Deterministic computing device based on physical mapping structure and causal inference method

Abstract

This invention addresses the problems of weight-dependent charge storage and computational latency constrained by the von Neumann architecture in existing deterministic computing devices for neural networks. It provides a physical mapping-based device and method for one-time deterministic computation using a physical mapping structure (Si / SiO2 multilayer dielectric, thickness 100-500nm, refractive index accuracy ±0.001), belonging to the field of physical computing technology. This invention replaces the storage of neural network weights with a physical mapping structure whose geometric parameters and / or material physical parameters are permanently fixed during manufacturing. It is suitable for real-time control scenarios requiring deterministic, low-latency reasoning, such as high-frequency trading, autonomous driving, and edge reasoning, involving linear equations, partial differential equations, and logical deductions. Latency is compressed to the order of physical propagation time (0.5-0.8ns).

Description

Technical Field

[0001] The technical solution of this invention focuses on solving operators for the solidification of passive physical medium parameters and performing deterministic calculations, without involving artificial intelligence algorithm features.

[0002] The algorithmic features of this invention are specifically technically related to the parameter distribution structure of the physical medium.

[0003] This invention relates to the field of physical computing technology, specifically to a deterministic computing device and method that does not require a programmable processor and directly maps output from a physical structure. Its international patent classification numbers are G06E1 / 00 (optical computing device), G02F3 / 00 (optical computing), and G06J1 / 00 (hybrid computing device). When the embodiment uses a surface acoustic wave physical medium, it may also be labeled G06G1 / 00 (analog computer) or H03H9 / 00 (network for electromechanical resonators).

[0004] When the embodiment uses a metasurface unit, it can also be labeled G02B1 / 00 (optical element).

[0005] All computational operations in this invention are performed at the physical level by wave field propagation, reflection, or diffraction processes within a passive physical medium, without relying on any software instruction stream or programmable processor execution. The physical medium and its processing technology are existing technologies. The innovation of this invention lies in permanently embedding the problem-solving operator into the medium in the form of a physical parameter distribution, forming a completely new computational architecture.

[0006] Preferably, the reverse design algorithm employs the gradient descent method. Preferably, the medium material of the diffractive optical element is glass or polymer to ensure nanosecond-level light field propagation delay. Background Technology

[0007] Furthermore, the 'deterministic computing device based on physical mapping structure' described in claim 1 of this invention has substantial differences from existing technologies (including optical neural networks, SAW filters, DOE optical slide rules, etc.): (1) Different problem domains: machine learning classification vs. deterministic computing; (2) Different physical mapping structures: trained diffraction units vs. reflection grating / phase plate directly solidified operators; (3) Different output carrying capacity: light intensity classification vs. complex amplitude distribution of light field; (4) Different post-processing requirements: CPU processing vs. direct output.

[0008] Furthermore, regarding the key term "surface acoustic wave computation," existing technologies mainly focus on SAW filters, SAW resonators, and SAW sensors (such as commercial products from manufacturers like Murata and TDK), used in the fields of radio frequency signal processing, sensors, and microfluidics. There are no public reports on the "solidification of SAW physical mapping structures for deterministic computation tasks."

[0009] Regarding the keyword (3) solving partial differential equations for diffractive optical elements, the AMiner search results show that existing technologies mainly focus on optical neural networks (for machine learning classification) and optical slide rules (for simulation calculations), and there are no public reports on 'DOE physical mapping structure directly solidified PDE solver'.

[0010] (a)+(e) Synergistic Argument: The core of this invention's inventiveness lies in the synergistic combination of (a) the permanent solidification of the solution operator in the physical medium parameters and (e) the instructionless passive execution of the physical medium in the solution process, rather than any single feature. A single (a) feature (such as programmable in-memory computation) still requires digital processor scheduling; a single (e) feature (such as a D²NN static diffraction network) is still used for classification tasks, not for solving deterministic operators. This invention's (a)+(e) synergy defines "non-volatile solidification of the solution operator" and "instructionless passive execution of the solution process" as an inseparable combination, realizing a paradigm shift from "algorithm acceleration" to "operator carrier".

[0011] c) Existing D²NN technology has approximately 2 million neurons (200×200=40,000 neurons per layer × 5 layers). The number of parameters in the physical mapping structure of this invention is proportional to the dimension of the problem to be solved (Nth-order linear equations correspond to N² reflective gratings). The number of parameters is controllable and matches the problem size. d) Existing D²NN technology reads the light intensity distribution through 5 photodetectors and then takes argmax. The output of this invention is directly carried by the complex amplitude distribution of the light field (containing both amplitude and phase information), requiring no post-processing.

[0012] (3) Lin X. et al., 'All-optical machine learning using diffractive deep neural networks', Science, 361(6406):1004-1008, published in September 2018 (DOI: 10.1126 / science.aat8084).

[0013] The essential difference between this invention and the prior art is that: a) the existing planar waveguide integrated DONN is also used for machine learning classification (MNIST handwritten digit recognition), rather than for the physical solidification of deterministic computation tasks; b) the existing technology uses an oblique incidence propagation model (achieved through diffraction and reflection in the waveguide space), while this invention uses a free space normal incidence propagation model (reflection grating array + phase plate + Fourier lens + detector surface), and the two have different optical structures;

[0014] c) Existing technologies discretize neuron phase values ​​into 4 levels (0, π / 2, π, 3π / 2) to simplify fabrication. The physical mapping structure parameters of this invention are continuously adjustable (reflection grating position accuracy 10 nm, phase plate phase accuracy 0.01 rad) to meet the stringent accuracy requirements of deterministic computing tasks; d) Existing technologies process 50 types of amplitude-coded handwritten digit images (32×32 pixels), while this invention processes 176-bit structured data packets or PDE coefficient vectors containing node identifiers and connection relationships.

[0015] (2) Feng JN et al., 'Compact planar-waveguide integrated diffractive optical neural network chip', Advanced Photonics Nexus, published in 2024.

[0016] This document discloses a planar waveguide integrated diffraction neural network chip, in which three diffraction layers are integrated on the same quartz wafer surface, with a processing space of 32 mm³ and a computing speed of 3.1 × 10⁻⁶. 9 TOPS, 250,000 neurons (500×500) achieved 73.4% experimental accuracy on the MNIST dataset, with 88.6% consistency over 10 cycles.

[0017] The essential difference between this invention and the present invention is that: a) the prior art DONN is used for machine learning classification tasks (Iris / MNIST pattern recognition), while the present invention is used for the physical solidification of deterministic computation tasks (solving linear equations, executing finite state machines, solving PDEs, and deducing logic diagrams), that is, 'signal → output' rather than 'image → category label';

[0018] b) The diffraction unit (DU) weights of the existing DONN are fixed after training through backpropagation, but the training process is 'softly adjustable' and not for deterministic computation operators. The physical mapping structure of the present invention directly solidifies the mapping relationship of 'problem to be solved → operator → output' through the distribution of the reflector grating position and / or the phase distribution of the phase plate, without any training process; c) The output of the existing DONN is sent to the CPU for subsequent processing after reading the light intensity distribution through a photodetector. The output of the present invention is directly carried by the light field distribution itself, without any photoelectric conversion or electronic computing intervention;

[0019] d) The number of neurons and topology of the existing DONN technology are optimized for specific datasets (250,000 neurons for MNIST). The physical mapping structure of the present invention can be arbitrarily adjusted according to the dimension of the problem to be solved (order of linear equations, number of states of finite state machine, dimension of PDE), and is not bound to a specific dataset.

[0020] (1) Fu TZ et al., 'Photonic machine learning with on-chip diffractive optics', Nature Communications, 14:70, published on January 4, 2023 (DOI: 10.1038 / s41467-022-35772-7).

[0021] This paper discloses an on-chip diffractive optical neural network (DONN) based on a silicon-on-insulator (SOI) platform. The operating wavelength is 1.55 μm, each layer contains 186 neurons (558 silicon trench-filled SiO2 units), the single hidden layer is 0.15 mm², and the triple hidden layer is 0.3 mm². It achieves a test accuracy of 86.7% / 90% on the Iris dataset and 96.3% (numerical calculation) / 86.0% (experimental test) accuracy on the MNIST dataset.

[0022] In July 2026, the inventors conducted a systematic prior art search on the AMiner academic search platform (Academic Collaboration Graph, SemanticScholar, arXiv, Nature Communications database) for the 'deterministic computing device and method based on physical mapping structure' described in claim 1 of this invention. The search keywords included:

[0023] (1) Optical Neural Network / Physical Mapping Structure; (2) Surface Acoustic Wave Computing; (3) DOE-based PDE Solver.

[0024] The following are three representative prior art works that are closest to the physical mapping structure of this invention, and these works are substantially different from the present invention:

[0025] Comparison of Existing Technologies and Novelty Counter-Evidence Based on AMiner Academic Search

[0026] Modern computers are all based on the von Neumann architecture, and the computation process must go through steps such as instruction fetching, decoding, execution, and write-back. Even when new computing paradigms such as in-memory computing (e.g., the in-memory computing architecture based on memristor array disclosed in CN112668702A) or optical neural networks (e.g., the optical interferometric neural network accelerator disclosed in US11062205B2) are adopted, they only transfer some matrix operations to the analog domain, and still require the deep involvement of digital processors and control of instruction flow.

[0027] The fundamental limitation of this type of solution is that the physical structure only serves as an accelerator for the algorithm and does not change the essence of "computation is the execution of instructions". Therefore, the latency and power consumption cannot break through the von Neumann bottleneck.

[0028] (3) The essential difference from recursive optics / adaptive optics schemes: Recursive optics solvers (such as adaptive optics iterative convergence systems and neural optical networks) achieve convergence through feedback loops or iterative optimization algorithms. This invention completes the non-iterative simulation domain solution in one go through the parameter space distribution of the physical medium. The solution delay depends only on the propagation time of the physical field, and the solution process does not depend on any feedback loop. Furthermore, the accuracy of recursive optics / adaptive optics is limited by the iterative convergence residual and feedback delay. The accuracy of this invention is uniquely determined by the processing accuracy of the physical medium (sub-nanometer accuracy of photolithography / etching / deposition processes), and there is no iterative residual. In addition, recursive optics schemes require digital processors to implement feedback control. The physical medium of this invention has no instruction flow, no clock, and no instruction fetch-decode-execution process in the solution process. It is a purely passive physical mapping. The core differences between the above three types (operational amplifiers / in-memory computing / recursive optics) and this invention can all be summarized as follows: Existing technologies use physical structures as "accelerators of algorithms" (requiring deep involvement of digital processors). This invention uses physical structures as "carriers of operators" (without any digital processor involvement). This embodiment corresponds to the device described in claim 1.

[0029] Furthermore, in the field of analog computing, it is known to use operational amplifier networks to solve linear equation systems (such as analog solvers based on resistor networks). However, the essential difference between this approach and the present invention lies in:

[0030] (1) Different computational paradigms: The operational amplifier solver relies on the negative feedback loop to achieve iterative convergence, which belongs to the "analog iterative" computational paradigm. The circuit needs time to establish a steady-state equilibrium. The present invention completes a one-time direct mapping through the propagation of the physical field. There is no iterative process, and the solution delay depends only on the propagation time of the physical field.

[0031] (2) Different accuracy mechanisms: The accuracy of the operational amplifier solver is limited by the resistor matching error (typically 1%), and the error accumulates as the circuit size increases; the optical phase accuracy of the present invention can reach λ / 100 (corresponding to subwavelength level phase control);

[0032] Based on an equivalent surface processing error of approximately 5.3 nm corresponding to a working wavelength of 532 nm, referencing typical surface roughness indices for modern photolithography processes (J. van Schoot et al., "EUV lithography: NXE platform performance overview and status", Proc. SPIE 11609, 2021), and considering that each spatial degree of freedom is processed in parallel, the error does not accumulate with the problem size.

[0033] (3) Different power consumption modes: The operational amplifier solver requires external power supply to maintain the working state of the feedback loop, resulting in continuous static power consumption; the physical mapping structure of the present invention (i.e. the solution medium referred to in the claims) is completely in a passive state without external control signals during the calculation process, requiring only light source or pulse excitation, and the dynamic power consumption is close to zero.

[0034] Current known optical simulation computing schemes (such as the superstructure for solving equations disclosed in US2019 / 0114012A1) typically employ recursive transformations, requiring multiple passes through the physical structure. Schemes based on tunable optical elements (such as optical modulators) rely on external control signals, essentially remaining within the command-controlled paradigm. Existing technologies lack a fully passive physical computing structure capable of permanently embedding the solver and completing the calculation in a single, direct mapping manner.

[0035] Furthermore, Chinese patent CN121998010A discloses a photon simulation iterative calculation device (application date November 6, 2024). This device achieves optical solution of matrix equations by evenly splitting the optical signal into multiple paths, performing photon multiplication and addition calculations using modulation voltage, and iteratively solving the variable vector according to an iterative formula containing proportional, original, and differential terms. This scheme belongs to the optical simulation iterative calculation paradigm and still requires continuous external modulation voltage to drive multiple iterations, which is fundamentally different from the one-time direct mapping of this invention.

[0036] In addition, Chinese patent CN110334804A (Tsinghua University) discloses a fixed all-optical diffraction neural network after training, which belongs to machine learning classification tasks. This is fundamentally different from the present invention, which fixes the solution operator during manufacturing and does not depend on the training process.

[0037] The above distinctions differentiate this invention from all known simulation computing schemes at the computational paradigm level.

[0038] For a large number of deterministic computational tasks (such as solving systems of linear equations, executing finite state machines, logic diagram derivation, solving partial differential equations, or logical reasoning), the steps of the algorithm can be pre-defined. However, existing technologies lack a computational structure that can directly physically solidify the solution operators for such problems without requiring runtime programmability.

[0039] A. Optical Neural Networks and Matrix / PDE Solving [Tait et al., 2014] "Broadcast and weight: an integrated network for scalable photonic spike processing", J.Lightwave Technol. 32(15):4029-4041 (Baseline literature on silicon photonic neural network multi-wavelength broadcast architecture);

[0040] [Shen et al., 2017] "Deep learning with coherent nanophotonic circuits", Nat. Photon. 11:441-446 (Foundational work on deep learning photonic circuits);

[0041] [Lin et al., 2018] "All-optical machine learning using diffractivedeep neural networks",

[0042] Sci. Rep. 8:12320 (Diffraction Deep Neural Network, 2018 Highly Cited Paper) — All three papers focus on "trainable / rewriteable" neural network weight systems. They train weights in the software layer through backpropagation algorithm and then load them into the optical path to achieve inference, or train and optimize the geometric parameters of the diffraction plate in the software layer through deep learning algorithm and then solidify them.

[0043] This is fundamentally different from the "passive permanent solidified mathematical solver (non-trained weights)" in this application at the paradigm level.

[0044] B. Passive Analog Optical Computing: A review comparing passive optical computing with memristor cross arrays, a review of analog optical computing architectures, optical recursive transformation architectures, tunable optical element architectures, etc. - The listed literature focuses on the "passive device + electronically controlled weight" or "recursive / tunable" paradigms, and has not achieved the simultaneous hardware implementation of "passive + permanent solidification + one-time direct mapping" in three dimensions.

[0045] C. Deterministic physical mapping computation [CN112668702A] (an in-memory computation architecture based on memristor arrays, storing weights through conductance states; belonging to the category of "rewritable charge / conductance states," which is physically completely different from "permanently fixed geometric structure parameters / material physical parameters" that are explicitly "not stored in programmable memory in the form of charge states, magnetization states, or phase transition states"); [US11062205B2] (an optical interferometric neural network accelerator, loading weights in real time through an electro-optic modulator).

[0046] It belongs to the combination of "optical structure + electronic weights", and essentially still requires command control); [US2019 / 0114012A1] (the superstructure used to solve the equations uses a recursive transformation method that requires multiple passes through the physical structure; unlike the "one-time direct mapping" of this application);

[0047] [J. van Schoot et al., 2021] "EUV lithography: NXE platform performance overview and status", Proc. SPIE 11609 (An overview of extreme ultraviolet lithography, providing evidence of the feasibility of the process for "subwavelength-level surface processing errors" (approximately 5.3 nm @ 532 nm operating wavelength); it is not a computational architecture document itself, but only serves as support for the feasibility of the manufacturing process in this application). Summary of the Invention

[0048] Purpose of the invention

[0049] The technical problem to be solved by this invention is that existing physical computing schemes all rely on the instruction intervention of tunable elements and digital controllers, and the physical structure is not used as a complete solver, resulting in instruction overhead and memory-computation separation delay.

[0050] Technical solution

[0051] Explanation of the superordinate concept: In this invention, "physical medium" is a superordinate concept, encompassing any physical medium that can permanently solidify the solver through geometric structural parameters and / or material physical parameters during the manufacturing process, including but not limited to any one of the following: multilayer diffractive optical elements, surface acoustic wave delay lines, memristor cross arrays, phase change memory cells, and piezoelectric substrate structures; "non-uniform parameter spatial distribution" encompasses any one of the following: refractive index distribution, thickness distribution, geometric structural distribution, and material composition distribution.

[0052] Furthermore, corresponding to the physical medium type described in claim 3, the physical medium of this invention can be at least one of the following: a diffractive optical multilayer structure (multilayer diffractive optical element), a surface acoustic wave delay line array, and a subwavelength metasurface unit array. The diffractive optical multilayer structure is composed of stacked multilayer phase plates, achieving optical calculations through multilayer phase accumulation. The subwavelength metasurface unit array is composed of subwavelength scale resonant units arranged according to a preset phase distribution, enabling phase control equivalent to that of a multilayer diffractive optical element on a single-layer structure. All of the above medium types are specific implementations of the overarching concept of the physical medium in this invention.

[0053] The aforementioned overarching concepts mean that the solution operator solidification module described in claim 1 is not limited to a single physical medium type, and those skilled in the art can select the most suitable physical medium implementation method according to the type of problem to be solved.

[0054] This invention provides a deterministic computing device and method based on direct mapping of physical structures. It embeds the solution operator for a specific problem into the parameter distribution of the physical medium at once. After inputting a problem instance, the physical action outputs the solution instantly. The entire process is without instruction flow, clock cycle, or instruction fetch-decode-execution process.

[0055] The device includes an input encoding module, a physical mapping structure, and an output decoding module.

[0056] It should be noted that the "input encoding module," "physical mapping structure," and "output decoding module" used in this embodiment and accompanying drawings correspond to the "incident physical field excitation module," "solution operator solidification module and physical solution module," and "output demapping module" in claim 1, respectively: wherein, the input encoding module is the incident physical field excitation module as described in claim 1, which is used to encode the problem instance into an incident physical field and apply it to the physical medium; the physical mapping structure is the solution operator solidification module and physical solution module as described in claim 1, which solidifies the solution operator in the non-uniform parameter space distribution of the physical medium and performs analog domain solution operations when the incident physical field passes through; the output decoding module is the output demapping module as described in claim 1, which is used to sense the measurable physical features related to the solution in the output physical field and output the numerical value or symbol representing the solution.

[0057] The input encoding module is used to convert the problem instance to be solved into an incident physical field with problem encoding properties. Depending on the type of physical medium, the incident physical field can be a coherent beam (optical embodiment), surface acoustic waves (acoustic embodiment), or other forms of physical waves.

[0058] The physical mapping structure consists of a medium with non-uniform parameter distributions (such as phase distribution, reflection grating distribution, and refractive index distribution) fixed according to the solver operator, and is in a passive state during computation. The incident physical field, encoded by the problem instance, is directly modulated into the output physical field through this structure.

[0059] It should be particularly noted that the parameter distribution and solidification method of the physical mapping structure of this invention differs from the data storage in traditional computers. Information in traditional memory exists in the form of charge states (DRAM), magnetization states (MRAM), or phase transition states (PCM), and can be rewritten via electrical signals;

[0060] The parameter distribution of this invention exists in the form of geometrical parameters of the physical medium (such as phase delay determined by etching depth, physical position and width of the reflective grating) and / or material physical parameters (such as refractive index distribution determined by material composition or doping distribution). These parameters are permanently formed during manufacturing through micro-nano fabrication processes and cannot be altered by any electrical signals during computation. This "fixation of geometrical structure and material parameters" is fundamentally different from "data storage" in physical terms—the former refers to the shape and compositional properties of the material, while the latter refers to the state properties of the material.

[0061] The output decoding module directly extracts the solution from the measurable features of the output field (spot position, pulse delay, intensity ratio, etc.), without the need for analog-to-digital conversion and digital post-processing by a general-purpose processor.

[0062] This distinguishing feature is not a combination of conventional techniques in the field: existing optical neural networks rely on external training to determine weights, SAW filters are used for radio frequency signal processing rather than computation, and DOE optical slide rules are used for simulation computation. None of these three methods directly solidify computation operators into passive physical structures for deterministic computation tasks, nor do they establish a one-to-one physical mapping between the wave field propagation process and the deterministic solution operators. Those skilled in the art have no motivation to combine the independent techniques of optical neural networks, SAW filters, and DOE optical slide rules into a device that directly outputs deterministic computation results with a physical mapping structure. The prior art has also not provided any technical inspiration for the above combination. Therefore, the technical solution of the present invention is not obvious, but has substantial features and progress.

[0063] It should be noted that in recent years, research has emerged in the direction of solving equations using optical metasurface inverse design and diffraction optics: Cordaro et al. (2023, Nature Nanotechnology) passively solved the Fredholm integral equation in free space using a reverse-designed supergrating. Although the solution process is passive propagation, it involves a single inverse design for a specific integral operator and does not form a general method for fixing the parameters of the solution operator. The structural parameters cannot be flexibly mapped for different equations. DIODES (2025) and other diffraction optics differential equation solvers are based on trainable or operator-specific structures. This invention maps the operators to be solved (algebraic equation systems, partial differential equations, finite state machines, logic diagram derivations, etc.) to the phase distribution of a multi-layer phase plate through angular spectrum propagation and gradient descent, forming a general passive solution architecture that can be directly solved once by manufacturing and solving any similar problems. This is substantially different from the above-mentioned schemes that involve inverse design for a single operator or are trainable and rewriteable in terms of operator fixing method, universality, and computational paradigm.

[0064] Based on the aforementioned distinguishing technical features, the technical problem actually solved by this invention is: how to complete deterministic computation tasks using a one-time direct mapping method with passive physical propagation without the need for software instruction streams and recursive iterations, thus avoiding the accumulation of latency and instruction overhead caused by software computation. By permanently fixing the solver operator to the medium with physical parameter distribution and mapping it out in one go, this invention eliminates the latency of recursive iterations and the instruction control dependency of tunable elements.

[0065] Beneficial effects

[0066] (1) Extremely low delay: Compared with US11062205B2, this invention compresses the computational delay to the pure physical action time—when solving a 1000-dimensional linear equation system, the physical propagation mapping time is about 0.5 nanoseconds (based on an equivalent optical path length of about 10 cm in the propagation medium, corresponding to a physical propagation time of 0.5 ns; this estimate is consistent with the representative value in the range of 30–50 cm for the total optical path length after folding the optical path of a 10-layer phase plate stack), including high-speed InGaAs / InP avalanche photodiodes.

[0067] Furthermore, in the folded optical path of this embodiment, the optical path difference between adjacent phase plate layers is controlled within 15cm to ensure the coherence of the phase superposition of the optical fields of each layer; the total optical path of the folded optical path of 30-50cm and the optical path difference between adjacent layers ≤15cm together define the spatial dimensions and interlayer consistency constraints of the folded optical path of this embodiment.

[0068] The total readout delay of the (APD) is approximately 0.8 nanoseconds (see H. Han et al., "High performance InGaAs / InP avalanche photodiode integrated with metal–insulator–metalmicrocavity", Opt. Quantum Electron. 53:307).

[0069] In 2021, its 3 dB bandwidth reached 5.8 GHz, corresponding to a response time on the order of hundreds of picoseconds;

[0070] The 0.3 ns readout time in this embodiment is within the range achievable for custom-designed devices, while optical neural network chips of the same scale require microsecond-level delays (see AN Tait et al., "Broadcast and weight: an integrated network for scalable photonic spike processing", J. LightwaveTechnol. 32(15):4029–4041, 2014;

[0071] Y. Shen et al., "Deep learning with coherent nanophotonic circuits", Nat. Photon. 11:441–446, 2017. Compared with the electronic operational amplifier solver, this invention does not require iterative convergence time, and the solution delay depends only on the propagation speed of the physical field.

[0072] (2) Extremely low power consumption: Because there is no clock, no instruction fetch, no instruction decoding, and no dynamic power consumption of the instruction cycle of traditional digital circuits, the total power consumption is determined by the static power of the light source, with a typical value of 100mW.

[0073] (3) No post-processing required: Compared with CN112668702A, this solution does not require analog-to-digital conversion and digital post-processing by a general-purpose processor. The output decoding module directly extracts the final solution from the measurable features of the physical field, eliminating the delay and accuracy loss of analog-to-digital conversion.

[0074] (4) Accuracy does not accumulate with scale: Unlike electrical simulation solvers, the optical / acoustic parallel processing mechanism of this invention enables the calculation of each spatial degree of freedom to be completed simultaneously, and the error does not accumulate as the problem scale increases.

[0075] (5) Compared with the recursive transformation superstructure disclosed in US2019 / 0114012A1, the present invention achieves one-time direct mapping through permanently fixed parameter distribution, eliminating the delay accumulation caused by recursive iteration; compared with the scheme based on tunable optical elements, the present invention does not require external control signals during the calculation process, realizing true zero instruction overhead calculation.

[0076] It should be noted that the passive state without external control signals referred to in this specification and claims means that the physical mapping structure (i.e., the solver solidification module and the physical solver module) does not receive external control signals for controlling the solver operation during the solver process, and completes the calculation by passive propagation of the light field; the spatial light modulator used in the input encoding module in the embodiment is only used to encode the known quantities of the problem instance into the incident light field before the solver starts, and does not participate in the control of the solver process, nor does it change the passive characteristics of the physical mapping structure during the solver process. Attached Figure Description

[0077] Figure 1 This is a block diagram illustrating the overall principle of the deterministic computing device of the present invention, showing the signal flow of the input encoding module, the physical mapping structure, and the output decoding module. The diagram shows a block diagram comprising three main modules (input encoding module 1, physical mapping structure 2, and output decoding module 3), with arrows indicating the direction of the physical field / signal flow. Input encoding module 1 encodes the problem instance into an incident physical field, which is then fed into physical mapping structure 2. Physical mapping structure 2 outputs an output physical field carrying solution information to output decoding module 3.

[0078] Figure 2This is a schematic diagram of the physical mapping structure of the multilayer diffractive optical element in Embodiment 1 (5-20 layers stacked), which is composed of multiple phase plates stacked together. Optical calculations are achieved through multilayer phase accumulation. M layers of transmissive phase plates (20, 21, 22) stacked along the optical axis are shown. The pixelated phase distribution (phase value from 0 to 2π) is represented by grayscale or color level on each phase plate. The optical axis direction is indicated by arrows.

[0079] Figure 3 This is a complete optical path diagram (6 nodes) of the optical solver in Embodiment 1, including a laser, beam expander, spatial light modulator, multi-layer phase plate, Fourier lens, and CCD camera. It shows that the beam emitted by laser 10 passes sequentially through beam expander 11, spatial light modulator 12, multi-layer phase plate stack (20, 21, 22), Fourier lens 30, and finally reaches CCD camera 31. The optical paths between the various optical components are connected by lines and arrows.

[0080] Figure 4 This is a top view of the surface acoustic wave delay line mapping structure in Embodiment 2, including the input IDT, the reflective grating array, and the output IDT. It shows the input interdigital transducer 41 (comb-shaped electrode structure), the reflective grating array 42 (several parallel strip reflective gratings), the output interdigital transducer 43 (comb-shaped electrode structure) fabricated on the upper surface of the piezoelectric substrate 40, and the schematic direction of the propagation of the surface acoustic wave 50 from the input IDT to the reflective grating array and the output IDT.

[0081] Figure 5 This is an example diagram of the surface acoustic wave output echo train in Example 2, showing the timing output of the echo pulses. With time as the horizontal axis and amplitude as the vertical axis, the arrival time of each echo pulse (such as t1, t2, t3) and its corresponding conclusion symbols (C1, C2, C3) are shown.

[0082] Figure 6 This is a schematic diagram of the working principle of the partial differential equation Green's function solver in Example 3 (3 layers), consisting of three layers: angular spectrum decomposition, multi-layer phase mapping, and output solution distribution. Three sub-figures are shown: the left side is a schematic diagram of Green's function angular spectrum decomposition (G(r) is decomposed into several spatial frequency components), the middle side is a schematic diagram of multi-layer phase mapping (each frequency component corresponds to the phase distribution of each phase plate), and the right side is a schematic diagram of the output plane solution distribution (the complex amplitude distribution of the diffraction pattern corresponds to the solution function u(r)).

[0083] Figure 7This is the overall flowchart of the method of the present invention (4 steps), which consists of four steps: solidification, input encoding, physical mapping, and output decoding. The flowchart shows the four steps: solidification step (creating a physical mapping structure) → input encoding step (encoding the known quantity into the incident physical field) → physical mapping step (generating the output physical field through the physical mapping structure) → output decoding step (extracting the solution from the output field), and the steps are connected by arrows.

[0084] Figure 8 This is a schematic diagram of the physical mapping of causal triples in Example 4, including cause node A, cause node B, and effect node C.

[0085] Figure 9 This is a timing diagram of the physical process of cause-and-effect deduction in Example 4, which shows the pulse timing and reflects the physical cause-and-effect deduction. Detailed Implementation

[0093] This embodiment further combines Figure 7 The overall method flowchart shown sequentially achieves a complete closed loop through four steps: solidification, input encoding, physical mapping, and output decoding, and works in conjunction with the device described in claim 1.

[0094] This implementation example Figure 1 As shown, by Figure 2 The multi-layer phase plate structure shown is... Figure 3 The optical path shown is implemented; Figures 4-6 Different implementation methods are shown.

[0095] (This embodiment corresponds to the device described in claim 1. As shown in the overall architecture in Figure 1, it is realized through the multi-layer phase plate shown in Figure 2 and the optical path shown in Figure 3; Figures 4-6 show different embodiments.)

[0096] (This embodiment corresponds to the device described in claim 1 and the method described in claim 2).

[0097] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0098] Example 1: Solver of Optical Diffraction Linear Equations

[0099] This embodiment uses optical diffraction to solve the linear equation system.

[0100] (a) Fabrication of physical mapping structures

[0101] During the manufacturing stage, the solution operator—i.e., the pseudo-inverse A⁺ of matrix A—is first decomposed into the phase distributions φ1(x,y), φ2(x,y), …, φ of M layers of transmission diffraction optical elements using angular spectrum propagation theory and gradient descent optimization algorithms. m (x,y). Specifically:

[0102] Let the system of linear equations to be solved be Ax = b, where A is an N×N coefficient matrix, b is a known constant vector, and x is an unknown vector to be solved.

[0103] Solve for the pseudoinverse A⁺ = (AᵀA)⁻¹Aᵀ with operator A. Perform angular spectral decomposition on the pseudoinverse A⁺ to obtain its propagation components in the spatial frequency domain. Each propagation component corresponds to the phase distribution of a single layer of transmission diffractive optical elements.

[0104] The reverse design algorithm can adopt an iterative optimization method based on gradient descent: the pixelated phase distribution of the multi-layer phase plate is used as the optimization variable, the mean square error between the overall angular spectrum response of the multi-layer structure and the angular spectrum propagation function of the pseudo-inverse A⁺ is used as the loss function, the gradient is calculated by the adjoint method, and the phase distribution is iteratively updated until the loss function converges.

[0105] Each element is an 8th-order quantization phase plate made of polymer or glass (each order corresponds to a phase increment of 2π / 8 = π / 4), with a pixel size of 1 μm (at a working wavelength of 532 nm, a 1 μm pixel size corresponds to 1.88λ; for an imaging system with NA = 0.5, this provides approximately twice the margin of the 0.53 μm Nyquist limit, satisfying the spatial Nyquist sampling requirements). The phase plates are stacked along the optical axis to form a physical mapping structure. The number of layers M is determined based on the required computational accuracy, typically ranging from 5 to 20 layers.

[0106] Preferably, the multilayer diffractive optical element can be composed of Si / SiO2 multilayer media with a single layer thickness of 100-500nm and a refractive index accuracy of ±0.001; the number of layers M of the multilayer phase plate is selected from at least one of 5, 10, 15, and 20 layers, and the optical path difference between adjacent layers is within 15cm in the folded optical path.

[0107] The pixelated phase distribution of each phase plate is obtained through iterative optimization using a reverse design algorithm. The optimization objective is to make the overall angular spectrum response of the multilayer structure approximate the angular spectrum propagation function of the pseudo-inverse A⁺.

[0108] (ii) Input encoding

[0109] The input encoding module includes a 532nm laser, a beam expander, and a spatial light modulator.

[0110] The spatial light modulator encodes the constant vector b of the linear equation system as the amplitude distribution of the incident light field. Specifically, the N components of the constant vector b are mapped to the amplitude values ​​at N spatial locations on the wavefront of the incident light field, and a linear phase tilt is added to the wavefront to spatially label the different components.

[0111] After the collimated beam passes through the physical mapping structure consisting of stacked M-layer phase plates, it forms a diffraction pattern at the back focal plane of the Fourier lens.

[0112] (III) Output Decoding

[0113] The output decoding module consists of a CCD camera and peak positioning logic.

[0114] The spatial frequency coordinates (fx, fy) with the highest intensity in the diffraction pattern directly correspond to the component values ​​of the solution vector x. For an N-dimensional system of equations, N peak spots will appear in the diffraction pattern (if the matrix is ​​full rank), and the spatial frequency coordinates of each peak spot correspond to the component values ​​of the solution vector.

[0115] Specifically, for a two-dimensional system of equations, the spatial frequency of the peak point in the x-direction is x1, and the spatial frequency in the y-direction is x2. For higher-dimensional systems of equations, the peak spot is decoded using its two-dimensional spatial coordinates on the focal plane.

[0116] The peak positioning logic calculates the precise position of each peak spot using the centroid algorithm, and then converts the spatial frequency coordinates into a numerical solution using a preset calibration mapping table.

[0117] (iv) Experimental verification

[0118] For a set of randomly generated full-rank linear equations of order 1000 (the elements of A are uniformly and randomly generated in the interval [-1,1], and the elements of b are uniformly and randomly generated in the interval [-1,1]), the physical propagation mapping time in this embodiment is about 0.5 nanoseconds, and the total delay including the photodetector readout time is about 0.8 nanoseconds; when using a customized high-speed event-driven photodetector array, the total delay can be controlled at the nanosecond level.

[0119] The equivalent matrix, when solved iteratively using LU decomposition on an Intel i9-13900K CPU (24 cores, 32 threads, maximum turbo frequency 5.8 GHz, 36 MB L3 cache; parameters refer to the Intel ARK specification), took approximately 2.1 milliseconds (typical time for 1000-dimensional LU decomposition at a single-core frequency of 5.5 GHz is in the range of 1–5 ms). The light source maintained a constant power of 100 mW, with no dynamic power consumption peaks.

[0120] Compared to operational amplifier-based electrical simulation solvers, the solution latency of this embodiment is reduced by approximately six orders of magnitude (the setup time of electrical solvers is typically in the range of microseconds to milliseconds), and the accuracy is not limited by resistor matching errors.

[0121] Example 2: Surface Acoustic Wave Logic Reasoning Solver

[0122] This embodiment uses a surface acoustic wave delay line array to implement logical reasoning.

[0123] (a) Fabrication of physical mapping structures

[0124] An input interdigital transducer, a reflective grating array, and an output interdigital transducer were fabricated on a 128°YX-cut lithium niobate (LiNbO3) piezoelectric substrate using photolithography and metal deposition processes.

[0125] The layout design of the reflective grating array is as follows: the production rule base (such as all rules in the form of "If A and B then C") is converted into a state transition diagram. Each state in the state transition diagram corresponds to a time window on the substrate, and each transition edge corresponds to a reflective grating.

[0126] The position encoding of the reflective grating is used to determine the start and end states of the transition edge.

[0127] Specifically, the distance *d* between the reflector grating and the input interdigital transducer satisfies the following condition: the round-trip propagation time of the surface acoustic wave, 2d / v, is equal to the time interval from the initial triggering state to the arrival of the final state, where *v* is the propagation speed of the surface acoustic wave on the substrate (for 128° YX-cut lithium niobate, *v* ≈ 3960 m / s, see D. Morgan, "Surface AcousticWave Materials and Applications").

[0128] In Surface Acoustic Waves, Elsevier, 2017, Table 3.4: Typical SAW velocity of LiNbO3 128°YX tangent propagating in the X direction is 3960 m / s.

[0129] The width of the reflective grating encodes the confidence of the transfer edge—the larger the width, the higher the reflection coefficient and the larger the echo pulse amplitude.

[0130] The finger width and finger spacing of the input and output interdigital transducers are determined by the operating frequency. In this embodiment, the center frequency is 200 MHz, and both the finger width and finger spacing are approximately 5 μm (λ / 4, λ = v / f ≈ 3960 / 200×10). 6= 19.8μm; Referring to the IDT design formula f = v_s / d, where d is the sum of the IDT finger pitch and finger width; in this embodiment, d = 2×5 μm = 10 μm, corresponding to f0 = 3960 / (10×10⁻ 6 = 396 MHz;

[0131] The operating frequency of 200 MHz corresponds to an IDT period of d = v / f = 19.8 μm, which means the finger width + finger spacing = 19.8 μm. Taking the finger width = finger spacing ≈ 5 μm at a 1:1 ratio, this is an approximation of the "half-cycle" under actual 200 MHz harmonic conditions.

[0132] (ii) Input encoding

[0133] The input encoding module converts the triggering facts (such as "A is true" or "B is true") into corresponding radio frequency pulse sequences.

[0134] Specifically, the fact "A is true" is encoded as a single-frequency pulse with frequency f_A, and "B is true" is encoded as a single-frequency pulse with frequency f_B. When both facts are true at the same time, a dual-frequency pulse sequence containing f_A and f_B is emitted.

[0135] The radio frequency pulse is converted into a surface acoustic wave by the input interdigital transducer and propagates along the substrate surface.

[0136] (III) Physical mapping process

[0137] Surface acoustic waves propagate on the substrate surface and generate reflected echoes upon encountering the various reflective gratings in the reflective grating array. Since the distances between different reflective gratings and the input interdigital transducer are different, the arrival times of each reflected echo at the output interdigital transducer are different, forming an echo train with a specific time distribution.

[0138] The arrival time of the reflected echo is determined by the position of the reflector grating, which in turn is determined by the transition edges in the state transition diagram. Therefore, the temporal distribution of the echo train directly encodes the reasoning path—which preconditions are triggered, which intermediate states are passed through, and which conclusion is ultimately reached.

[0139] (iv) Output Decoding

[0140] The output decoding module detects the echo train at the output interdigital transducer and identifies the arrival time of the first valid pulse that exceeds the preset threshold by using a threshold judgment.

[0141] The arrival time is directly mapped to a symbolic identifier for the reasoning conclusion through a pre-defined lookup table. For example, the corresponding conclusion C1 is the arrival time in the interval [t1, t2), the corresponding conclusion C2 is the arrival time in the interval [t2, t3), and so on.

[0142] This embodiment achieves zero-clock-cycle rule reasoning: the delay from the input fact pulse to the output conclusion pulse depends only on the propagation time of the surface acoustic wave on the substrate (microseconds for typical substrate sizes), requiring no microprocessor intervention whatsoever.

[0143] Example 3: Optical Green's Function Solver for Partial Differential Equations

[0144] This embodiment uses optical diffraction to solve partial differential equations.

[0145] (I) Overview of Principles

[0146] The technical problem solved by this embodiment is: for partial differential equations of the form L[u(r)] = f(r) (where L is a linear differential operator and f is a known source term), the value of the unknown function u in a specified region can be directly solved using optical diffraction structures.

[0147] The core principle of this embodiment is the Green's function method—representing the solution of a partial differential equation as the convolution of the source term and the Green's function: u(r) = ∫G(r, r')f(r')dr'. In the spatial frequency domain, this convolution operation is transformed into a multiplication operation, which can be implemented in parallel in the angular spectral domain using multilayer diffractive optical elements.

[0148] (ii) Fabrication of physical mapping structures

[0149] The specific process for the curing step is as follows:

[0150] The first step is to construct the Green's function G(r, r') based on the linear differential operator L of the partial differential equation, which satisfies L[G(r, r')] = δ(r - r').

[0151] The second step is to expand the Green's function in the spatial frequency domain (angular spectrum domain) into a superposition of spatial frequency components:

[0152] G(r, r') = Σ_k Ĝ(k) e^{ik·(r-r')}

[0153] Where Ĝ(k) is the angular spectrum (spatial spectrum) of the Green's function.

[0154] The third step is to map each angular spectral component Ĝ(k) e^{ik·r} to a local phase modulation of a certain layer in a multilayer diffractive optical element—the spatial frequency of the phase modulation corresponds to k, and the modulation amplitude corresponds to the amplitude of Ĝ(k).

[0155] The fourth step is to superimpose the phase modulations of all layers according to the weights of the angular spectral components to form a complete phase plate stack. Each phase plate is a transmissive element made of polymer or glass, with a pixel size of 1 μm and a quantization order of 8 (each order has a phase increment of 2π / 8 = π / 4).

[0156] (III) Input Encoding and Output Decoding

[0157] In the input encoding stage, the source term f(r) is encoded as the wavefront amplitude distribution of the incident light beam. Specifically, the values ​​of f(r) at each point in space are mapped to the amplitude transmittance distribution of the incident light wavefront.

[0158] After the light beam passes through the stacked phase plates, the complex amplitude distribution of the diffraction pattern formed on the output plane directly corresponds to the solution u(r) of the partial differential equation. This relationship is guaranteed by the angular spectrum propagation theory—the overall angular spectrum response of the multilayer phase structure is equal to the angular spectrum of the Green's function G. Therefore, after the incident light field (encoded f) passes through this structure, the output light field (encoded u) satisfies u = G*f.

[0159] The output decoding module extracts the phase and amplitude information of the complex amplitude from the diffraction pattern through interferometry or holographic reconstruction methods, and restores the values ​​of the solution function u at each spatial location.

[0160] (iv) Differences between this embodiment and the prior art

[0161] The difference between electrical simulation solvers and electronic simulation solvers: Electronic simulation solvers (such as resistor networks based on operational amplifiers) simulate the solution of equations by bringing the circuit to steady-state equilibrium through negative feedback. The solution process is an iterative convergence process - the circuit needs time to establish a steady state, and the accuracy is limited by the matching error of the resistor elements (typically 1%).

[0162] The optical Green's function solver in this embodiment can form a solution distribution on the output plane through a single propagation of light waves in a multi-layer phase structure. The solution delay depends only on the propagation time of light in the phase plate stack (nanosecond level), and the accuracy is limited by the manufacturing accuracy of the phase plates (up to λ / 100, corresponding to the subwavelength level).

[0163] Based on the equivalent surface error of approximately 5.3 nm corresponding to a working wavelength of 532 nm (refer to MC Troparevsky et al., "Modeling and design of metasurfaces and diffractive optical elements", Adv. Opt. Mater. 5:1600679, 2017), and there is no uncertainty issue regarding iterative convergence.

[0164] The difference from numerical solvers: Traditional numerical solvers for partial differential equations (such as the finite element method and the finite difference method) are executed on digital processors, requiring discretization of the mesh and iterative solutions. Their computational complexity increases with the number of mesh points (typically O(N²) to O(N³)). In contrast, the computational complexity of the optical solver in this embodiment is independent of the number of mesh points—the calculation of all spatial grid points is performed in parallel as light passes through the stacked phase plates.

[0165] The difference between this embodiment and Embodiment 1 (linear equation solver) lies in the different objects being solved—Embodiment 1 deals with the algebraic equation Ax=b, whose solution operator is the pseudo-inverse of a matrix; this embodiment deals with the differential equation L[u]=f, whose solution operator is the Green's function of the differential operator. Mathematically, they belong to different problem types, but both are implemented through the same physical framework of angular spectral decomposition and multi-layer phase mapping.

[0166] Explanation of reference numerals in the attached figures

[0167] Example 4: Causal Triple Physics Derivation Solver

[0168] This embodiment is based on the physical mapping structure of Embodiment 1 to realize the physical deduction of causal triples. A causal triple (cause node A, cause node B, effect node C) represents the causal rule "cause A and cause B, then effect C". (I) Solidification of causal rules: The causal rule library is converted into a state transition diagram. Each causal triple corresponds to a transition edge. The state of the cause node corresponds to the input condition, and the state of the effect node corresponds to the deduction conclusion. The position and width of the transition edge are encoded in the non-uniform parameter space distribution of the physical mapping structure (optical phase plate or surface acoustic wave reflector array), and are permanently solidified during manufacturing. (II) Input encoding: The current fact (cause node A is true, cause node B is true) is encoded as the incident physical field (optical field amplitude / phase distribution or radio frequency pulse sequence). (III) Physical deduction: The incident physical field is solved by the simulation domain of the causal combination through the solidified physical mapping structure and the non-uniform parameter space distribution. The output physical field directly encodes the causal deduction result, and there is no instruction flow throughout the process. (iv) Output Decoding: Extract the state of the result node from the measurable features (spot position, pulse timing) of the output physical field to realize causal inference. The physical propagation time of the entire causal inference process is the solution delay (nanosecond to microsecond level), without the instruction fetch-decode-execution process.

[0169] 1—Input encoding module; 2—Physical mapping structure; 3—Output decoding module;

[0170] 10—Laser; 11—Beam expander; 12—Spatial light modulator;

[0171] 20—Transmission phase plate (layer 1); 21—Transmission phase plate (layer 2); 22—Transmission phase plate (layer M);

[0172] 30—Fourier lens; 31—CCD camera;

[0173] 40—Piezoelectric substrate; 41—Input interdigital transducer; 42—Reflective grating array; 43—Output interdigital transducer;

[0174] 50—Propagation direction of surface acoustic wave; 51—Incident pulse; 52—Reflected echo train;

[0175] 60—Schematic diagram of Green's function angular spectrum decomposition; 61—Schematic diagram of multi-layer phase mapping; 62—Distribution of solution in the output plane.

[0176] The core value of this application lies in directly solidifying the problem-solving operators into the parameter distribution of the physical medium through a physical mapping structure, thereby achieving deterministic computation with nanosecond-level latency and no instruction flow.

[0177] [Patent Family Synergy Statement]

[0178] This application is one of 11 deterministic computing architecture patent families filed by the applicant on the same day. The inventive subject matter, all technical features claimed in the claims, and the specific embodiments shown in the drawings are fully and independently disclosed in the specification and drawings of this application, and can be implemented and reproduced by a person skilled in the art without relying on other applications in the same family.

[0179] The technical solutions claimed in this application are structurally independent from other applications in the same family at the physical implementation level, each containing a complete physical structure, signal flow path, range of technical parameters, and specific embodiments.

[0180] The distinguishing feature of the adjacent application "Inference Acceleration and Weight Storage Device Based on Solid-State and Acoustic Physical Media" is as follows: This application solidifies a general solution operator (algebraic equation system / partial differential equation / logic graph derivation) with physical medium parameters and performs passive simulation domain solution on the incident physical field, focusing on the solidification of the solution operator and causal derivation; while the other application in this family achieves inference acceleration through the synergy of heterogeneous physical effects (first physical effect inference + second physical effect weight storage, interconnected via FPGA). This application focuses on the solidification of operators and causal derivation of a single physical medium, while the other application in this family focuses on the synergy and weight storage of heterogeneous physical effects. The two differ in their physical medium organization, processing objects, and technical means.

[0181] The distinguishing feature of the adjacent application "Deterministic Symbolic Computation Device and Method Based on Optical Physical Fields" is that: this application uses physical medium parameters to solidify general solution operators (covering algebraic equation systems, finite state machines, partial differential equations, and logic diagram derivations), without limiting it to a specific physical medium; while the other application uses an optical medium to pre-solidify a symbolic correlation matrix, targeting only discrete symbolic computation tasks. The technical solution of this application does not depend on an optical medium; the two differ in physical medium type, solution operator type, and computational task scope.

[0182] [Industry Application]

[0183] This invention belongs to the field of physical computing hardware technology. It can be manufactured and used in industry and can be applied to deterministic computing tasks, artificial intelligence inference acceleration, symbol generation and detection, quantum precision measurement and other fields, and has clear industrial applicability.

[0184] [Data Source Explanation] The specific performance data (including but not limited to latency, power consumption, sensitivity, etc.) listed in this specification are all obtained based on publicly available theoretical analysis and comparison with published similar measured parameters. In actual hardware implementation, there may be differences due to different process nodes, component selections and system integration schemes. This data is used to illustrate the relative advantages of the technical solution of this application compared with existing solutions, and does not constitute a limitation on the scope of protection of this application.

[0185] [Defensive Disclosure]

[0186] This invention can also work in conjunction with a lightweight, edge-embedded natural language understanding model, where the small model acts as a semantic receiver to convert user intent into structured instructions, and the physical computing device acts as an executor to respond to physical instructions in real time. The device described in this invention can be connected to existing computer systems as a coprocessor in the form of a PCIe accelerator card, without modifying the instruction flow of the host operating system, achieving a human-machine collaborative division of labor of 99% physical execution and 1% human decision-making. It is suitable for drones, autonomous mobile robots, vehicle terminals, or edge computing devices.

Claims

1. A deterministic computing device based on the permanent solidification of physical medium parameters, characterized in that... , include: The operator solidification module is used to permanently solidify the solution operator of the deterministic computational problem to be solved during the manufacturing process through the geometric structural parameters and / or material physical parameters of the physical medium, forming a non-uniform parameter spatial distribution; An incident physical field excitation module is used to receive a problem instance, encode known quantities as at least one property of the incident physical field, and apply it to the physical medium; The physics solution module is used to pass the incident physical field through the physical medium, perform simulation domain solution operations on the incident physical field based on the non-uniform parameter spatial distribution, and generate the output physical field; The solving operator is at least one of the following: a set of algebraic equation solving operators, a set of finite state machine execution operators, a set of partial differential equation solving operators, and a set of logic diagram derivation operators, and is permanently solidified by the non-uniform parameter spatial distribution of the physical medium; The output demapping module is used to sense measurable physical features in the output physical field that are related to the solution, and output a numerical value or symbol representing the solution; The physical medium is in a passive state without any external control signals during the solution process. It does not execute any form of instruction flow internally, has no clock cycle, and has no instruction fetch-decode-execute process. Furthermore, the light source of the incident physical field excitation module is an external excitation signal. During the solution process, there are no external electrical or control signals inside the medium being solved. The "excitation" of the light source and the "passive solution" of the medium are decoupled at the physical level.

2. A deterministic solution method based on permanent curing of physical medium parameters, characterized in that... This includes the following steps: Based on the deterministic computational problem to be solved, the solution operator is obtained analytically or numerically and mapped to the non-uniform parameter spatial distribution of the physical medium; Receive a problem instance, encode the known quantities into an incident physical field, and apply it to the physical medium; The incident physical field is passed through the physical medium, and the incident physical field is directly solved by the simulation domain based on the non-uniform parameter spatial distribution; The sensor detects measurable physical features in the output physical field that are related to the solution, and outputs a numerical or symbolic representation of the solution. During the solution process, no external electrical or control signals are applied to the physical medium to change its parameter distribution.

3. The apparatus or method according to claim 1 or 2, characterized in that... The physical medium includes at least one of the following: a diffractive optical multilayer structure, a surface acoustic wave delay line array, and a subwavelength metasurface unit array.

4. The apparatus or method according to claim 1 or 2, characterized in that... The solution operator is selected from at least one of the following operator groups: algebraic equation solution operator group, finite state machine execution operator group, partial differential equation solution operator group, and logic diagram derivation operator group.

5. The apparatus or method according to claim 1 or 2, characterized in that... The non-uniform spatial distribution of the physical medium is solidified in the form of geometric structural parameters, including at least one of thickness distribution, refractive index distribution, and reflective grating position distribution. These parameters are permanently formed during the manufacturing process through photolithography, etching, or deposition processes and cannot be changed by electrical signals during the calculation process.

6. The apparatus or method according to claim 1 or 2, characterized in that... The number of layers M of the physical medium is selected from at least one of 5, 10, 15, and 20 layers, and the optical path difference between adjacent layers is within 15 cm in the folded optical path.

7. The apparatus or method according to claim 1 or 2, characterized in that... The non-uniform parameter spatial distribution cannot be rewritten by electrical signals after manufacturing, and the analog domain solution operation is directly completed through the geometric parameters and material physical parameters solidified in the physical medium.

Images