A photon solver and wind farm reconstruction method
By implementing linear mapping and nonlinear activation of the hidden layer of partial differential equations through a photonic solver, the problems of high computational complexity and low energy efficiency in existing wind field reconstruction methods are solved, achieving efficient and real-time wind field reconstruction, which is suitable for edge computing and scenarios with high timeliness requirements.
Patent Information
- Application Number
- CN202611275689.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing wind field reconstruction methods suffer from high computational complexity and low energy efficiency when dealing with high-resolution, large-scale, or strongly nonlinear wind field problems, making it difficult to meet real-time requirements. Furthermore, existing optical solutions are insufficient in terms of hardware implementation for physical information extreme learning machines.
By employing a photonic solver, a multi-channel tunable light source module, a photonic chip, a photodetector array, and an analog nonlinear activation circuit, combined with host computer control, the hidden layer linear mapping and nonlinear activation of partial differential equations are realized. The number of hidden layer nodes is expanded by using time-series wavelength scanning, thereby reducing computational complexity and power consumption.
It achieves efficient and real-time wind field reconstruction, reduces hardware costs and power consumption, is suitable for edge computing scenarios, has high solution efficiency and stability, and is suitable for scenarios with high timeliness requirements such as real-time monitoring of tropical cyclones and aviation meteorological support.
Smart Images

Figure CN123115573A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of optoelectronic hybrid computing and meteorological detection technology, specifically to a photonic solver and a wind field reconstruction method, which can be used to solve the wind field control equations and predict and reconstruct wind field data, belonging to the interdisciplinary fields of scientific computing, optical computing and meteorology. Background Technology
[0002] Wind field reconstruction is a core technology in meteorological forecasting, tropical cyclone monitoring, wind energy resource assessment, and aviation meteorological support. The evolution of the atmospheric wind field is controlled by a set of nonlinear partial differential equations, and their accurate solution directly affects the accuracy and timeliness of severe weather warnings. Currently, wind field reconstruction mainly relies on traditional numerical methods such as the finite volume method, finite element method, and finite difference method. These methods discretize the continuous solution domain into a grid, transforming the partial differential equations into a system of algebraic equations for solution.
[0003] When dealing with high-resolution, large-scale, or strongly nonlinear wind field problems, the above methods face the following significant drawbacks: (1) Traditional methods require multiple linear algebra operations, especially in three-dimensional and higher-dimensional cases, where the computational complexity increases exponentially, requiring extremely high computing resources and storage capacity, making it difficult to meet real-time computing needs. (2) As semiconductor processes approach physical limits, traditional electronic computing systems face serious quantum tunneling and thermal effects, resulting in a sharp increase in power density, making it difficult to deploy on power-constrained edge computing platforms (such as drones, smart sensors, and portable devices). (3) The solution accuracy of traditional methods heavily depends on the quality of mesh generation. For irregular solution domains or multi-scale coupled problems, mesh generation itself constitutes a major technical bottleneck.
[0004] In recent years, Physical Information Extreme Learning Machines (PILMs) have been proposed for solving partial differential equations. The core idea is to directly embed the governing equations, boundary conditions, and initial conditions of the partial differential equations as physical constraints into the neural network training process, approximating the solution through a single hidden layer feedforward network. The advantage of this method is that it eliminates the need for backpropagation iterative training, resulting in significantly higher training efficiency than traditional deep neural networks. In recent years, it has been preliminarily explored for applications such as wind field reconstruction and other physical field inversion problems.
[0005] However, the computational overhead of the Physical Information Extreme Learning Machine (PIEM) is still mainly concentrated on the linear mapping of hidden layers and nonlinear activation operations. As the number of hidden layer nodes increases with the accuracy requirements of wind field reconstruction, the electronic processor faces problems such as large computational latency and high energy consumption, making it difficult to meet the needs of real-time reconstruction of high-resolution wind fields. Optical computing technology, due to its inherent high parallel processing capability and extremely low heat dissipation, has been explored for accelerating related computational processes. However, existing optical solutions still have shortcomings in terms of integration, nonlinear processing capability, or adaptability to wind field reconstruction tasks. There is still a lack of a hardware solution for wind field reconstruction that can efficiently realize the linear mapping of hidden layers and nonlinear activation operations in the PIEM. Summary of the Invention
[0006] To address the issues of high computational complexity and low energy efficiency in existing wind field reconstruction methods using purely electronic schemes, which make it difficult to meet real-time requirements, and the shortcomings of existing optical schemes in hardware implementation of physical information extreme learning machines, this invention provides a photonic solver and a wind field reconstruction method based on it. The photonic device realizes the linear mapping of the hidden layer in solving the partial differential equation of the wind field, nonlinear activation is achieved through analog circuits, and regression solution is achieved through a host computer, thereby improving the computational efficiency and real-time performance of wind field reconstruction and reducing the power consumption and cost of edge deployment.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a photon solver and a wind field reconstruction method based thereon.
[0008] The structure of the photon decoder includes the following parts:
[0009] A multi-channel tunable light source module, centrally controlled by a host computer, generates optical signals and modulates the independent variable information of partial differential equations onto the light intensity. The module includes at least two output channels, each connected to one of the multiple input waveguide ports of a photonic chip. The number of output channels is greater than or equal to the number of independent variables in the partial differential equation to be solved. The module outputs light of a single wavelength simultaneously, with all output channels having the same wavelength. The host computer sends wavelength switching commands to the module according to a preset wavelength sequence, controlling it to synchronously switch to the target wavelength. Simultaneously, the host computer maps the multi-dimensional independent variable values of the current sampling point to the intensity modulation signals of each output channel, synchronously controlling the output light intensity of each channel. Intensity modulation and wavelength switching are strictly time-locked and are both completed by synchronously issuing control commands from the host computer.
[0010] The photonic chip has multiple input waveguide ports at its front end and an output waveguide array at its rear end. The number of input waveguide ports matches the number of output channels of the multi-channel tunable light source module, used to receive optical signals from each channel. The photonic chip also includes a multimode waveguide region, into which optical signals from each input waveguide port propagate. Since each output port has a different light intensity transmission coefficient for optical signals from different input ports, the intensity of each output light is equivalent to a linear superposition of the light intensities of each input port multiplied by their respective transmission coefficients. The linear mapping relationship of the optical signal propagating in the multimode waveguide region is determined by the physical structure of the multimode waveguide region and is fixed after manufacturing. The output waveguide array is used to spatially sample the optical field output from the multimode waveguide region, outputting multi-channel optical intensity signals.
[0011] The multimode waveguide region of the photonic chip has a width of not less than 20 micrometers and a length of not less than 4 millimeters. The output waveguide array is a tapered waveguide array, used to perform spatially selective coupling sampling of the transverse optical field at the output end face of the multimode waveguide region, projecting the continuous optical field into a finite-dimensional output.
[0012] A photodetector array is used to convert the multi-channel optical intensity signal into a multi-channel electrical signal. The photodetector array is matched to the number of output channels of the photonic chip.
[0013] An analog nonlinear activation circuit is used to perform a nonlinear transformation on the multi-channel electrical signal and output nonlinear hidden layer features. The analog nonlinear activation circuit is preferably an analog circuit based on a differential amplifier pair, whose output-input relationship satisfies the hyperbolic tangent function form.
[0014] The host computer controls the wavelength switching and intensity modulation of the multi-channel tunable light source module and receives the nonlinear hidden layer features. For any sampling point in the solution domain, i.e., a set of determined independent variable values, the host computer keeps the independent variable values unchanged and controls the multi-channel tunable light source module to sequentially traverse the preset n wavelengths, acquiring the light intensity signals of m output channels at each wavelength; the n sets of m-dimensional output vectors obtained from the n wavelength switching are concatenated in the column direction to form an N = m×n-dimensional extended hidden layer output vector, which serves as the hidden layer feature of the sampling point. After completing the scanning of all wavelengths at the sampling point, the host computer controls the multi-channel tunable light source module to move to the next sampling point, changes the independent variable values, and repeats the above complete process. Subsequently, based on the physical information extreme learning machine framework, the host computer constructs physical constraints according to the control equations, boundary conditions, and initial conditions of the partial differential equation, solves the output layer weights through regression calculation, and outputs the numerical solution of the partial differential equation based on the output layer weights.
[0015] The number of wavelength scans, n, is not a fixed constant, but a dynamic parameter adaptively determined by the host computer based on the computational complexity and preset accuracy requirements of the partial differential equation to be solved. For strongly nonlinear or multi-scale coupled partial differential equations, the host computer configures a larger value for n, corresponding to more hidden layer nodes to ensure the model's expressive power. This mechanism allows the solver to flexibly adapt to a wide range of problem categories, from low-dimensional linear to high-dimensional nonlinear, by adjusting the number of wavelength scans without changing any hardware structure.
[0016] In the preset n wavelength sequences, the interval between adjacent wavelengths is sufficient to effectively expand the number of output nodes in the hidden layer, that is, the wavelength interval is greater than or equal to the wavelength resolution of the photonic chip, so as to ensure that there is sufficient distinguishability between each group of output vectors.
[0017] The wind field reconstruction method based on photon solvers includes the following steps:
[0018] S1. The host computer obtains the solution domain in the wind field reconstruction task, selects a wavelength from the preset wavelength sequence, and controls the multi-channel tunable light source module to switch to the wavelength synchronously. At the same time, the host computer assigns the multi-dimensional independent variable values of the current sampling point to the intensity modulation signal of each output channel, and synchronously controls the output light intensity of each output channel.
[0019] The independent variables of the wind field control equation correspond to different output channels of the multi-channel tunable light source module, with each channel carrying one independent variable component. The host computer drives the intensity modulation port of each output channel to map the normalized variable values to the corresponding channel's output light intensity change. All output channels of the multi-channel tunable light source module output light of the same wavelength at the same time.
[0020] S2. The optical signals from each output channel are injected in parallel into multiple input waveguide ports at the front end of the photonic chip and propagate into the multimode waveguide region. The optical signals at each input port undergo diffraction and interference in the multimode waveguide region. Since each output port has a different light intensity transmission coefficient for optical signals from different input ports, the output light intensity of each output port is a linear superposition of the light intensities of each input port multiplied by their corresponding transmission coefficients.
[0021] S3. Multi-channel optical intensity signals are acquired through the output waveguide array and transmitted to the analog nonlinear activation circuit after photoelectric conversion.
[0022] S4. The analog nonlinear activation circuit performs a nonlinear transformation on the received electrical signal and then transmits it to the host computer.
[0023] S5. The host computer keeps the value of the independent variable at the current sampling point unchanged, switches the multi-channel tunable light source module to the next wavelength in the wavelength sequence, and repeats steps S1 to S4 until all n preset wavelengths are traversed.
[0024] S6. The host computer concatenates the n sets of m-dimensional output vectors obtained from n wavelength switching in the column direction to form an N = m×n-dimensional extended hidden layer output vector, which serves as the final hidden layer feature of the sampling point.
[0025] S7. The host computer moves to the next sampling point, changes the value of the independent variable, and repeats steps S1 to S6 until the data acquisition and hidden layer feature extraction of all sampling points in the wind field solution domain are completed.
[0026] S8. The host computer uses the nonlinearly transformed feature signal as a physical information limit learning machine framework to construct a set of physical constraint equations based on the wind field control equations, boundary conditions and initial conditions, and solves the output layer weights through regression calculation.
[0027] S9. Using the output layer weights, perform inference calculations on the newly input independent variables, output wind field distribution data, and complete the wind field reconstruction of the target area.
[0028] In step S5, the spacing between adjacent wavelengths in the preset n wavelength sequences should be sufficiently large to effectively expand the number of hidden layer output nodes, i.e., the wavelength spacing is greater than or equal to the wavelength resolution of the photonic chip, to ensure sufficient distinguishability between each group of output vectors. The number of hidden layer nodes N is determined by the specific needs of the problem being solved.
[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0030] (1) This invention assigns the task of linear mapping of the hidden layer with the largest computational load and the highest parallelism in solving the physical information extreme learning machine to a passive photonic chip. Using a Raspberry Pi as the host computer, the solution speed can be achieved with the same speed as an Intel i7 processor. This proves that this invention maintains high solution efficiency while having extremely low requirements for the computing power of the host computer, which can significantly reduce hardware costs and power consumption. It is suitable for edge computing and power-constrained scenarios.
[0031] (2) This invention applies the photon solver to the field of wind field reconstruction, which can complete the task of solving wind field partial differential equations in seconds, which would take several minutes or even hours using traditional numerical methods. It provides a feasible technical solution for scenarios with extremely high time requirements, such as real-time monitoring of tropical cyclones, aviation meteorological support, and rapid assessment of wind energy resources.
[0032] (3) The photonic chip used in this invention is a pure passive structure. The linear mapping of light in the multimode waveguide is entirely dependent on the physical process of light propagation. It does not require any adjustable optical components and has the characteristics of simple structure and extremely low power consumption.
[0033] (4) This invention utilizes a time-series wavelength scanning technique with unified scheduling by a host computer to flexibly expand the number of hidden layer nodes without increasing chip area or power consumption. By switching different wavelengths, it is equivalent to obtaining multiple independent optical linear mapping matrices, and the number of hidden layer nodes is determined by the specific problem being solved. This mechanism does not require changes to the hardware structure; a dynamic balance between solution accuracy and computational efficiency can be achieved simply through wavelength configuration.
[0034] (5) The present invention enables multiple independent variable information under the same wavelength to enter the photonic chip for linear mapping through a multi-port parallel injection architecture, making full use of the spatial parallelism of light propagation; combined with the time-series wavelength scanning mechanism, it realizes the joint reuse of spatial dimension and wavelength dimension, significantly improving the utilization efficiency of hardware resources.
[0035] (6) The present invention has strong robustness to the initial configuration represented by different random weights, small fluctuation range of solution accuracy, strong robustness to hardware non-ideal factors, and high stability and reliability. Attached Figure Description
[0036] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:
[0037] Figure 1 This is a schematic diagram of the overall architecture of the photon solver provided in Embodiment 1 of the present invention.
[0038] Figure 2 This is a schematic diagram of the photonic chip waveguide structure provided in Embodiment 1 of the present invention.
[0039] Figure 3 The input-output characteristic curve of the analog nonlinear activation circuit provided in Embodiment 1 of the present invention is shown.
[0040] Figure 4 This is a schematic diagram of the wind field reconstruction method based on a photon solver provided in Embodiment 2 of the present invention.
[0041] Figure 5 This is a comparison chart of the tropical cyclone wind field reconstruction results in Embodiment 3 of the present invention.
[0042] Figure 6 This is a schematic diagram illustrating the robustness solution in Embodiment 3 of the present invention.
[0043] Figure 7 This is a graph showing the relationship between the mean absolute error and the number of hidden layer nodes in Embodiment 4 of the present invention.
[0044] Figure 8 This is a graph showing the wavelength correlation coefficient of the photonic chip provided in Embodiment 4 of the present invention. Detailed Implementation
[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0046] In the description of this invention, the term "comprising" and any variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.
[0047] Example 1: The specific structure of the photon solver.
[0048] like Figure 1 As shown, the photonic solver provided in this embodiment includes: a multi-channel tunable light source module 1, a photonic chip 2, a photodetector array 3, an analog nonlinear activation circuit 4, and a host computer 5.
[0049] The multi-channel tunable light source module 1 includes multiple tunable semiconductor lasers and their corresponding intensity modulators. Each laser has wavelength tuning and intensity modulation capabilities. The wavelength adjustment range of each laser is C-band, with an output power of 10 to 20 milliwatts, a linewidth of approximately 10 MHz, and a side-mode rejection ratio greater than 35 dB. Each laser is equipped with a heat sink and temperature control circuitry to maintain a constant operating temperature, thus avoiding the impact of temperature fluctuations on wavelength and light intensity.
[0050] The number of output channels in the multi-channel tunable light source module 1 is greater than or equal to the number of independent variables in the partial differential equation to be solved. Each output channel is connected to one of the multiple input waveguide ports at the front end of the photonic chip 2. In this embodiment, taking a three-dimensional space-time problem as an example, the independent variables of the partial differential equation are x, y, and t, which correspond to three output channels, and each output channel is connected to one of the three input waveguide ports of the photonic chip.
[0051] Photonic chip 2 is manufactured using standard SOI technology, with a waveguide thickness of 220 nanometers. For example... Figure 2As shown, the chip has multiple input waveguide ports at the front end and an output waveguide array at the rear end. Each input waveguide port receives optical signals from each output channel of the multi-channel tunable light source module 1. The optical signals are transmitted through the input waveguides and then enter the multimode waveguide region. Since each output port in the multimode waveguide region has a different light intensity transmission coefficient for optical signals from different input ports, the output light intensity of each path is equivalent to the linear superposition of the light intensities of each input port multiplied by the corresponding transmission coefficient. The optical signal propagates in the multimode waveguide region, and its linear mapping relationship is determined by the physical structure of the multimode waveguide region and is fixed after manufacturing. A tapered waveguide array is set at the output end to convert the output light field into multi-channel light intensity signals. In this embodiment, the number of output channels m = 8.
[0052] In this embodiment, the multimode waveguide region has a width of not less than 20 micrometers and a length of not less than 4 millimeters to improve wavelength resolution within a limited chip area. The longer the path length of the optical signal propagating in the multimode waveguide region, the greater the phase accumulation difference between different wavelengths, thereby enabling the chip to resolve smaller wavelength intervals.
[0053] Preferably, the photonic chip is equipped with a temperature control system after packaging. The temperature control system includes a semiconductor cooler, a thermistor, a heat sink, and a temperature control circuit. It maintains a constant chip operating temperature through closed-loop feedback control, avoiding refractive index changes and interference drift caused by temperature fluctuations.
[0054] The photodetector array 3 uses photodetectors corresponding to the number of output channels of the photonic chip 2 to convert light intensity signals into voltage signals.
[0055] Analog nonlinear activation circuit 4 employs an analog circuit based on a differential amplifier pair to achieve hyperbolic tangent nonlinear activation. Under constant current source bias, the currents in the two branches of the differential pair undergo nonlinear distribution with the input differential voltage, forming a saturated output voltage across the load resistor. Its output-input relationship is as follows: U in and U out R represents the input and output voltages. c Where I is the load resistance, I0 is the constant current source current, and U... T Given the thermal voltage constant, the input-output characteristic curve of the analog nonlinear activation circuit is shown below. Figure 3 As shown.
[0056] The host computer 5 uses the low-power embedded platform Raspberry Pi CM4, which connects to the data acquisition card and front-end hardware via an expansion board. The acquisition card has 8 channels, 16-bit resolution, and a 100MS / s sampling rate, supporting a ±5V analog input range. The host computer 5 performs regression solving on the received nonlinear hidden layer output based on the Physical Information Extreme Learning Machine (PIEM) framework. The host computer 5 is also responsible for controlling the wavelength switching and intensity modulation of the multi-channel tunable light source module 1: for any sampling point in the solution domain, the host computer 5 keeps the values of that set of independent variables unchanged, controls the multi-channel tunable light source module 1 to sequentially traverse the preset n wavelengths, acquires the light intensity signals of m output channels at each wavelength, and concatenates the n sets of m-dimensional output vectors obtained from the n wavelength switching in the column direction to form an N = m×n-dimensional extended hidden layer output vector. The host computer 5 develops data acquisition and hardware control programs, as well as the PIEM solution algorithm, based on the Python environment.
[0057] Example 2: Specific process of wind field reconstruction method based on photon solver.
[0058] like Figure 4 As shown, the wind farm reconstruction method provided in this embodiment includes the following steps:
[0059] S1. The host computer obtains the solution domain in the wind field reconstruction task, selects a wavelength from the preset wavelength sequence, and controls the multi-channel tunable light source module to switch to the wavelength synchronously. At the same time, the host computer maps the multi-dimensional independent variable values of the current sampling point to the intensity modulation signal of each output channel, and synchronously controls the output light intensity of each output channel.
[0060] Taking two-dimensional spatial wind field reconstruction as an example, let the independent variables of the wind field control equation be x, y, and t, corresponding to two orthogonal directions in space and the time dimension, respectively. The multi-channel tunable light source module contains three output channels, each connected to one of the three input waveguide ports of the photonic chip. Each output channel carries one independent variable component. The lasers in each output channel employ direct current modulation, with the intensity modulation port of the laser driven by an analog voltage signal generated by a host computer or programmable logic device.
[0061] In this embodiment, the operating wavelength of each laser covers a tuning range of approximately 35 nanometers in the C-band, with a continuous wave output power of 10 to 20 milliwatts, a linewidth of approximately 10 MHz, and a side-mode rejection ratio greater than 35 dB. A temperature control circuit monitors the operating temperature of each laser in real time and maintains it at a constant value to avoid wavelength shift and power fluctuations caused by temperature drift. The multi-channel tunable light source module outputs light of a single wavelength simultaneously, with all output channels having the same output wavelength. This wavelength is determined by the host computer from a preset sequence of n wavelengths λ1, λ2, ..., λ3. nThe wavelengths are selected sequentially, and the entire array is switched via synchronous control commands. Simultaneously, the host computer sends wavelength switching commands and maps the three-dimensional independent variable values x, y, and t of the current sampling point to intensity modulation signals for the three output channels, synchronously controlling the output light intensity of the three channels.
[0062] S2. The optical signals from each output channel are injected in parallel into multiple input waveguide ports at the front end of the photonic chip, and enter the multimode waveguide region to perform linear mapping.
[0063] The modulated optical signals from the three output channels are transmitted through optical fibers to the three input grating couplers of the photonic chip, and then vertically coupled into the single-mode waveguide inside the chip via the grating couplers.
[0064] After the optical signals from each input port enter the photonic chip, they are first transmitted through a single-mode waveguide to the entrance of the multimode waveguide region. The width of the single-mode waveguide is approximately 450 nanometers, supporting fundamental mode transmission. When the optical signal enters the multimode waveguide from the single-mode waveguide, the input optical field interferes within the multimode waveguide.
[0065] The optical signals at each input port propagate in the multimode waveguide region. Let the optical intensity at the i-th input port be I. i The transmission coefficient from the j-th output port to the i-th input port is T. ij For different wavelengths λ, the transmission coefficient matrix T(λ) is different. The light intensity at the j-th output port is... , where l is the number of input ports. Therefore, when single-wavelength optical signals are injected simultaneously from multiple input ports, the system performs spatial linear mapping at that wavelength.
[0066] Mathematically, this propagation process is equivalent to a fixed-weight stochastic linear mapping layer: the input vector x carrying the wind field independent variable information is mapped to the linear output vector of the hidden layer. W eq Let b be a linear mapping matrix determined by the transmission coefficient matrix. eq This is the offset term equivalent to the phase difference determined by the interference result. Since the propagation and interference results in the multimode waveguide are both determined by the physical structure of the waveguide and are fixed after manufacturing, this linear mapping remains unchanged during system operation, which corresponds perfectly to the setting of fixed hidden layer weights after random initialization in the physical limit learning machine algorithm.
[0067] At the output end of the multimode waveguide region, the optical field enters an output waveguide array composed of tapered waveguides. In this embodiment, the output waveguide array contains m = 8 tapered waveguide channels, each of which extracts optical field information from different lateral positions on the output end face of the multimode waveguide. The width of each tapered waveguide gradually narrows along the propagation direction, gradually compressing the collected optical field energy to the width of a single-mode waveguide, and finally outputting it to the corresponding single-mode waveguide. The function of this tapered waveguide array is to spatially sample the lateral optical field on the output end face of the multimode waveguide, projecting the continuous optical field onto a finite-dimensional space. The light intensity value of each output channel retains part of the integral information of the entire optical field on the output end face, rather than sampling only the light intensity of a single spatial point.
[0068] S3. The optical signals from the eight output channels of the photonic chip are coupled to the photodetector array via an optical fiber array. Each photodetector converts the incident light power into a photocurrent, which is then converted into a voltage signal by a transimpedance amplifier. This voltage signal is an analog quantity, and its amplitude is proportional to the output light intensity of the corresponding channel.
[0069] S4. The voltage signal input of eight channels is used to simulate nonlinear activation circuits for nonlinear transformation.
[0070] This circuit is based on a differential amplifier pair, with its core structure being an emitter-coupled differential pair. The emitters of the two transistors in the differential pair are connected to the same constant current source, while their bases receive the input signal and a reference level, respectively. The collectors are connected to the power supply through a load resistor. The circuit uses the differential input voltage as the control quantity, which is the base-emitter voltage of the two transistors in the differential pair.
[0071] Under constant current source bias, the collector currents of the two branches of the differential pair exhibit a nonlinear distribution with respect to the differential input voltage. When using differential output, the relationship between the output voltage and the input voltage is as follows: U in and U out These are the input and output voltages, respectively; I0 is the constant current source current; R... c U is the load resistance. T It is the thermoelectric voltage constant, which is approximately 26 millivolts at room temperature.
[0072] This relationship indicates that the analog nonlinear activation circuit maps the input differential voltage into a amplitude-limited, continuous, and smooth saturated nonlinear output, which is highly consistent with the hyperbolic tangent activation function commonly used in neural networks. By adjusting the constant current source current and the load resistance, the saturation amplitude and slope of the nonlinear curve can be adjusted so that the circuit's operating point falls within the effective range of the nonlinear response.
[0073] In the actual operation of this embodiment, the signals from the eight channels are processed sequentially through the aforementioned analog nonlinear activation circuit. The bias voltage of each channel is finely adjusted to compensate for the DC offset difference between the photodetector and the differential amplifier, ensuring that each channel operates near the center point of the nonlinear region at zero input. This design guarantees the consistency of the output characteristics of different channels in terms of scale and dynamic range after nonlinear mapping.
[0074] The output of the analog activation circuit is an analog voltage signal with limited amplitude, ranging from approximately ±2.5V. It is synchronously acquired by the data acquisition card at a 16-bit resolution and a sampling rate of 100MS / s, converted into a digital signal, and then transmitted to the host computer.
[0075] Preferably, the nonlinear activation function σ is a hyperbolic tangent function. Since this function is differentiable, the spatial and temporal derivatives in the subsequent wind field control equations can be obtained analytically without numerical difference.
[0076] S5. Expand the number of hidden layer nodes by switching wavelengths. If the number of hidden layer nodes needs to be expanded to N, then the required number of wavelength scans is n = N / m. In this embodiment, the number of waveguide output channels is m = 8.
[0077] After the host computer obtains a hidden layer nonlinear output vector H = σ(Z), it keeps the value of the independent variable at the current sampling point unchanged, switches the multi-channel tunable light source module to the next wavelength in the wavelength sequence, and repeats steps S1 to S4 until all n preset wavelengths are traversed.
[0078] Let the preset wavelength sequence be λ1, λ2, ..., λ n For the independent variable values x, y, and t at the current sampling point, the host computer first controls the multi-channel tunable light source module to switch to λ1, and simultaneously modulates x, y, and t onto the light intensity of the three output channels respectively, obtaining the eight-dimensional output vector H. (1) Then, keeping x, y, and t constant, switch the array to λ2 and obtain H. (2) And so on, until all n wavelengths have been scanned.
[0079] The host computer concatenates the n sets of m-dimensional output vectors obtained from n wavelength switching in the column direction to obtain the N = m×n dimensional extended hidden layer nonlinear output vector H. ext .
[0080] The interval Δλ between adjacent wavelengths must be greater than or equal to the wavelength resolution δλ of the photonic chip to ensure sufficient distinguishability between the feature vectors generated by different wavelengths. The definition of wavelength resolution and the calculation method of the correlation coefficient are described in Example 4. When Δλ is greater than or equal to δλ, the chip can effectively distinguish different wavelengths, and the distinguishability between each group of output vectors is sufficient. When Δλ is less than δλ, the chip cannot effectively distinguish adjacent wavelengths, the output vectors are highly correlated, the hidden layer node expansion effect is significantly reduced, and thus the accuracy of wind field reconstruction is affected.
[0081] S6. After completing the full wavelength scan of a sampling point, the host computer moves to the next sampling point, changes the value of the independent variable, and repeats steps S1 to S5 until the data acquisition of all sampling points in the wind field reconstruction solution domain is completed, and the nonlinear output of the extended hidden layer is obtained.
[0082] S7. The nonlinear output of the extended hidden layer is based on the physical information extreme learning machine framework. A set of physical constraint equations is constructed according to the wind field control equation, boundary conditions and initial conditions. The output layer weights are solved by regression calculation.
[0083] Let the solution domain for wind field reconstruction be Ω, the boundary be ∂Ω, and the governing equations for the wind field be: Here, x is the input matrix of the neural network, and the boundary conditions are... The initial conditions are The wind field distribution u(x) is approximated by a single hidden layer neural network: W i and b i The input weights and biases are randomly set and fixed, corresponding to the fixed characteristics of the optical linear mapping. σ(·) is the nonlinear activation function, and c i The output layer weights are to be solved, and N is the number of hidden layer nodes, which is determined by step S5.
[0084] During the training phase, the host computer first selects M within the solution domain Ω. Ω M sampling points are selected on the boundary ∂Ω. b There are several boundary points, and M is selected at the initial time. I There are several initial condition points. For each sampling point, the host computer calculates the residual based on the wind field control equation operator: , where x i These are the sampling points within the solution domain of the neural network. Since the activation function is differentiable, the spatial and temporal derivatives in the wind field control equations can be obtained analytically without numerical differencing. Similarly, boundary residuals are calculated at the boundary points. Calculate the initial residual at the initial point. .
[0085] The host computer constructs the global loss function: .
[0086] residual L Ω L b and L I The output layer weights *c* are nonlinear functions, and the global loss function is nonconvex. The host computer uses a pseudo-Newton method for iterative solution. The computational complexity of the pseudo-Newton method is O(N), which is significantly better than the O(N) time complexity of the Newton method. 2 ).
[0087] S8. Using the obtained output layer weights, perform inference calculations on the newly input independent variables, output wind field distribution data, and complete the wind field reconstruction of the target area.
[0088] Example 3: Solution verification and energy efficiency comparison of photon solver.
[0089] This embodiment uses the system architecture from Embodiment 1 and the wind field reconstruction method from Embodiment 2 to reconstruct the wind field of Hurricane "Lee" to verify the solver's accuracy and feasibility. The mean absolute error of multiple independent repeated experiments is used to verify the system's robustness. The energy efficiency of this embodiment is also compared with that of two other configurations.
[0090] (2) Experimental verification of tropical cyclone wind field reconstruction.
[0091] To verify the solver's ability to reconstruct wind fields, this experiment uses tropical cyclone hurricane "Lee" as the verification task. The governing equations of a typhoon wind field include a continuity equation and a momentum equation, determined by the conservation of momentum under the influence of surface drag and the Coriolis effect. The specific form of the equations is as follows: , , , where ū and v̄ represent the average components of zonal and meridional winds, respectively, and f is the Coriolis parameter.
[0092] The trajectory data for Hurricane Lee was obtained from the International Best Track Climate Management Archive Database, the wind field data from high-fidelity simulations of meteorological research and forecasting models, and the driving data from the final operational global analysis data provided by the U.S. National Center for Environmental Prediction. The experimental configuration used a photonic chip with m = 8 output channels, n = 15 wavelength scans, corresponding to N = 120 hidden layer nodes.
[0093] The solution results are as follows Figure 5 As shown, the system achieves high-precision reconstruction of the wind field of Hurricane Lee at a specified time. The location of the central low-speed zone and the wind speed gradient of the outer ring zone are highly consistent with the true values, and the error distribution is concentrated and always less than 8%. This result verifies the effectiveness of the solver in real complex wind field reconstruction problems.
[0094] (3) Repeated experiments were conducted independently multiple times to verify the robustness of the solution.
[0095] With a fixed number of hidden layer nodes, different transmission coefficient matrices were generated by changing the random seed, effectively resulting in different initialization weights. One hundred independent and repeated experiments were conducted on a tropical cyclone wind field reconstruction verification task. The mean absolute error of each experiment was recorded, and the results are as follows: Figure 6 As shown in the figure, the mean absolute error (MAE) of the tropical cyclone wind field reconstruction task is a scatter plot showing the relationship between the random weight seed and the MAE. The figure indicates that in the validation task, the MAE corresponding to the random seed fluctuates slightly around its mean, with a small range of fluctuation. No significant deviation in accuracy due to different weight initializations was observed. This suggests that the solver is robust to the weight initialization configuration, and the solution accuracy mainly stems from the expansion of the hidden layer node count brought about by the temporal wavelength scanning, rather than the accidental advantage of a specific weight configuration.
[0096] (4) Comparison of energy efficiency of different configurations.
[0097] To verify the high-speed and low-power characteristics of this solver, the solution speed of this embodiment was compared with two other configurations. Comparative Example 1 uses the same physical extreme learning machine framework, but without a photonic chip, relying entirely on the low-power embedded platform Raspberry Pi CM4 for solution. Comparative Example 2 also relies entirely on the high-end desktop processor Intel i7-10750H for solution. Multiple high-precision reconstructions of Hurricane "Li" were performed, and the average time for each solution task was recorded.
[0098] Experimental results show that the mean absolute error of the solution obtained in this embodiment and with Intel i7 remains at the same level of accuracy, while the solution speed is comparable. Specifically, the average time to complete a single solution using only Raspberry Pi CM4 is 3.6 seconds, while using only Intel i7 is 0.55 seconds, and this embodiment is 0.52 seconds. These results fully verify that after the present invention uses a photonic chip to handle the linear mapping calculation, the host computer only needs to perform simple scalar operations and regression solutions, with no significant dependence on processor computing power. This achieves extremely low system power consumption and hardware cost while maintaining high-speed solution.
[0099] Example 4: The impact of the number of hidden layer nodes and the wavelength resolution of the photonic chip on the solver performance.
[0100] This embodiment analyzes the impact of the number of hidden layer nodes and the wavelength resolution of the photonic chip on the performance of solving partial differential equations through system experiments, in order to clarify the effectiveness of expanding the number of hidden layer nodes through time-series wavelength scanning and the constraints of wavelength selection.
[0101] (1) The impact of the number of hidden layer nodes on the solution accuracy.
[0102] This experiment used the same experimental configuration as in Example 3. In the tropical cyclone wind field reconstruction task, the number of hidden layer nodes N was adjusted by changing the number of wavelength scans (n), and the solution accuracy under different node numbers was recorded. The experimental results are as follows: Figure 7 As shown, as the number of hidden layer nodes increases from 8 to 120, the mean absolute error shows a continuous downward trend; when the number of nodes reaches 112 to 120, the mean absolute error has been significantly reduced to an acceptable level, and the effect of further increasing the number of nodes on improving accuracy gradually weakens.
[0103] The experimental results above demonstrate that expanding the number of hidden layer nodes through temporal wavelength scanning is an effective means to improve solution accuracy, verifying the core innovation of this invention. There exists an optimal range for the number of nodes, within which solution accuracy significantly improves with increasing node count. Beyond this range, the accuracy improvement tends to saturate, and further increasing the number of wavelength scans will incur additional measurement time overhead, resulting in diminishing returns. In practical applications, the number of wavelength scans can be flexibly selected to balance solution accuracy and computational efficiency based on the complexity and accuracy requirements of the specific problem.
[0104] (2) Wavelength resolution constrains the selection of wavelength combinations.
[0105] Wavelength correlation coefficient of photonic chips Where I(x,λ) represents the light intensity at the waveguide receiver at position x and wavelength λ, and Δλ represents the wavelength shift. λ This represents the average value over wavelength, <·> x It represents the average value over a spatial location.
[0106] like Figure 8 The correlation coefficient curve shown indicates that when C(Δλ) drops to half of its maximum value, the corresponding wavelength interval δλ represents the chip's wavelength resolution. When expanding the number of hidden layer nodes by switching wavelength combinations, the interval between the selected wavelengths must be greater than δλ to ensure sufficient distinguishability between hidden layer nodes generated by different wavelengths. If the wavelength interval is less than δλ, the chip cannot effectively distinguish between different wavelengths, significantly reducing the hidden layer node expansion effect and thus affecting the solution accuracy.
[0107] Those skilled in the art will understand that the specific numerical values described in the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In practical applications, the number of output channels, the number of wavelength combinations, and the feature dimension can be appropriately adjusted according to the specific type of partial differential equation, the required solution accuracy, and computational resource limitations.
[0108] Furthermore, the photonic solver described in this invention is not only applicable to the applications described in the embodiments, but can also be extended to solving other types of linear partial differential equations and nonlinear partial differential equations, as well as multiphysics coupling problems.
Claims
1. A photonic solver, characterized in that, include: A multi-channel tunable light source module is used to generate light signals and modulate digital information onto light intensity; The photonic chip has multiple input waveguide ports at its front end and an output waveguide array at its rear end. The output channels of the multi-channel tunable light source module are respectively connected to the input waveguide ports. The photonic chip also includes a multimode waveguide region where optical signals from each input waveguide port propagate. Each output port has a different light intensity transmission coefficient for optical signals from different input ports, and the light intensity of each output port is a linear superposition of the light intensities of each input port multiplied by their respective transmission coefficients. A photodetector array is used to convert the multi-channel light intensity signal into an electrical signal; An analog nonlinear activation circuit is used to perform nonlinear transformation on the multi-channel electrical signal and output a nonlinear characteristic signal. The host computer is used to control the wavelength switching and intensity modulation of the multi-channel tunable light source module, and to receive the output electrical signal of the analog nonlinear activation circuit. The host computer controls the multi-channel tunable light source module to switch multiple wavelengths sequentially, and integrates the outputs obtained at each wavelength to form the resources required for solving partial differential equations of configurable scale. The host computer is based on the physical information extreme learning machine framework, which can construct physical constraints according to the control equations, boundary conditions and initial conditions of the partial differential equations, solve the output layer weights through regression calculation, and output the numerical solution of the partial differential equations.
2. The photonic solver according to claim 1, characterized in that, The multimode waveguide region of the photonic chip has a width of not less than 20 micrometers and a length of not less than 4 millimeters.
3. The photonic solver according to claim 1, characterized in that, The photonic chip has one or more ports in its input multimode waveguide region, and the output end of the multimode waveguide is a tapered waveguide array with eight or more ports, which is used to perform spatial selective coupling sampling of the transverse optical field at the output end face of the multimode waveguide region, projecting the continuous optical field into a finite-dimensional output.
4. The photonic solver according to claim 1, characterized in that, The nonlinear transformation performed in the simulated nonlinear activation circuit is a hyperbolic tangent function transformation.
5. A wind field reconstruction method based on the photon solver according to any one of claims 1 to 4, characterized in that, Includes the following steps: The independent variables of the wind field control equation are numerically encoded into the light intensity information of each output channel of the multi-channel tunable light source module, and each output channel outputs light of the same wavelength at the same time. The optical signals from each output channel are injected in parallel into the multimode waveguide region of the photonic chip. The optical signals from each input port interfere in the multimode waveguide region. The light intensity of each output port is a linear superposition of the light intensities of each input port multiplied by the corresponding transmission coefficient. Multi-channel light intensity signals are collected and transmitted to an analog nonlinear activation circuit after photoelectric conversion. The analog nonlinear activation circuit performs nonlinear transformation on the received electrical signals and outputs them to the host computer. The host computer receives the nonlinearly transformed electrical signal and obtains a hidden layer output vector. Keeping the value of the independent variable at the current sampling point unchanged, switch to the next wavelength and repeat the encoding, linear mapping, acquisition and nonlinear transformation operations until all preset wavelengths are traversed. Then, the output vectors obtained at each wavelength are concatenated to form the output vector of the extended hidden layer. Move to the next sampling point, change the value of the independent variable, and repeat the above steps until data acquisition of all sampling points in the solution domain is completed. Based on the physical information extreme learning machine framework, a set of physical constraint equations is constructed according to the control equations, boundary conditions and initial conditions of the wind field control equations, and the output layer weights are solved by regression calculation. Using the output layer weights, inference calculations are performed on the newly input independent variables to output the solution of the wind field control equation and reconstruct the wind field distribution of the target area.
6. The wind farm reconstruction method according to claim 5, characterized in that, The host computer adaptively determines the number of wavelength switching times based on the computational complexity and preset accuracy requirements of the wind field control equation to be solved; for strongly nonlinear or large-scale wind field control equations, a larger number of wavelength switching times is configured to increase the number of hidden layer nodes.
7. The wind farm reconstruction method according to claim 5, characterized in that, The spacing between adjacent wavelengths switched is sufficient to effectively expand the number of output nodes in the hidden layer.
8. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, implements the wind field reconstruction method as described in any one of claims 6 to 7.
