waveguide 214; (b) a second arm waveguide 215; (c) an input-side optical coupler 218; and (d) an output-side optical coupler 219. As depicted, the input-side optical coupler 218 receives a first input from the coupling matrix to the bistable Ising node waveguide 220 and receives a second input from the feedback loop waveguide 216. Then, the input-side optical coupler 218 outputs a first output to the first arm waveguide 214 and outputs a second output to the second arm waveguide 215. The output-side optical coupler 219 receives a first input from the first arm waveguide 214 and receives a second input from the second arm waveguide 215. Then, the output-side optical coupler 219 outputs a first output to the bistable Ising node to the coupling matrix waveguide 222 and outputs a second output to the feedback loop waveguide 216. The input-side optical coupler 218 and the output-side optical coupler 219 may be various types of optical couplers, including a 50-50 splitter, a multi-mode interferometer (MMI), a directional coupler, and the like.
[0038] As depicted, the bistable Ising node 200 also includes a feedback loop waveguide 216 and a bias pump 230. The output of the bias pump 230 can be optically coupled to the feedback loop waveguide 216. As will be described in more detail below, the bias pump 230 and the optical resonators 210(a) and 210(b) can be configured to store two anti-correlated states associated with the bistable Ising node 200. That is, through the feedback loop and the MZI structure, the bistable Ising node 200 can selectively modify the circulating power of the symmetric super modes of the optical resonators 210(a) and 210(b) so that the optical resonators 210(a) and 210(b) exhibit two anti-correlated states.
[0039] In various examples, the bistable Ising node 200 may also include bias field couplers 250 and 252. The bias field couplers 250 and 252 may inject a bias field for a spin variable (eg, one of the four non-relativistic coordinates of an electron) and facilitate monitoring of the current spin amplitude.
[0040] Since the examples of the presently disclosed technology are designed for understanding, the asymmetry of the bistability of a single optical resonator can be avoided by encoding the spin states in two optical resonators (e.g., optical resonators 210(a) and 210(b)), which are configured so that their internal states are anti-correlated. Therefore, when the drive from the bias pump 230 increases the power of the bias field above a threshold level (the phase of the bias field may be weakly correlated with the length of the feedback loop waveguide 216), the bistable Ising node 220 may exhibit a fork-like bifurcation. This may be similar to the bifurcation that occurs in a degenerate optical parametric oscillator (DOPO). In this sense, the bistable Ising node 200 can emulate certain traditional CIMs based on DOPO, but has the added advantage that the input field of the bias pump 230 has the same wavelength as the signal field of the bias pump 230. In addition, the bistable Ising node 200 does not rely on any particular type of optical nonlinearity and can also be adaptive to devices with nonlinear losses.
[0041] As implied above, a strong bias field into a first MZI input (e.g., a first input of the input-side optical coupler 218) can produce a tunable phase-sensitive gain for transmitting a small signal from a second MZI input (e.g., a second input of the input-side optical coupler 218) to an MZI output (e.g., an output from the input-side optical coupler 218). The maximum gain can be determined by detuning the bias (and signal) drive from a common optical resonator resonant frequency. There may be a threshold detuning beyond which the optical resonators 210(a) and 210(b) become bi-stable, respectively, within a certain input power range. In order to implement two states for encoding spins, coherent feedback can be used to couple the resonator modes of the optical resonators 210(a) and 210(b). For example, using an appropriately selected bias feedback phase, the two metastable states can be made unstable so that the optical resonators 210(a) and 210(b) can only exhibit anti-correlated internal states.
[0042] As implied above, the bistable Ising node 200 can be viewed as a tunable amplifier with feedback (TAFB) characterized by a forked bifurcation at a certain threshold bias field power. Above this threshold bias field power, the anti-correlated state described above may exist. The amplitude of the bias pump 230 can be used as a parameter for driving the bias field power to and / or up to the threshold bias field power. The spin variables can be encoded in a specific quadrature as a function of the pump amplitude.
[0043] FIG. 3A to FIG. 3C An example tensorized integrated CIM 300 is depicted in accordance with various examples of the present technique.
[0044] As depicted, the tensorized integrated CIM 300 includes: (a) a multi-wavelength Ising node set (e.g., multi-wavelength Ising node sets 312, 314, and 316); (b) a tensorized optical coupling matrix 330; (c) a multi-wavelength Ising node set to a tensorized optical coupling matrix waveguide 320 that optically connects outputs from the multi-wavelength Ising node set to inputs of the tensorized optical coupling matrix 330; and (d) a tensorized optical coupling matrix to a multi-wavelength Ising node set waveguide 325 that optically connects outputs from the tensorized optical coupling matrix 330 to inputs of the multi-wavelength Ising node set. Thus, the multi-wavelength Ising node set to the tensorized optical coupling matrix waveguide 320 and the tensorized optical coupling matrix to the multi-wavelength Ising node set waveguide 325 optically connect the multi-wavelength Ising node set and the tensorized optical coupling matrix 330 to form an Ising machine feedback loop. As depicted, the tensor integrated CIM 300 may also include a laser 360 (eg, a comb laser) that introduces at most g wavelengths of light into the tensor integrated CIM 300 .
[0045] Will combine Figure 5 The example multi-wavelength Ising node set of the present technology is described in more detail. FIG. 6A to FIG. 6D An example tensorized optical coupling matrix is described in more detail. Before proceeding to these figures, however, it may be useful to describe the composition and operation of the tensorized integrated CIM 300 at a high level.
[0046] As implied above, examples of the present technology provide a “tensorized” integrated CIM (e.g., tensorized integrated CIM 300) that improves scalability by utilizing a tensorized optical coupling matrix (e.g., tensorized optical coupling matrix 330) that includes multi-wavelength photonic tensor train (TT) core layers (e.g., a first layer including multi-wavelength photonic TT cores 332(a) and 332(b), a second layer including multi-wavelength photonic TT cores 334(a) and 334(b), a third layer including multi-wavelength photonic TT cores 336(a) and 336(b), a fourth layer including multi-wavelength photonic TT cores 338(a) and 338(b), etc.) cascaded together via passive optical cross-connects (e.g., passive optical cross-connect 335). A multi-wavelength photonic TT core (e.g., multi-wavelength photonic TT core 332(a)) may include a grid of Mach-Zehnder interferometers (MZIs) (i.e., a lattice / array of interconnected MZIs—see, e.g., Figure 7 for more details).
[0047] The tensorized optical coupling matrix / cascaded multi-wavelength photonic TT core can implement a tensor string (TT) decomposition algorithm that efficiently compresses the over-parameterized coupling matrix (especially the low-rank sparse coupling matrix) used to solve the Ising model. In addition, by cascading the multi-wavelength photonic TT cores via passive cross-connections, further reductions in hardware (e.g., fewer MZIs) and further reductions in occupied space can be achieved. Therefore, the tensorized integrated CIM of the present technology (e.g., the tensorized integrated CIM 300) can be extended to solve complex NP-hard problems with less hardware and a smaller occupied space compared to existing integrated CIMs. Relatedly, the manufacturing cost of the tensorized integrated CIM can be lower, the power consumed can be less, and the control complexity required can be less than that of existing integrated CIMs.
[0048] As implied above, the tensor-integrated CIM can achieve further scalability optimization through intelligent / specialized design of bistable Ising nodes. For example, the bistable Ising nodes of the tensor-integrated CIM can be implemented via one or more multi-wavelength Ising node sets (e.g., multi-wavelength Ising node sets 312, 314, 316, etc.). The multi-wavelength Ising node set (e.g., multi-wavelength Ising node set 312) can include a plurality of bistable Ising nodes, wherein each bistable Ising node in the multi-wavelength Ising node set is associated with a separate optical wavelength (see, e.g., Figure 5 for more details). A single multi-wavelength Ising node set to tensorized optical coupling matrix waveguide (see, for example, the separate multi-wavelength Ising node set to tensorized optical coupling matrix waveguide in multi-wavelength Ising node set to tensorized optical coupling matrix waveguide 320) can optically connect the output from the multi-wavelength Ising node set to the input multi-wavelength photonic TT core (e.g., multi-wavelength photonic TT core 332(a)) of the tensorized optical coupling matrix. Relatedly, a single tensorized optical coupling matrix to multi-wavelength Ising node set waveguide (see, for example, the separate tensorized optical coupling matrix to multi-wavelength Ising node set waveguide in tensorized optical coupling matrix to multi-wavelength Ising node set waveguide 325) can optically connect the output multi-wavelength photonic TT core (e.g., multi-wavelength photonic TT core 338(a)) of the tensorized optical coupling matrix to the input of the multi-wavelength Ising node set. Therefore, the multi-wavelength Ising node set to the tensorized optical coupling matrix waveguide and the tensorized optical coupling matrix to the multi-wavelength Ising node set waveguide can optically connect the multi-wavelength Ising node set and the tensorized optical coupling matrix to form an Ising machine feedback loop.
[0049] As implied above, a multi-wavelength Ising node set can reduce the hardware and footprint required to implement a tensor integrated CIM of the presently disclosed technology. That is, a tensor integrated CIM utilizing a multi-wavelength Ising node set architecture can achieve greater scalability than potential alternative tensor integrated CIM designs.
[0050] To illustrate, FIG. 4A to FIG. 4D An example alternative tensor integration CIM 400 is depicted that does not utilize a multi-wavelength Ising node set.
[0051] As depicted, the tensor integrated CIM 400 includes: (a) bistable Ising nodes (e.g., bistable Ising nodes 411-416), wherein each of the bistable Ising nodes is associated with a separate optical wavelength; (b) a tensor optical coupling matrix 430, wherein the tensor optical coupling matrix includes multi-wavelength photonic tensor train (TT) core layers (e.g., a first layer including multi-wavelength photonic TT cores 432(a) and 432(b), a second layer including multi-wavelength photonic TT cores 434(a) and 434(b), a third layer including multi-wavelength photonic TT cores 436(a) and 436(b), a fourth layer including multi-wavelength photonic TT cores 438(a) and 438(b), etc.) cascaded together via passive optical cross-connects (e.g., passive optical cross-connect 435); c) a multiplexer 470 that multiplexes optical signals received from the bistable Ising node to multiplexer waveguide 440 onto an input waveguide of the tensor coupling matrix 430; (d) a demultiplexer 480 that demultiplexes signals received from the output waveguides of the tensor coupling matrix 430 onto a demultiplexer to bistable Ising node waveguide 445; (e) a bistable Ising node to multiplexer waveguide 440 that optically connects outputs from the bistable Ising nodes (e.g., outputs from the bistable Ising nodes 411-416) to the multiplexer 470; and (f) a demultiplexer to bistable Ising node waveguide 445 that optically connects outputs from the demultiplexer 480 to the bistable Ising nodes. As depicted, each bistable Ising node of the tensorized integrated CIM 400 can be implemented on its own MZI and have its own separate bistable Ising node-to-multiplexer waveguide and demultiplexer-to-bistable Ising node waveguide. Figure 2 The bistable Ising node 200 is implemented in a similar manner.
[0052] As depicted, tensorized integrated CIM 400 includes some features that are the same / similar to tensorized integrated CIM 300 (e.g., tensorized optical coupling matrix 430 including a structure that is the same / similar to tensorized optical coupling matrix 330, laser 460 that introduces at most g wavelengths into the system, etc.), but with some key differences. For example, due to the lack of the multi-wavelength Ising node set architecture of tensorized integrated CIM 300, tensorized integrated CIM 400 includes a greater number of waveguides between its Ising nodes and its tensorized coupling matrix. For example, if tensorized integrated CIM 300 and tensorized integrated CIM 400 both include N Ising nodes, and each multi-wavelength Ising node set of tensorized integrated CIM 300 includes g Ising nodes, then tensorized integrated CIM 400 may include 2g times more waveguides between its Ising nodes and the tensorized optical coupling matrix. Lacking the multi-wavelength Ising node set architecture of the tensor integrated CIM 300 , the tensor integrated CIM 400 further includes a multiplexer 470 (which may include one or more separate multiplexers) and a demultiplexer 480 (which may include one or more separate demultiplexers).
[0053] As implied above, the bistable Ising nodes in existing integrated CIMs are not grouped into sets of multi-wavelength Ising nodes. Instead, each bistable Ising node is associated with its own MZI, its own bistable Ising node to coupling matrix waveguide, and its own coupling matrix to bistable Ising node waveguide (see, e.g. Figure 2 bistable Ising node 200). Therefore, using such an existing bistable Ising node arrangement to implement a tensorized integrated CIM - of which the tensorized integrated CIM 400 is an example - typically requires a multiplexer to multiplex the optical signals received from the bistable Ising node to the waveguides associated with the input multi-wavelength photonic TT cores of the tensorized optical coupling matrix. Relatedly, a demultiplexer will be required to demultiplex the signals received from the output multi-wavelength photonic TT cores of the tensorized optical coupling matrix to the waveguides going to the bistable Ising node. Depending on the size of the tensorized integrated CIM (e.g., the number of multi-wavelength photonic TT cores composed of each layer), the number and / or size of these multiplexers and demultiplexers may also increase. Furthermore (and as implied above), a tensorized integrated CIM implementing this simple / existing bistable Ising node arrangement (e.g., tensorized integrated CIM 400) may require more waveguides between the bistable Ising nodes and the tensorized optical coupling matrix than a tensorized integrated CIM utilizing a set of multi-wavelength Ising nodes (e.g., tensorized integrated CIM 300).
[0054] Thus, by utilizing a multi-wavelength Ising node set rather than a simple / conventional bistable Ising node arrangement, the tensorized integrated CIM of the presently disclosed technology (e.g., the tensorized integrated CIM 300) can: (a) reduce / eliminate the aforementioned multiplexer(s) and demultiplexer(s); (b) reduce the number of waveguides required to connect the bistable Ising nodes and the tensorized coupling matrix; and (c) improve the on-chip arrangement / merging of the bistable Ising node hardware. Thus, a tensorized integrated CIM implementing a multi-wavelength Ising node set can achieve greater compactness (i.e., smaller footprint) with less hardware than a potential alternative (and simpler) tensorized integrated CIM design. Through these footprint and hardware reductions, larger (and more powerful systems) can be implemented on smaller chips.
[0055] Figure 5 An example multi-wavelength Ising node set 500 is depicted in accordance with various examples of the presently disclosed technology. The multi-wavelength Ising node set 500 may be an example of a multi-wavelength Ising node set of the tensorized integrated CIM 300 of FIG.
[0056] As depicted, a multi-wavelength Ising node set 500 includes g bistable Ising nodes implemented on a Mach-Zehnder interferometer (MZI), wherein each bistable Ising node in the multi-wavelength Ising node set is associated with a separate wavelength of light. A given bistable Ising node may include two identical optical resonators configured to modulate light of the associated wavelength of the given bistable Ising node. For example, a first bistable Ising node in the multi-wavelength Ising node set 500 may include optical resonators 510(1)(a) and 510(1)(b). Optical resonators 510(1)(a) and 510(1)(b) may be configured to modulate light of a first wavelength. Together with a bias pump 530 (which may be a multi-wavelength bias pump associated with g wavelengths), optical resonators 510(1)(a) and 510(1)(b) may also be configured to store two bistable / anti-correlated states. Relatedly, the g-th bistable Ising node in the multi-wavelength Ising node set 500 may include optical resonators 510(g)(a) and 510(g)(b). The optical resonators 510(g)(a) and 510(g)(b) may be configured to modulate light at the g-th wavelength. Together with the bias pump 530, the optical resonators 510(g)(a) and 510(g)(b) may also be configured to store two bistable / anti-correlated states.
[0057] As implied above (and as depicted), the g bistable Ising nodes in the multi-wavelength Ising node set 500 are implemented on a common MZI.
[0058] The MZI on which g bistable Ising nodes are implemented includes: (a) a first arm waveguide 514; (b) a second arm waveguide 515; (c) an input-side optical coupler 518; and (d) an output-side optical coupler 519. As depicted, the input-side optical coupler 518 receives a first input from the tensorized optical coupling matrix to multi-wavelength Ising node set waveguide 520 and receives a second input from the feedback loop waveguide 516. Then, the input-side optical coupler 518 outputs a first output to the first arm waveguide 514 and outputs a second output to the second arm waveguide 515. The output-side optical coupler 519 receives a first input from the first arm waveguide 514 and receives a second input from the second arm waveguide 515. Then, the output-side optical coupler 519 outputs a first output to the multi-wavelength Ising node set to the tensorized optical coupling matrix waveguide 522 and outputs a second output to the feedback loop waveguide 516. The input-side optical coupler 518 and the output-side optical coupler 519 may be various types of optical couplers, including a 50-50 splitter, a multi-mode interferometer (MMI), a directional coupler, and the like.
[0059] Through the feedback loop / MZI structure, the multi-wavelength Ising node set 500 can selectively modify the circulating power of the symmetric super mode of the optical resonator implemented on the arm of the MZI so that the optical resonator exhibits two anti-correlated states. The coupler inside the feedback loop waveguide 516 can be used to add the bias signal from the bias pump 530 (as implied above, these bias signals can be associated with g wavelengths).
[0060] In various examples, the multi-wavelength Ising node set 500 may also include a bias field coupler (not depicted) implemented on each of the tensorized optical coupling matrix to multi-wavelength Ising node set waveguide 520 and the multi-wavelength Ising node set to tensorized optical coupling matrix waveguide 522. Such a bias field coupler can inject a bias field for the spin variable and help monitor the current spin amplitude.
[0061] Since the examples of the presently disclosed technology are designed for the sake of understanding, the asymmetry of the bistability of a single optical resonator can be avoided by encoding the spin states in two optical resonators (e.g., optical resonators 510(1)(a) and 510(1)(b)), which are configured so that their internal states are anti-correlated. Therefore, a given bistable Ising node in the multi-wavelength Ising node set 500 can exhibit a fork-like bifurcation as the drive from the multi-wavelength bias pump 530 increases the power of the bias field above a threshold level. This may be similar to the bifurcation that occurs in a degenerate optical parametric oscillator. In this sense, a given bistable Ising node can emulate certain traditional CIMs based on a degenerate optical parametric oscillator (DOPO), but with the added advantage that the input field of the bias pump 530 has the same wavelength as the signal field of the bias pump 530. In addition, a given bistable Ising node does not rely on any particular type of optical nonlinearity and can also be adaptive to devices with nonlinear losses.
[0062] As implied above, a strong bias field entering the first MZI input (e.g., the first input of the input-side optical coupler 518) can produce a tunable phase-sensitive gain for transmitting a small signal from the second MZI input (e.g., the second input of the input-side optical coupler 518) to the second MZI output (e.g., the output from the input-side optical coupler 418). The maximum gain can be determined by detuning the bias (and signal) drive from the common optical resonator resonant frequency. For example, there may be a threshold detuning beyond which the optical resonators 510(1)(a) and 510(1)(b) become bistable over a certain input power range, respectively. To implement two states for encoding spin, coherent feedback can be used to couple the resonator modes of the optical resonators 510(1)(a) and 510(1)(b). For example, using a first appropriately selected bias feedback phase (driven by bias pump 530), the two metastable states can be made unstable so that optical resonators 510(1)(a) and 510(1)(b) can only exhibit anti-correlated internal states. Optical resonators 510(2)(a) and 510(2)(b) can be configured so that a second bias feedback phase (driven by bias pump 530) causes optical resonators 510(2)(a) and 510(2)(b) to exhibit anti-correlated states. Relatedly, optical resonators 510(g)(a) and 510(g)(b) can be configured so that a third bias feedback phase (driven by bias pump 530) causes optical resonators 510(g)(a) and 510(g)(b) to exhibit anti-correlated states, and so on.
[0063] It should be understood here that the multi-wavelength Ising node set 500 is only one example implementation of the multi-wavelength Ising node set. For example, another multi-wavelength Ising node set implementation may include a bus waveguide and multiple micro-ring resonators with internal reflection. Another multi-wavelength Ising node set implementation may include multiple directly coupled dual micro-ring resonators.
[0064] FIG. 6A to FIG. 6D An example tensorized optical coupling matrix 600 is depicted in accordance with various examples of the presently disclosed technology. It should be understood that the tensorized optical coupling matrix 600 merely depicts an example, and other implementations may include differently configured tensorized optical coupling matrices embodying the same / similar principles.
[0065] That is, examples of the present technology provide a "tensorized" integrated CIM that improves scalability by utilizing a tensorized optical coupling matrix (e.g., tensorized optical coupling matrix 600) that includes multi-wavelength photonic tensor train (TT) core layers (e.g., photonic TT core layer 4, photonic TT core layer 3, photonic TT core layer 2, and photonic TT core layer 1) cascaded together via passive optical cross-connections. The multi-wavelength photonic TT cores may include a grid of Mach-Zehnder interferometers (MZIs) (i.e., a lattice / array of interconnected MZIs - see e.g., Figure 7 for more details).
[0066] The tensorized optical coupling matrix / cascaded multi-wavelength photonic TT core can implement a tensor string (TT) decomposition algorithm that efficiently compresses the over-parameterized coupling matrix (especially the low-rank sparse coupling matrix) used to solve the Ising model. In addition, by cascading the multi-wavelength photonic TT cores via passive cross-connections, further reductions in hardware (e.g., fewer MZIs) and further reductions in occupied space can be achieved. Therefore, the tensorized integrated CIM of the present technology can be extended to solve complex NP-hard problems with less hardware and a smaller occupied space compared to existing integrated CIMs. Relatedly, the manufacturing cost of the tensorized integrated CIM can be lower, the power consumed can be less, and the control complexity required can be less than that of existing integrated CIMs.
[0067] Now refer to the TT decomposition algorithm, because the examples of this technique are designed for understanding, the weight matrix W∈R M×N It can be represented in TT format. The matrix dimensions M and N can be assumed to be factorized as and where d is defined as the number of factors of M and N. μ and v can be indices from index (i, j) of W to the second-order-2d weight tensor W [μ1(i), v1(j)..., μd (i), v d (j)] is a natural bijection. So, W(i, j) = W(μ1(i), v1(j)..., μ d (i), v d (j)). The TT decomposition can be interpreted as the singular value decomposition (SVD) of a multidimensional matrix. Fig. 9 As can be seen in , TT decomposition can represent the weight tensor W as a series of tensor products as depicted in Equation 3 below:
[0068]
[0069] Here, the four-way tensor are TT cores, and the total number of tensor cores is d. The vector (R0, R1, ... R d ) is the TT rank, and R0 = R d = 1. In this way, the total number of parameters can be reduced from M×N to the sum of the parameters of the TT core, that is,
[0070] Referring now to the example tensorized optical coupling matrix 600 of FIG. 6 , the tensorized optical coupling matrix 600 can be configured to receive inputs from 1024 bistable Ising nodes associated with 32 separate wavelengths of light. Here, 1024×1024 is factorized into 8×4×4×8×8×4×4×8. As depicted, the total number of photonic TT core layers is four (i.e., d=4), and the total number of wavelengths is 32 (i.e., g=32). Thus, in various examples, the sets of multi-wavelength Ising nodes coupled to the tensorized optical coupling matrix 600 can each include 32 bistable Ising nodes associated with 32 separate wavelengths.
[0071] Here, the TT rank can be set to R0=R2=R4=1 and R1=R3=2.
[0072] Therefore, the tensorized optical coupling matrix 600 includes four photonic TT core layers—photonic TT core layer 4, photonic TT core layer 3, photonic TT core layer 2, and photonic TT core layer 1—whose sizes are G4∈R 1×8×4×2 , G3∈R 2×4×8×1 , G2∈R 1×8×4×2 and G1∈R 2×4×8×1. The passive optical cross-connection between the photonic TT core layer 4 and the photonic TT core layer 3 can implement index switching between the spatial domains of the intermediate signals. The passive optical cross-connection between the photonic TT core layer 2 and the photonic TT core layer 1 can implement similar index switching. As depicted, the passive optical cross-connection between the photonic TT core layer 3 and the photonic TT core layer 2 can implement different functions. In particular, they can include a wavelength-space cross-connection that effectively switches the index between the spatial domain and the wavelength domain of the intermediate signal. The wavelength-space cross-connection between the photonic TT core layer 3 and the photonic TT core layer 2 can be achieved by using a wavelength division multiplexing (WDM) transponder. In some examples, the wavelength-space cross-connection can be implemented using O / E / O conversion and passive electrical cross-connections.
[0073] As depicted, each multi-wavelength photonic TT core of the tensorized optical coupling matrix 600 may include an 8×8 MZI grid. Thus (and as depicted), the tensorized optical coupling matrix 600 includes sixteen 8×8 MZI grids, and passive optical cross-connections between the multi-wavelength photonic TT cores (i.e., 8×8 MZI grids) of different photonic TT core layers, resulting in a total of sixteen 8×8 MZI grids, 448 MZIs, and 32 cascaded MZI levels. An input optical signal received by the tensorized optical coupling matrix 600 from a bistable Ising node may first be modulated using thirty-two 32-wavelength wavelength division multiplexing (WDM) micro-ring modulator arrays. The signal may then be multiplied by the multi-wavelength photonic TT core of each photonic TT core layer before the tensorized optical coupling matrix 600 outputs the optical signal back to the bistable Ising node.
[0074] As depicted, a 32-wavelength comb laser can provide the light source for the system's 32-wavelength optical signals.
[0075] Figure 7 An example Mach-Zehnder interferometer (MZI) grid 700 is depicted in accordance with various examples of the presently disclosed technology.
[0076] As implied above, the multi-wavelength photonic TT core of the presently disclosed technology may include a MZI grid (ie, a lattice / array of interconnected MZIs). Thus, the MZI grid 700 may represent the multi-wavelength photonic TT core of the presently disclosed technology.
[0077] As depicted, the MZI grid 700 may implement an N×N identity matrix represented by a “rectangular” MZI grid having 2×2 MZIs as building blocks. That is, each constituent element (e.g., U 1,1 ) can include 2×2MZI.
[0078] Figure 8An example MZI grid element 800 according to various examples of the presently disclosed technology is depicted. The MZI grid element 800 may include constituent MZI grid elements 1 and 2 of the MZI grid 700. i,j one.
[0079] As depicted, the MZI grid element 800 may include a 2×2 MZI. The 2×2 MZI may include: (a) a first optical coupler 810 and a second optical coupler 812; and (b) a first phase shifter 820 and a second phase shifter 822. The optical couplers 810 and 812 may be various types of optical couplers, including a 50-50 splitter, a multi-mode interferometer (MMI), a directional coupler, and the like.
[0080] As implied above, Fig. 9 Depicted are examples of weight matrix TT decomposition for parameter compression according to various examples of the presently disclosed technology.
[0081] As used herein, the term "optically connected" (and variations thereof "operatively connected," etc.) may refer to a direct or indirect connection between two components that allows optical signals to pass from one component to another.
[0082] As used herein, the term "or" may be interpreted in an inclusive or exclusive sense. In addition, descriptions of resources, operations, or structures in the singular should not be understood to exclude the plural. Unless explicitly stated otherwise, or otherwise understood within the context as used, conditional language (such as, among others, "can," "could," "might," or "may") is generally intended to convey that certain embodiments include and other embodiments do not include certain features, elements, and / or steps.
[0083] Unless otherwise expressly stated, the terms and phrases used in this document and their variations should be interpreted as open-ended rather than restrictive. Adjectives such as "conventional", "traditional", "normal", "standard", "known" and terms of similar meaning should not be interpreted as limiting the items described to items available for a given time period or before a given time, but should be understood to cover traditional, conventional, normal or standard technologies that may be available or known at any time now or in the future. In some cases, the presence of broad words and phrases (such as "one or more", "at least", "but not limited to" or other similar phrases) should not be understood to mean that narrower situations are intended or required where such broad phrases may not be present.
Claims
1. A device, include: a multi-wavelength Ising node set including bistable Ising nodes associated with individual optical wavelengths; Tensorized optical coupling matrix; a multi-wavelength Ising node set to a tensorized optical coupling matrix waveguide optically connecting outputs from the multi-wavelength Ising node set to inputs of the tensorized optical coupling matrix; as well as A tensorized optical coupling matrix is provided to a multi-wavelength Ising node set waveguide, which optically connects outputs from the tensorized optical coupling matrix to inputs of the multi-wavelength Ising node set.
2. The device according to claim 1, in, The multi-wavelength Ising node set to the tensorized optical coupling matrix waveguide and the tensorized optical coupling matrix to the multi-wavelength Ising node set waveguide optically connect the multi-wavelength Ising node set and the tensorized optical coupling matrix to form an Ising machine feedback loop.
3. The device according to claim 1, in, The bistable Ising node is implemented on a Mach-Zehnder interferometer (MZI).
4. The device as claimed in claim 3, in, The MZI of the multi-wavelength Ising node set includes: First arm waveguide; The second arm waveguide; Input side optical coupler that performs the following operations: receiving a first input from the tensorized optical coupling matrix to a multi-wavelength Ising node set waveguide and receiving a second input from a feedback waveguide, and outputting a first output to the first arm waveguide and outputting a second output to the second arm waveguide; and an output-side optical coupler that performs the following operations: receiving a first input from the first arm waveguide and receiving a second input from the second arm waveguide, and Outputting a first output to the multi-wavelength Ising node set to a tensorized optical coupling matrix waveguide, and outputting a second output to the feedback loop waveguide; The bistable Ising node in the multi-wavelength Ising node set is implemented between the input-side optical coupler and the output-side optical coupler on the MZI.
5. The device as claimed in claim 4, in: The first bistable Ising node in the multi-wavelength Ising node set comprises: a first optical resonator optically coupled to the first arm waveguide between the input-side optical coupler and the output-side optical coupler, and a second optical resonator optically coupled to the second arm waveguide between the input-side optical coupler and the output-side optical coupler; and The second bistable Ising node in the multi-wavelength Ising node set includes: a third optical resonator coupled to the first arm waveguide between the input-side optical coupler and the output-side optical coupler, and A fourth optical resonator is coupled to the second arm waveguide between the input-side optical coupler and the output-side optical coupler.
6. The device as claimed in claim 5, in: The first optical resonator and the second optical resonator are configured to modulate light of an associated wavelength of the first bistable Ising node; and The third optical resonator and the fourth optical resonator are configured to modulate light of a wavelength associated with the second bistable Ising node.
7. The device of claim 5, further comprising a pump, in: The output of the pump is optically connected to the feedback loop waveguide; the pump, the first optical resonator, and the second optical resonator being configured to store two anti-correlated states associated with the first bistable Ising node; and The pump, the third optical resonator, and the fourth optical resonator are configured to store two anti-correlated states associated with the second bistable Ising node.
8. The device according to claim 1, in, The tensorized optical coupling matrix includes multi-wavelength photonic tensor train (TT) core layers cascaded together by passive optical cross-connections.
9. The device as claimed in claim 8, in, One of the multi-wavelength photonic TT core layers includes one or more multi-wavelength photonic TT cores.
10. The device according to claim 8, in, A multi-wavelength photonic TT core of the one or more multi-wavelength photonic TT cores comprises an MZI grid.
11. A device, include: a set of multi-wavelength Ising nodes comprising bistable Ising nodes implemented on the MZI, wherein the bistable Ising nodes are associated with separate optical wavelengths; Tensorized optical coupling matrix; a multi-wavelength Ising node set to a tensorized optical coupling matrix waveguide optically connecting outputs from the multi-wavelength Ising node set to inputs of the tensorized optical coupling matrix; and a tensorized optical coupling matrix to a multi-wavelength Ising node set waveguide optically connecting outputs from the tensorized optical coupling matrix to inputs of the multi-wavelength Ising node set; The multi-wavelength Ising node set to the tensorized optical coupling matrix waveguide and the tensorized optical coupling matrix to the multi-wavelength Ising node set waveguide optically connect the multi-wavelength Ising node set and the tensorized optical coupling matrix to form an Ising machine feedback loop.
12. The device according to claim 11, in, The MZI of the multi-wavelength Ising node set includes: First arm waveguide; The second arm waveguide; Input side optical coupler that performs the following operations: receiving a first input from the tensorized optical coupling matrix to a multi-wavelength Ising node set waveguide and receiving a second input from a feedback waveguide, and outputting a first output to the first arm waveguide and outputting a second output to the second arm waveguide; and an output-side optical coupler that performs the following operations: receiving a first input from the first arm waveguide and receiving a second input from the second arm waveguide, and Outputting a first output to the multi-wavelength Ising node set to a tensorized optical coupling matrix waveguide, and outputting a second output to the feedback loop waveguide; The bistable Ising node in the multi-wavelength Ising node set is implemented between the input-side optical coupler and the output-side optical coupler on the MZI.
13. The device according to claim 12, in: The first bistable Ising node in the multi-wavelength Ising node set comprises: a first optical resonator optically coupled to the first arm waveguide between the input-side optical coupler and the output-side optical coupler, and a second optical resonator optically coupled to the second arm waveguide between the input-side optical coupler and the output-side optical coupler; and The second bistable Ising node in the multi-wavelength Ising node set includes: a third optical resonator coupled to the first arm waveguide between the input-side optical coupler and the output-side optical coupler, and A fourth optical resonator is coupled to the second arm waveguide between the input-side optical coupler and the output-side optical coupler.
14. The device according to claim 13, in: The first optical resonator and the second optical resonator are configured to modulate light of an associated wavelength of the first bistable Ising node; and The third optical resonator and the fourth optical resonator are configured to modulate light of a wavelength associated with the second bistable Ising node.
15. The device of claim 13, further comprising a pump, in: The output of the pump is optically connected to the feedback loop waveguide; the pump, the first optical resonator, and the second optical resonator being configured to store two anti-correlated states associated with the first bistable Ising node; and The pump, the third optical resonator, and the fourth optical resonator are configured to store two anti-correlated states associated with the second bistable Ising node.
16. The device according to claim 11, in, The tensorized optical coupling matrix comprises multi-wavelength photonic TT core layers cascaded together by passive optical cross-connections.
17. The device according to claim 16, in, One of the multi-wavelength photonic TT core layers includes one or more multi-wavelength photonic TT cores.
18. The device according to claim 16, in, A multi-wavelength photonic TT core of the one or more multi-wavelength photonic TT cores comprises an MZI grid.
19. A device, include: bistable Ising nodes, which are associated with individual wavelengths of light; Tensorized optical coupling matrix; a multiplexer that multiplexes optical signals received from the bistable Ising nodes to the multiplexer waveguides onto input waveguides of the tensorized optical coupling matrix; a demultiplexer that demultiplexes signals received from output waveguides of the tensorized optical coupling matrix onto a demultiplexer onto a bistable Ising node waveguide; The bistable Ising node to multiplexer waveguide optically connects an output from the bistable Ising node to the multiplexer; and The demultiplexer to bistable Ising node waveguide optically connects the output from the demultiplexer to the bistable Ising node.
20. The device according to claim 18, in, The bistable Ising node to multiplexer waveguide and the demultiplexer to bistable Ising node waveguide connect the bistable Ising nodes and the tensorized optical coupling matrix to form an Ising machine feedback loop.
Citation Information
Patent Citations
Optical ising machines and optical convolutional neural networks
CN111095303A
Optical Isin model calculation method and system based on wavelength division multiplexing
CN116184739A
Apparatus for solving ising problems
US20180267937A1
Hybrid analog-digital matrix processors
US20200272795A1
Photonic tensor core matrix vector multiplier
US20230152667A1
