Optimization method and system for multi-port coupling of optical waveguide chip based on deep learning

Abstract

This invention provides a deep learning-based method and system for optimizing multi-port coupling of optical waveguide chips, relating to the field of optical waveguide chip technology. The method includes acquiring port data and extracting mode features, constructing a matching relationship matrix between ports, screening candidate coupling paths and calculating priority scores, then predicting waveguide structure parameters using a neural network, and finally optimizing the structure parameters by iteratively adjusting the matching threshold until preset coupling efficiency and crosstalk requirements are met. This invention achieves automated and intelligent optimization design of coupling paths and structural parameters for multi-port optical waveguide chips.

Description

Technical Field

[0001] This invention relates to the field of optical waveguide chip technology, and in particular to a method and system for optimizing multi-port coupling of optical waveguide chips based on deep learning. Background Technology

[0002] As a core component of integrated photonic systems, the performance of optical waveguide chips is highly dependent on the coupling efficiency of optical signals between different functional regions. In complex optical waveguide networks, achieving efficient, low-crosstalk optical coupling between multiple ports is crucial for improving the overall chip performance. Current technologies for addressing multi-port coupling problems primarily rely on design flows based on analytical models or finite element simulations.

[0003] Conventional design flows typically begin with the physical layout of the chip. Designers pre-define the coupling structure parameters between waveguides, such as coupling spacing and coupling region length, based on experience or simplified coupling models. Electromagnetic simulation software is then used to simulate the designed structure to evaluate its coupling efficiency and crosstalk level. If the simulation results do not meet the design requirements, the structural parameters must be manually adjusted and the simulation repeated. This iterative process is repeated until an acceptable design is obtained. This approach relies heavily on the designer's professional experience and prior knowledge of specific waveguide modes.

[0004] However, the above approach has significant limitations. The design process relies heavily on human experience and trial and error, resulting in low efficiency and difficulty in finding the globally optimal solution. For optical waveguide chips with numerous ports and complex topologies, the number of possible coupling path combinations is enormous, making it almost impossible to manually traverse and evaluate all potential efficient coupling schemes. This easily leads to getting trapped in local optima, resulting in the final design's coupling performance falling short of its potential limit. Methods based on simplified analytical models or local simulations often struggle to accurately characterize complex physical effects such as higher-order mode excitation, mode mismatch, and crosstalk between multiple paths. This leads to discrepancies between the predicted coupling efficiency and the actual chip fabrication and testing performance, resulting in insufficient design reliability. Often, multiple expensive tape-out iterations are required to achieve the target performance, significantly increasing R&D costs and time. Summary of the Invention

[0005] This invention provides a method and system for optimizing multi-port coupling of optical waveguide chips based on deep learning, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a deep learning-based method for optimizing multi-port coupling of optical waveguide chips, comprising: Acquire port topology data and port optical field distribution data of the optical waveguide chip; Wavefront reconstruction and mode decomposition are performed on the port optical field distribution data to identify the fundamental mode component and higher-order mode component of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set. The set of port pairs is determined by waveguide path tracing. The difference between the fundamental mode coupling transmission efficiency and the energy vector of higher-order modes is calculated based on the port mode feature set. Mode matching coefficients are generated, and a matching relationship matrix is ​​constructed. From the matching relationship matrix, filter port pairs whose pattern matching coefficients meet the preset matching threshold, establish a candidate path set, calculate the distance penalty factor based on the port spacing, calculate the priority score in combination with the pattern matching coefficients, select a preset number of coupled paths, and obtain the coupled path configuration; The set of structural parameters is obtained by predicting the waveguide spacing and coupling region length of each coupling path in the coupling path configuration using a neural network. The target port coupling efficiency and non-target port crosstalk level are calculated based on the structural parameter set. When the target port coupling efficiency is lower than the preset efficiency threshold or the non-target port crosstalk level exceeds the preset crosstalk threshold, the matching threshold is adjusted and the path planning step is returned until the threshold requirements are met, and the structural parameter set is output.

[0007] In one optional embodiment, wavefront reconstruction and mode decomposition are performed on the port optical field distribution data to identify the fundamental mode component and higher-order mode components of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set including: Wavefront reconstruction is performed on the optical field distribution data of the ports to obtain the phase distribution and amplitude distribution of each port. The principal curvature direction of the phase distribution and the principal symmetry axis of the amplitude distribution are extracted by the adaptive principal component analysis method. The propagation direction of the fundamental mode component is determined based on the consistency between the principal curvature direction and the principal symmetry axis. A mode decomposition coordinate system is established with the propagation direction of the fundamental mode component as the axis. The eigenmodes of the light field distribution are expanded in the mode decomposition coordinate system. The expansion weights of each mode component are obtained by the singular value decomposition method. The fundamental mode component and each higher-order mode component are identified based on the numerical distribution of the expansion weights. The phase information of the fundamental mode component and each higher-order mode component during propagation is extracted. The difference between the phase information of the fundamental mode component and the phase information of each higher-order mode component is calculated as the inter-mode phase difference. The proportion of the square of the unfolded weight of each mode component to the total sum of squares of the weights is calculated to determine the energy distribution ratio. For each port, the phase difference and energy distribution ratio are bound to the spatial location information and propagation direction information of that port, respectively, to form a port mode feature set containing mode coupling path characteristics.

[0008] In an optional embodiment, extracting the principal curvature direction of the phase distribution and the principal symmetry axis of the amplitude distribution using an adaptive principal component analysis method includes: The phase distribution is subjected to second-order partial derivative calculation to obtain the phase curvature tensor. Based on the spatial distribution characteristics of the phase curvature tensor, a weighted covariance matrix is ​​constructed, where the weight coefficients are determined by the phase gradient magnitude. The weighted covariance matrix is ​​subjected to eigenvalue decomposition, and the eigenvector corresponding to the largest eigenvalue is extracted as the principal curvature direction. The amplitude distribution is measured and symmetry is calculated. The direction of the principal axis of symmetry is identified by rotation invariance test, and consistency is determined by combining the principal curvature direction.

[0009] In one optional embodiment, the expanded weights of each mode component are obtained through singular value decomposition, and the fundamental mode component and higher-order mode components are identified based on the numerical distribution of the expanded weights, including: In the mode decomposition coordinate system, the light field distribution is constructed as a light field matrix. Singular value decomposition is performed on the light field matrix to obtain a sequence of singular values ​​and a set of corresponding singular vectors. The cumulative energy contribution rate of the singular value sequence is calculated, the minimum number of singular values ​​required to make the cumulative energy contribution rate reach a preset contribution rate threshold is determined, and the number of effective mode components is determined. The singular vector set is sorted according to the magnitude of the singular values. The singular vector corresponding to the largest singular value is identified as the fundamental mode component. Other singular vectors within the range of the number of effective mode components after sorting are identified as higher-order mode components in descending order of singular values. The singular values ​​corresponding to each mode component are extracted as expansion weights.

[0010] In one optional embodiment, the port pair set is determined by waveguide path tracing, the difference between the fundamental mode coupling transmission efficiency and the higher-order mode energy vector is calculated based on the port mode feature set, mode matching coefficients are generated, and a matching relationship matrix is ​​constructed, including: Analyze the port topology data, construct a port spatial distribution map, perform waveguide path tracing on the port spatial distribution map, identify port pairs with direct waveguide connections or port pairs with overlapping near-field coupling regions, and form a set of port pairs; The fundamental mode component phase information of the port pair is extracted from the port mode feature set. The difference between the fundamental mode component phase information of the port pair is calculated. The difference between the phase information is compared with the preset coherence conditions to determine the fundamental mode coupling transmission efficiency of the port pair. Combined with the energy distribution ratio of the fundamental mode component of the port pair, the fundamental mode correlation index of the port pair is generated. Extract the energy distribution ratio of higher-order mode components of port pairs from the port mode feature set, construct the higher-order mode energy vector of port pairs, and calculate the Euclidean distance between the higher-order mode energy vectors as the higher-order mode energy difference index of the port pairs. The fundamental mode correlation index is used as the coupling strength weight, and the reciprocal of the higher-order mode energy difference index is used as the mode purity weight. The geometric mean is then used to generate the mode matching coefficient. Fill the pattern matching coefficients of each port pair into the corresponding row and column positions to construct the matching relationship matrix.

[0011] In one optional embodiment, port topology data is parsed to construct a port spatial distribution map. Waveguide path tracing is performed on the port spatial distribution map to identify port pairs with direct waveguide connections or port pairs with overlapping near-field coupling regions, forming a port pair set including: The port topology data is analyzed to extract the spatial coordinates and propagation direction information of each port. The spatial coordinates are mapped to a spatial coordinate system, and the propagation vectors of each port are marked according to the propagation direction information to construct a port spatial distribution map. In the port spatial distribution diagram, starting from the spatial coordinates of each port, a ray is projected along the direction of the propagation vector to detect the intersection of the ray with the propagation vectors of other ports, and the ports with intersections are identified as port pairs with direct waveguide connections. In the port spatial distribution map, the range of the near-field coupling region is determined based on the near-field coupling distance, with the spatial coordinates of each port as the center. It is then detected whether the near-field coupling regions of different ports have spatial overlap, and ports with overlapping near-field coupling regions are identified as port pairs with overlapping near-field coupling regions. The port pairs with direct waveguide connections are combined with the port pairs that overlap with the near-field coupling region to form a port pair set.

[0012] In one optional embodiment, port pairs whose pattern matching coefficients satisfy a preset matching threshold are filtered from the matching relationship matrix to establish a candidate path set. A distance penalty factor is calculated based on the port spacing, and a priority score is calculated in combination with the pattern matching coefficients. A preset number of coupled paths are selected to obtain the coupled path configuration, which includes: Traverse the matching relationship matrix, extract matrix elements whose pattern matching coefficient is greater than a preset matching threshold, record the row and column indices corresponding to the matrix elements, extract the corresponding port pairs from the port pair set according to the row and column indices, mark them as candidate paths, and establish a candidate path set. The spatial coordinates of each candidate path in the candidate path set are extracted from the port topology data. The three-dimensional straight-line distance between ports in each candidate path is calculated as the port spacing. Based on the attenuation characteristics of electromagnetic waves in the transmission medium, the port spacing is multiplied by the medium absorption coefficient to obtain the transmission loss value. The transmission loss value is exponentially transformed to obtain the distance penalty factor. The mode matching coefficients corresponding to each candidate path are extracted. The mode matching coefficients and the reciprocal of the distance penalty factor are weighted and geometrically averaged to calculate the priority score of each candidate path. The priority scores of all candidate paths in the candidate path set are sorted in descending order, and a preset number of candidate paths with the highest priority scores are extracted as coupling paths to obtain the coupling path configuration.

[0013] A second aspect of this invention provides a deep learning-based multi-port coupling optimization system for optical waveguide chips, comprising: The data acquisition unit is used to acquire port topology data and port optical field distribution data of the optical waveguide chip; The mode analysis unit is used to perform wavefront reconstruction and mode decomposition on the port optical field distribution data, identify the fundamental mode component and higher-order mode component of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set. The matching calculation unit is used to determine the set of port pairs by waveguide path tracing, calculate the difference between the fundamental mode coupling transmission efficiency and the energy vector of higher-order modes based on the port mode feature set, generate mode matching coefficients, and construct a matching relationship matrix. The path filtering unit is used to filter port pairs whose pattern matching coefficients meet the preset matching threshold from the matching relationship matrix, establish a candidate path set, calculate the distance penalty factor based on the port spacing, calculate the priority score in combination with the pattern matching coefficient, select a preset number of coupled paths, and obtain the coupled path configuration. The parameter prediction unit is used to predict the waveguide spacing and coupling region length of each coupling path in the coupling path configuration through a neural network, thereby obtaining the structural parameter set. The optimization verification unit is used to calculate the coupling efficiency of the target port and the crosstalk level of the non-target port based on the structural parameter set. When the coupling efficiency of the target port is lower than the preset efficiency threshold or the crosstalk level of the non-target port exceeds the preset crosstalk threshold, the matching threshold is adjusted and the path planning step is returned until the threshold requirement is met, and the structural parameter set is output.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] In this embodiment of the invention, automated data processing and path planning avoid the tedious process of relying on manual experience and repeated trial and error, shortening the design cycle and reducing development costs. Wavefront reconstruction and mode decomposition of the port optical field can accurately identify and quantify the fundamental and higher-order mode components of each port. Extracting the phase difference and energy distribution ratio between modes forms a comprehensive port mode feature set, thereby profoundly revealing the essence of mode matching between ports and providing a precise physical basis for evaluating coupling potential. It realizes rapid and quantitative evaluation of coupling compatibility between ports of the entire chip. By screening and establishing a candidate path set through preset thresholds, and calculating priority scores by combining distance penalty factors and mode matching coefficients, it can intelligently select the preset number of coupling paths with the best overall performance. A closed-loop feedback optimization mechanism is introduced to calculate the actual coupling efficiency and crosstalk level based on the predicted structural parameters and compare them with preset thresholds. This mechanism ensures that the final output structural parameter set can simultaneously meet the stringent performance indicators of high coupling efficiency and low crosstalk, realizing self-verification and self-optimization of design results, and significantly improving the success rate and reliability of the design. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the deep learning-based multi-port coupling optimization method for optical waveguide chips according to an embodiment of the present invention. Figure 2 Build a logical flowchart for port matching and matrix construction. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the deep learning-based multi-port coupling optimization method for optical waveguide chips according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire port topology data and port optical field distribution data of the optical waveguide chip; Wavefront reconstruction and mode decomposition are performed on the port optical field distribution data to identify the fundamental mode component and higher-order mode component of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set. The set of port pairs is determined by waveguide path tracing. The difference between the fundamental mode coupling transmission efficiency and the energy vector of higher-order modes is calculated based on the port mode feature set. Mode matching coefficients are generated, and a matching relationship matrix is ​​constructed. From the matching relationship matrix, filter port pairs whose pattern matching coefficients meet the preset matching threshold, establish a candidate path set, calculate the distance penalty factor based on the port spacing, calculate the priority score in combination with the pattern matching coefficients, select a preset number of coupled paths, and obtain the coupled path configuration; The set of structural parameters is obtained by predicting the waveguide spacing and coupling region length of each coupling path in the coupling path configuration using a neural network. The target port coupling efficiency and non-target port crosstalk level are calculated based on the structural parameter set. When the target port coupling efficiency is lower than the preset efficiency threshold or the non-target port crosstalk level exceeds the preset crosstalk threshold, the matching threshold is adjusted and the path planning step is returned until the threshold requirements are met, and the structural parameter set is output.

[0021] In one specific implementation, when optimizing the multi-port coupling of an optical waveguide chip, port topology data and port optical field distribution data of the optical waveguide chip are acquired. The port topology data includes port spatial coordinates, relative positional relationships, and connection status. For example, the design may include 32 input / output ports with a coordinate accuracy of 0.01 micrometers, and the port positions cover a 10mm × 10mm chip area. The port optical field distribution data includes electromagnetic field intensity and phase distribution information, with a sampling resolution of 0.05 micrometers and approximately 100 × 100 sampling points per port.

[0022] After acquiring the data, wavefront reconstruction and mode decomposition are performed on the port optical field distribution data. Wavefront reconstruction reconstructs the complete electromagnetic field distribution at each port by analyzing the optical field intensity and phase data. The reconstruction process employs phase unwinding technology to recover continuous phase changes with an accuracy of 2π / 100. Subsequently, adaptive principal component analysis is used to extract the principal curvature direction of the phase distribution and the principal axis of symmetry of the amplitude distribution. In the phase curvature tensor calculation, the second-order partial derivative of the phase function is calculated, and the weighting coefficient is set as a proportional function of the phase gradient amplitude, with the coefficient ranging from 0 to 1. When calculating the symmetry measurement of the amplitude distribution, a rotation invariance index is used, and the analysis is performed with a 1-degree step within the range of 0 to 180 degrees, with the angle error controlled within ±0.5 degrees. After determining the fundamental mode propagation direction, a mode decomposition coordinate system is established, and the eigenmode expansion of the optical field distribution is performed in this coordinate system. In singular value decomposition, the cumulative energy contribution rate threshold is set to 99%, which is usually satisfied by the first 3 to 5 singular values. For a standard single-mode waveguide, the fundamental mode typically accounts for over 95% of the total energy, while higher-order modes account for a smaller proportion. When extracting the phase difference information between the fundamental and higher-order modes, the calculation accuracy reaches 2π / 360°, and the energy distribution ratio accuracy reaches 0.1%. The resulting port mode feature set includes port location, propagation direction, fundamental mode energy proportion, higher-order mode energy distribution, and phase difference data.

[0023] Subsequently, the port pair set was determined through waveguide path tracing. In the port spatial distribution map, the propagation vector accuracy was ±0.01 radians, and the ray projection detection accuracy was 0.1 micrometers. The near-field coupling distance was set to twice the wavelength, typically 3.1 micrometers for a wavelength of 1.55 micrometers. For the determined port pairs, the fundamental mode phase information was extracted from the port mode feature set, the phase difference was calculated, and compared with the coherence condition. A phase difference within ±π / 4 was considered high coherence, corresponding to a fundamental mode coupling transmission efficiency exceeding 90%. The higher-order mode energy vector was represented by a 5-dimensional vector to represent the energy distribution of the first five modes, with an Euclidean distance less than 0.2 indicating high mode matching. The fundamental mode correlation index weight was set to 0.7, and the reciprocal weight of the higher-order mode energy difference index was set to 0.3. The mode matching coefficient was obtained through geometric averaging. The mode matching coefficient ranged from 0 to 1, with typical values ​​between 0.5 and 0.95. It was filled into the matching relationship matrix to construct an n×n matrix, where n is the total number of ports.

[0024] Port pairs whose mode matching coefficients meet a preset matching threshold are selected from the matching matrix; a typical preset threshold is 0.75. Port pairs meeting the threshold are marked as candidate paths, and a candidate path set is established. The three-dimensional linear distance between ports within the candidate paths is calculated. The dielectric absorption coefficient is determined based on the waveguide material properties; for example, a typical value for silicon-based waveguides is 0.5 dB / cm. The transmission loss value is converted to a distance penalty factor using an exponential function with a base of 2.71828. In the priority scoring calculation, the weighting ratio of the mode matching coefficient to the reciprocal of the distance penalty factor is 7:3, which can be adjusted according to the application scenario. After sorting by priority score in descending order, the highest-scoring coupling path is selected; the number of paths selected depends on design requirements, typically ranging from 1 / 4 to 1 / 3 of the total number of ports.

[0025] This study uses a neural network to predict the waveguide spacing and coupling region length of each coupling path in a coupled path configuration. The neural network employs a five-layer feedforward structure. The input layer contains port pattern features and relative position parameters, while the hidden layer contains 128, 256, 128, and 64 neurons respectively. The output layer outputs the waveguide spacing and coupling region length parameters. The training sample size is 10,000 sets, derived from electromagnetic field simulation and experimental measurement data. The waveguide spacing prediction accuracy reaches ±0.02 micrometers, and the coupling region length prediction accuracy reaches ±1 micrometer. The prediction results form a structural parameter set containing the specific design parameters for all coupling paths.

[0026] The target port coupling efficiency and non-target port crosstalk level are calculated based on the structural parameter set. The preset threshold for target port coupling efficiency is 85%, and the preset threshold for non-target port crosstalk level is -20dB. If the threshold requirements are not met, the matching threshold is adjusted and the process returns to the path planning step. Typically, the initial matching threshold is set to 0.75, and if the requirements are not met, it is adjusted by 0.05 each time, ranging from 0.65 to 0.85. After 2 to 5 iterations, a structural parameter set that meets the requirements can usually be obtained, including the waveguide spacing (typically 0.2 to 1.0 μm) and coupling region length (typically 10 to 500 μm) for each coupling path.

[0027] In practical applications, when dealing with optical waveguide chips with 64 ports, the 20 optimal coupling paths determined by the above method achieve an average coupling efficiency of 90.2% and a crosstalk level of -23.5dB, which fully meets the design requirements of high-density optical waveguide chips.

[0028] In one optional embodiment, wavefront reconstruction and mode decomposition are performed on the port optical field distribution data to identify the fundamental mode component and higher-order mode components of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set including: Wavefront reconstruction is performed on the optical field distribution data of the ports to obtain the phase distribution and amplitude distribution of each port. The principal curvature direction of the phase distribution and the principal symmetry axis of the amplitude distribution are extracted by the adaptive principal component analysis method. The propagation direction of the fundamental mode component is determined based on the consistency between the principal curvature direction and the principal symmetry axis. A mode decomposition coordinate system is established with the propagation direction of the fundamental mode component as the axis. The eigenmodes of the light field distribution are expanded in the mode decomposition coordinate system. The expansion weights of each mode component are obtained by the singular value decomposition method. The fundamental mode component and each higher-order mode component are identified based on the numerical distribution of the expansion weights. The phase information of the fundamental mode component and each higher-order mode component during propagation is extracted. The difference between the phase information of the fundamental mode component and the phase information of each higher-order mode component is calculated as the inter-mode phase difference. The proportion of the square of the unfolded weight of each mode component to the total sum of squares of the weights is calculated to determine the energy distribution ratio. For each port, the phase difference and energy distribution ratio are bound to the spatial location information and propagation direction information of that port, respectively, to form a port mode feature set containing mode coupling path characteristics.

[0029] In one specific implementation, wavefront reconstruction and mode decomposition of the port optical field distribution data are key steps. The port optical field distribution data typically contains amplitude and phase information, formatted as a two-dimensional or three-dimensional array, with a resolution typically of 0.05 micrometers and a sampling number of 128 × 128 points per port. The wavefront reconstruction process utilizes the electric field amplitude and phase information to reconstruct the complete complex electric field distribution. Specifically, a phase unwinding algorithm is used to convert the phase folded within the range of -π to π into a continuously varying phase distribution. The unwinded phase distribution exhibits distinct propagation characteristics, with the phase gradient pointing in the direction of waveguide mode propagation.

[0030] After wavefront reconstruction, the principal curvature directions of the phase distribution are extracted using adaptive principal component analysis. The second-order partial derivatives of the phase distribution are calculated to obtain the curvature tensor for each spatial point. The curvature tensor contains the values ​​of the second-order partial derivatives in the x, y, and xy intersection directions. Spatial statistics are performed on the curvature tensor of the phase distribution to construct a weighted covariance matrix. The weighting coefficients are set as an exponential function of the phase gradient magnitude; the larger the magnitude, the higher the weight. In a typical implementation, the weighting coefficients are set as the square of the phase gradient magnitude divided by a normalization factor. Eigenvalue decomposition is performed on the weighted covariance matrix, and the eigenvector corresponding to the largest eigenvalue is extracted as the principal curvature direction. The principal curvature direction is usually represented as an angle with an accuracy of 0.1 degrees.

[0031] Simultaneously, the principal symmetry axis of the amplitude distribution is extracted, and symmetry metrics are calculated at different angles. The symmetry metric is obtained by calculating the mirror symmetry error of the amplitude distribution along a specific direction; the smaller the error, the higher the symmetry. The symmetry metric is calculated by rotating the amplitude distribution within a range of 0 to 180 degrees at 0.5-degree intervals, and the angle with the smallest symmetry metric is identified as the principal symmetry axis of the amplitude distribution. For an ideal fundamental mode distribution, the angle between the principal symmetry axis direction and the principal curvature direction should be less than 5 degrees; the smaller the angle, the higher the mode purity.

[0032] The propagation direction of the fundamental mode component is determined by comparing the consistency between the principal curvature direction and the principal axis of symmetry. The consistency criterion is that the angle between the two directions is less than a threshold (usually set to 10 degrees). When the angle is less than the threshold, the angle bisector between the principal curvature direction and the principal axis of symmetry is taken as the propagation direction of the fundamental mode component; when the angle is greater than the threshold, it indicates that the port optical field contains a complex mode structure. In this case, the principal curvature direction is used as the reference, and the port is marked as having a high content of higher-order modes. In practical applications, a typical propagation direction accuracy of ±1 degree is sufficient to meet subsequent processing requirements.

[0033] A mode decomposition coordinate system is established with the propagation direction of the fundamental mode component as the axis. In this system, the x-axis is perpendicular to the propagation direction and along the waveguide cross-section, the y-axis is perpendicular to the waveguide plane, and the z-axis is aligned with the propagation direction. The original optical field distribution is mapped to the mode decomposition coordinate system via coordinate transformation, achieved using a three-dimensional rotation matrix with a rotation accuracy of 0.01 radians. Under this system, the optical field distribution exhibits distinct mode characteristics: the fundamental mode typically displays a symmetrical distribution, while higher-order modes exhibit complex nodal structures.

[0034] The transformed light field distribution is constructed as a light field matrix, typically with dimensions m×n, where m and n represent the number of sampling points along the x and y directions, respectively. Singular value decomposition (SVD) is performed on the light field matrix to obtain a sequence of singular values ​​and a corresponding set of singular vectors. SVD is implemented using an iterative algorithm, typically with 100 iterations and a convergence threshold of 1e-6. The cumulative energy contribution rate of the singular value sequence is calculated by dividing the square of each singular value by the sum of the squares of all singular values, and then summing them sequentially. The minimum number of singular values ​​required to achieve a preset contribution rate threshold is determined as the number of effective mode components. The preset contribution rate threshold is typically set to 99%, corresponding to 3 to 7 effective mode components in practical applications.

[0035] The singular vector set is sorted according to the magnitude of the singular values. The singular vector corresponding to the largest singular value is identified as the fundamental mode component. Other singular vectors within the range of the number of effective mode components after sorting are identified as higher-order mode components in descending order of singular values. The singular values ​​corresponding to each mode component are extracted as expansion weights. Typically, the singular value of the fundamental mode component is about 2 to 5 times that of the second-highest-order mode, indicating that the fundamental mode dominates the energy distribution.

[0036] Phase information of the fundamental mode component and higher-order mode components during propagation is extracted. This phase information is obtained by extracting the phase of the complex amplitude of each mode component at different propagation distances. The difference between the phase information of the fundamental mode component and the phase information of each higher-order mode component is calculated as the inter-mode phase difference. Phase difference calculation employs phase difference normalization, mapping the phase difference to the range of -π to π. The proportion of the square of the expanded weight of each mode component to the sum of the squares of the total weights is calculated to determine the energy distribution ratio. The energy distribution ratio is expressed as a percentage value with an accuracy of 0.1%.

[0037] For a practical example of a 32-port waveguide chip, after wavefront reconstruction, the optical field distribution at a certain input port has a principal curvature direction of 37.2 degrees for phase distribution and a principal symmetry axis of 36.8 degrees for amplitude distribution, with an angle of 0.4 degrees between them, determining the propagation direction to be 37.0 degrees. In the established mode decomposition coordinate system, singular value decomposition of the optical field matrix yields a singular value sequence [8.76, 3.21, 0.95, 0.42, 0.18], with corresponding cumulative energy contribution rates of [0.841, 0.954, 0.991, 0.997, 1.000]. The number of effective mode components is determined to be 3. The singular vector corresponding to the maximum singular value 8.76 is identified as the fundamental mode component, while the singular vectors corresponding to singular values ​​3.21 and 0.95 are the first-order and second-order higher-order mode components, respectively. The phase difference between the fundamental mode and the first-order higher-order mode is calculated to be 0.72π, and the phase difference between the fundamental mode and the second-order higher-order mode is 1.38π. The energy distribution ratios of each mode component are [84.1%, 11.3%, 3.7%].

[0038] For all 32 ports on the chip, the phase difference and energy distribution ratio are bound to the spatial location and propagation direction information of each port, respectively, to form a port mode feature set. The port spatial location is represented by three-dimensional coordinates (x, y, z) with an accuracy of 0.01 micrometers; the propagation direction is represented by an angle with an accuracy of 0.1 degrees. The final port mode feature set is organized as structured data, including port identifier, spatial location, propagation direction, fundamental mode energy ratio, higher-order mode energy distribution, and phase difference data, providing accurate mode matching basis for subsequent coupling path planning.

[0039] In an optional embodiment, extracting the principal curvature direction of the phase distribution and the principal symmetry axis of the amplitude distribution using an adaptive principal component analysis method includes: The phase distribution is subjected to second-order partial derivative calculation to obtain the phase curvature tensor. Based on the spatial distribution characteristics of the phase curvature tensor, a weighted covariance matrix is ​​constructed, where the weight coefficients are determined by the phase gradient magnitude. The weighted covariance matrix is ​​subjected to eigenvalue decomposition, and the eigenvector corresponding to the largest eigenvalue is extracted as the principal curvature direction. The amplitude distribution is measured and symmetry is calculated. The direction of the principal axis of symmetry is identified by rotation invariance test, and consistency is determined by combining the principal curvature direction.

[0040] In one specific implementation, a key step in the multi-port coupling optimization method for optical waveguide chips is calculating the second-order partial derivatives of the phase distribution to obtain the phase curvature tensor. The phase data is typically represented as a two-dimensional array of size M×N, with a resolution of 0.05 micrometers, and a sampling range covering a region of 10 micrometers × 10 micrometers. The second-order partial derivatives are calculated using the finite difference method, calculating the second-order differences of the phase data in the horizontal, vertical, and cross directions respectively. Specifically, for each sampling point (i, j), the second-order partial derivative in the horizontal direction is calculated using the phase values ​​of three adjacent points; the second-order partial derivative in the vertical direction is calculated similarly; and the second-order partial derivative in the cross direction is calculated using the phase values ​​of four diagonal points. The calculation accuracy is related to the sampling interval, with a typical accuracy reaching 0.01 radians per square micrometer.

[0041] After calculating the second-order partial derivatives, the resulting phase curvature tensor contains three components, corresponding to the horizontal curvature, vertical curvature, and shear curvature, respectively. For a typical single-mode waveguide port, the phase curvature tensor has a larger value in the central region of the waveguide, gradually decreasing outwards, forming a Gaussian-like distribution. Based on the spatial distribution characteristics of the phase curvature tensor, a 2×2-dimensional weighted covariance matrix is ​​constructed. The matrix elements are obtained by weighted statistics of each component of the phase curvature tensor over the entire analysis region. The weighting coefficients are determined by the phase gradient magnitude. Specifically, the first-order difference of the phase data in the horizontal and vertical directions is first calculated, and the square root of the sum of the squares of the first-order differences in the two directions is taken as the phase gradient magnitude. For the measured data of a 32-channel waveguide chip, the phase gradient magnitude typically ranges from 0.1 to 10 radians per micrometer, with the largest gradient magnitude in the central region and gradually decreasing at the edges.

[0042] The weighting coefficients are designed using a non-linear mapping to convert the phase gradient magnitude into weight values ​​within the range of 0 to 1. The mapping function is set to divide the phase gradient magnitude by a reference value and then take the exponent. The reference value is typically set as the average of the phase gradient magnitudes multiplied by an adjustment factor of 1.5. The exponent is set to 2, making the weight of high-gradient regions significantly higher than that of low-gradient regions, thus better highlighting the propagation direction characteristics. In practical applications, a gradient magnitude of 5 radians per micrometer corresponds to a weight of approximately 0.9; a gradient magnitude of 1 radian per micrometer corresponds to a weight of approximately 0.3.

[0043] When constructing the weighted covariance matrix, the three components of the phase curvature tensor are summed using weights and then used to fill the four element positions of the covariance matrix. The weighted sum of the horizontal curvature is filled in the upper left corner of the matrix, the weighted sum of the vertical curvature in the lower right corner, and the weighted sum of the shear curvature in the other two diagonal elements. After filling, the covariance matrix exhibits symmetry, reflecting the main direction of phase distribution change. For single-mode waveguides, the covariance matrix typically has one eigenvalue that is significantly larger than the others, indicating a significant change in phase distribution along a specific direction.

[0044] Eigenvalue decomposition is performed on the weighted covariance matrix using the Jacobi iterative method, with an upper limit of 100 iterations and a convergence threshold of 0.000001. The decomposition yields two eigenvalues ​​and their corresponding eigenvectors. The eigenvalues ​​represent the degree of change in phase curvature along the corresponding eigenvector direction, and the eigenvectors represent the principal direction of curvature change. The eigenvector corresponding to the largest eigenvalue is extracted as the principal curvature direction. This eigenvector is a two-dimensional unit vector, and its corresponding angle is calculated using the arctangent function, ranging from 0 to 180 degrees with an accuracy of 0.1 degrees.

[0045] In a typical application scenario of optical waveguide chips, when processing a square waveguide port with a core size of 8 μm × 8 μm and a working wavelength of 1550 nm, the calculated second-order partial derivatives of the phase distribution show that the average curvature in the horizontal direction is 2.3 radians per square micrometer, the average curvature in the vertical direction is 2.1 radians per square micrometer, and the average shear curvature is 1.4 radians per square micrometer. The maximum phase gradient amplitude reaches 7.8 radians per micrometer, and the average value is 3.2 radians per micrometer. The constructed weighted covariance matrix is ​​[[0.76, 0.42], [0.42, 0.68]], and the eigenvalue decomposition yields eigenvalues ​​of 1.21 and 0.23, corresponding to a principal curvature direction of 37.2 degrees.

[0046] Symmetry measurement of the amplitude distribution is performed, and rotational invariance is used to identify the direction of the principal axis of symmetry. The amplitude distribution typically exhibits a Gaussian or Gaussian-like pattern, with the highest intensity at the center and gradually decreasing outwards. The symmetry measurement first normalizes the amplitude distribution to a maximum value of 1. Then, a rotational scan is performed within a range of 0 to 180 degrees in 1-degree increments, calculating the symmetry error along both sides of the rotation axis at each angle. The symmetry error is defined as the integral of the absolute value of the amplitude difference at the symmetrical position over the entire region. For an ideal symmetrical distribution, the symmetry error in a specific direction should be 0; in practical applications, due to measurement noise and mode impurities, the symmetry error is always greater than 0.

[0047] During the rotational invariance test, the symmetry error value at each rotation angle is recorded, forming a symmetry error curve. The symmetry error curve typically exhibits one or more local minima, corresponding to the potential axis of symmetry direction. The global minimum of the symmetry error is precisely located using polynomial interpolation, thus determining the direction of the principal axis of symmetry. The typical range of the minimum symmetry error is 0.05 to 0.2; a smaller value indicates better symmetry. The determination accuracy of the principal axis of symmetry direction reaches 0.5 degrees, meeting the requirements for subsequent processing.

[0048] For the aforementioned waveguide port, after symmetry measurement, the minimum symmetry error curve of the amplitude distribution is 0.083, occurring at an angle of 36.8 degrees, which is determined to be the principal symmetry axis direction. Compared with the principal curvature direction of the phase distribution at 37.2 degrees, the angle between the two is 0.4 degrees, which is less than the preset threshold of 5 degrees, and is therefore judged to be highly consistent. The consistency determination is achieved by calculating the angle between the principal curvature direction and the principal symmetry axis direction. An angle less than the threshold indicates that the two characteristic directions are consistent, while an angle greater than the threshold indicates the presence of mode aliasing or waveguide defects.

[0049] In the multi-port waveguide chip, the above analysis was performed on all 32 ports. It was found that the angle between the principal curvature direction and the principal symmetry axis direction was less than 5 degrees for 28 ports, indicating good mode purity. Four ports had an angle greater than 5 degrees, reaching a maximum of 12.7 degrees. These ports are marked as requiring special optimization and will receive more attention in subsequent coupling optimization. The statistical results of the consistency between the principal curvature direction and the principal symmetry axis serve as an important indicator for port mode quality assessment, directly affecting coupling path planning and parameter optimization strategies.

[0050] The entire analysis process, processing 32-port data on a standard computing platform, took no more than 2 seconds, ensuring the real-time performance and efficiency of the optimization method. By accurately extracting the principal curvature direction of the phase distribution and the principal symmetry axis of the amplitude distribution, a solid foundation was laid for subsequent mode decomposition and coupling path planning, significantly improving the accuracy and reliability of multi-port coupling optimization.

[0051] In one optional embodiment, the expanded weights of each mode component are obtained through singular value decomposition, and the fundamental mode component and higher-order mode components are identified based on the numerical distribution of the expanded weights, including: In the mode decomposition coordinate system, the light field distribution is constructed as a light field matrix. Singular value decomposition is performed on the light field matrix to obtain a sequence of singular values ​​and a set of corresponding singular vectors. The cumulative energy contribution rate of the singular value sequence is calculated, the minimum number of singular values ​​required to make the cumulative energy contribution rate reach a preset contribution rate threshold is determined, and the number of effective mode components is determined. The singular vector set is sorted according to the magnitude of the singular values. The singular vector corresponding to the largest singular value is identified as the fundamental mode component. Other singular vectors within the range of the number of effective mode components after sorting are identified as higher-order mode components in descending order of singular values. The singular values ​​corresponding to each mode component are extracted as expansion weights.

[0052] In one specific implementation, after acquiring the port optical field distribution data, the optical field needs to be represented in a matrix form within the established mode decomposition coordinate system. The optical field intensity distribution at the port cross-section is arranged into a two-dimensional optical field matrix according to the spatial sampling point positions. The rows and columns of this matrix correspond to the horizontal and vertical spatial coordinates, respectively. Singular value decomposition (SVD) is performed on the optical field matrix. This decomposition represents the matrix as a product of three matrices, where the middle diagonal matrix contains the singular values ​​arranged in descending order, and the column vectors of the left and right matrices form a set of singular vectors. The square of each singular value represents the energy carried by the corresponding mode component.

[0053] When calculating the cumulative energy contribution rate of the singular value sequence, the squares of each singular value are summed sequentially and then divided by the sum of the squares of all singular values. When the cumulative energy contribution rate first reaches or exceeds a preset contribution rate threshold, such as 95% or 98%, the corresponding singular value index is the number of effective mode components. This number determines the main mode components that need to be retained, eliminating noise modes with minimal energy contribution. The number of effective mode components is typically between 3 and 8, depending on the complexity of the port optical field and the mode support capability of the waveguide structure.

[0054] After determining the number of valid mode components, the singular vector set is sorted from largest to smallest based on the singular values. The singular vector corresponding to the largest singular value exhibits a Gaussian or near-Gaussian profile in spatial distribution, with energy concentrated in the central region of the port; this vector is identified as the fundamental mode component. The fundamental mode component carries the main energy of the optical field, and its expansion weight typically accounts for more than 70% of the total energy. Other singular vectors within the range of valid mode components after sorting are identified as first-order, second-order, third-order, and other higher-order mode components in descending order of singular value. Higher-order mode components exhibit a multi-peak structure or eccentric distribution in spatial distribution, with their expansion weights decreasing sequentially.

[0055] The singular values ​​corresponding to each mode component are extracted as the expansion weight values. The numerical distribution of the expansion weights directly reflects the mode purity of the optical field. When the expansion weight of the fundamental mode exceeds 85%, it indicates that the optical field at that port has good single-mode characteristics; when the expansion weight values ​​of multiple higher-order modes are close, it indicates the existence of mode mixing. The differences in the numerical distribution of expansion weights for different ports constitute the basic data for subsequent matching relationship calculations. By comparing the expansion weight vectors of two ports, the degree of mode matching can be quantitatively evaluated, providing a basis for prioritizing coupling paths.

[0056] In practice, the spatial sampling density of the light field matrix affects the mode decomposition accuracy. The sampling point spacing is set to 1 / 10 to 1 / 5 of the wavelength to ensure sufficient spatial resolution. For rectangular ports, the light field matrix is ​​constructed as a square matrix to facilitate singular value decomposition (SVD) calculations; for irregular ports, interpolation methods are used to map the light field data to a regular grid. SVD calculations are implemented using standard algorithms from numerical algebra libraries. The computation time increases linearly with the matrix dimension; for a typical 128×128 sampling matrix, a single decomposition takes milliseconds.

[0057] In one optional embodiment, the port pair set is determined by waveguide path tracing, the difference between the fundamental mode coupling transmission efficiency and the higher-order mode energy vector is calculated based on the port mode feature set, mode matching coefficients are generated, and a matching relationship matrix is ​​constructed, including: Analyze the port topology data, construct a port spatial distribution map, perform waveguide path tracing on the port spatial distribution map, identify port pairs with direct waveguide connections or port pairs with overlapping near-field coupling regions, and form a set of port pairs; The fundamental mode component phase information of the port pair is extracted from the port mode feature set. The difference between the fundamental mode component phase information of the port pair is calculated. The difference between the phase information is compared with the preset coherence conditions to determine the fundamental mode coupling transmission efficiency of the port pair. Combined with the energy distribution ratio of the fundamental mode component of the port pair, the fundamental mode correlation index of the port pair is generated. Extract the energy distribution ratio of higher-order mode components of port pairs from the port mode feature set, construct the higher-order mode energy vector of port pairs, and calculate the Euclidean distance between the higher-order mode energy vectors as the higher-order mode energy difference index of the port pairs. The fundamental mode correlation index is used as the coupling strength weight, and the reciprocal of the higher-order mode energy difference index is used as the mode purity weight. The geometric mean is then used to generate the mode matching coefficient. Fill the pattern matching coefficients of each port pair into the corresponding row and column positions to construct the matching relationship matrix.

[0058] In one specific implementation, during the port pair identification and matching evaluation stage of the optical waveguide chip, a comprehensive analysis of the chip's internal physical connections and optical transmission characteristics is required. The analysis step begins with port topology data, which includes the coordinates of each port (x, y, y). i y i ), port orientation angle θ i and port width w i Geometric parameters are defined. By constructing a port spatial distribution map, all ports are mapped to a two-dimensional plane coordinate system, forming a visual representation of the port positional relationships.

[0059] The waveguide path tracing process employs a simplified form of the ray tracing algorithm. Virtual rays are extended along the normal direction from each port, and the intersections of these rays with other port regions are detected. When there is a direct physical waveguide connection between two ports, the tracing line segment will be found in the waveguide structure data with the corresponding waveguide number. For port pairs that are not directly connected but have overlapping near-field coupling regions, the minimum spacing between the port edges is calculated. If the spacing is less than 5 times the wavelength and the optical field distributions of the two ports spatially overlap, they are considered to have the potential for near-field coupling. Using these two criteria, all port pairs meeting the physical coupling conditions are selected, forming the initial set of port pairs.

[0060] For each pair of ports in the set, the normalized electric field distribution and phase information of its fundamental mode component are extracted from the port mode feature set. Fundamental mode phase difference. Obtained by direct subtraction of phase information, with preset coherence requirements. Efficient coupling is considered to occur when the frequency falls within the interval [-π / 4, π / 4]. The fundamental mode coupling transmission efficiency is estimated using a simplified form of the mode overlap integral, which involves point-by-point multiplication and summation of the normalized amplitudes of the fundamental mode electric field distribution at both ports, combined with a phase difference correction factor. The coupling efficiency estimate is obtained. Combining the proportions r1 and r2 of the fundamental mode components of each port to the total energy, the calculation... A fundamental mode correlation index is generated, which comprehensively reflects the fundamental mode energy concentration and phase matching degree.

[0061] Energy distribution analysis of higher-order modes requires the construction of a multidimensional energy vector. The energy percentage of each higher-order mode (e.g., TE1, TE2, etc.) is extracted from the port mode feature set and arranged by mode number to form an energy vector v. i = [e1, e2, ..., e n ], where e k Let represent the normalized energy of the k-th order mode. Calculate the Euclidean distance between the two energy vectors of the port pair. This distance value quantifies the degree of difference in the distribution of higher-order modes. A smaller distance value indicates that the distributions of higher-order modes at both ports are similar, which is beneficial for reducing mode conversion losses. This distance value is used as an indicator of the energy difference of higher-order modes.

[0062] The mode matching coefficients are generated using a geometric mean strategy to ensure a balance between fundamental mode coupling performance and mode purity. The fundamental mode correlation index is directly used as the coupling strength weight w. c The reciprocal of the energy difference index of higher-order models, after normalization, is used as the model purity weight w. p The formula for calculating the geometric mean is: The value of this coefficient ranges from 0 to 1. The closer the value is to 1, the better the coupling matching condition of the port pair.

[0063] The matching matrix is ​​constructed in port number order, with dimensions M×M, where M is the total number of ports. For each port pair (i, j) in the port pair set, the calculated pattern matching coefficient γ is... ij Fill the matrix up to the i-th row and j-th column. Due to the symmetry of the coupling relationship, fill the j-th row and i-th column with the same value. For port combinations not in the port pair set, fill the corresponding matrix position with zero value, indicating that the port pair does not have physical coupling conditions. The matching relationship matrix formed after filling is a symmetric matrix, and the distribution of non-zero elements in the matrix reflects the feasible coupling path network topology inside the chip.

[0064] like Figure 2 The diagram shows the logical flowchart for port matching and matrix construction.

[0065] In one optional embodiment, port topology data is parsed to construct a port spatial distribution map. Waveguide path tracing is performed on the port spatial distribution map to identify port pairs with direct waveguide connections or port pairs with overlapping near-field coupling regions, forming a port pair set including: The port topology data is analyzed to extract the spatial coordinates and propagation direction information of each port. The spatial coordinates are mapped to a spatial coordinate system, and the propagation vectors of each port are marked according to the propagation direction information to construct a port spatial distribution map. In the port spatial distribution diagram, starting from the spatial coordinates of each port, a ray is projected along the direction of the propagation vector to detect the intersection of the ray with the propagation vectors of other ports, and the ports with intersections are identified as port pairs with direct waveguide connections. In the port spatial distribution map, the range of the near-field coupling region is determined based on the near-field coupling distance, with the spatial coordinates of each port as the center. It is then detected whether the near-field coupling regions of different ports have spatial overlap, and ports with overlapping near-field coupling regions are identified as port pairs with overlapping near-field coupling regions. The port pairs with direct waveguide connections are combined with the port pairs that overlap with the near-field coupling region to form a port pair set.

[0066] In one specific implementation, after acquiring the port topology data of the optical waveguide chip, the three-dimensional spatial coordinates (x, y, z) of each port are extracted from the data. i y i , z i ), where i represents the port number. Simultaneously, the propagation direction information for each port is extracted, expressed as a unit vector. This indicates that its modulus is 1, pointing to the main propagation direction of the light wave at that port. The extracted spatial coordinates are mapped to a unified Cartesian coordinate system, with a fixed reference point on the chip as the origin, ensuring that all port positions are expressed within the same coordinate framework. The positions of each port are marked in the coordinate system, and a corresponding propagation vector arrow is drawn at each port position. The length of the arrow can be scaled according to visualization needs, but the direction must strictly remain consistent with the propagation vector, thus forming a complete port spatial distribution map.

[0067] After constructing the port spatial distribution map, a ray casting operation is performed for each port. A port is selected as the starting port, and a semi-infinite ray is emitted along its propagation vector direction, using the spatial coordinates of that port as the ray's origin. Mathematically, this ray can be represented by parametric equations. Where t is a non-negative real number parameter. Traverse all other ports in the port spatial distribution map, calculating the spatial relationship between each ray and the lines containing the propagation vectors of other ports. When two lines intersect in three-dimensional space or the distance is less than a preset waveguide width threshold, they are determined to have an intersection point. Record the port pairs that produce the intersection point and mark them as having a direct waveguide connection. Perform this ray projection process sequentially on all ports in the port spatial distribution map to obtain a list of all port pairs with direct waveguide connections.

[0068] Based on identifying direct waveguide connections, port pairs interacting through near-field coupling are further identified. For each port, the effective range of near-field coupling is calculated according to the refractive index of the waveguide material, the operating wavelength, and the waveguide geometry. This range is typically the lateral distance corresponding to the waveguide evanescent field attenuating to 1 / e intensity, with typical values ​​ranging from several micrometers to tens of micrometers. A circular or elliptical region is defined in a plane perpendicular to the propagation direction, centered on the spatial coordinates of each port. The radius or semi-axial length of this region is equal to the near-field coupling distance, constituting the near-field coupling region of that port. For all ports in the chip, their near-field coupling regions are calculated sequentially, and then region overlap detection is performed. During detection, the spatial distance between any two ports is calculated. If the distance is less than the sum of their near-field coupling distances, the near-field coupling regions of the two ports are determined to overlap, and this port pair is recorded as a port pair with overlapping near-field coupling regions.

[0069] The previously identified list of port pairs with direct waveguide connections is merged with the list of port pairs whose near-field coupling regions overlap. During the merging process, duplicate port pairs are removed from both lists; if a port pair exists in both lists, only one record is retained. The merged list constitutes the port pair set, encompassing all possible port combinations on the chip that interact with the optical field via direct waveguide connection or near-field coupling. Each element in the port pair set is stored as a binary tuple, recording the numbers of the two ports involved in the coupling. This port pair set provides a clear scope for subsequent calculations of mode matching coefficients and the construction of the matching relationship matrix, avoiding redundant computational overhead caused by performing fully connected calculations on all ports.

[0070] In one optional embodiment, port pairs whose pattern matching coefficients satisfy a preset matching threshold are filtered from the matching relationship matrix to establish a candidate path set. A distance penalty factor is calculated based on the port spacing, and a priority score is calculated in combination with the pattern matching coefficients. A preset number of coupled paths are selected to obtain the coupled path configuration, which includes: Traverse the matching relationship matrix, extract matrix elements whose pattern matching coefficient is greater than a preset matching threshold, record the row and column indices corresponding to the matrix elements, extract the corresponding port pairs from the port pair set according to the row and column indices, mark them as candidate paths, and establish a candidate path set. The spatial coordinates of each candidate path in the candidate path set are extracted from the port topology data. The three-dimensional straight-line distance between ports in each candidate path is calculated as the port spacing. Based on the attenuation characteristics of electromagnetic waves in the transmission medium, the port spacing is multiplied by the medium absorption coefficient to obtain the transmission loss value. The transmission loss value is exponentially transformed to obtain the distance penalty factor. The mode matching coefficients corresponding to each candidate path are extracted. The mode matching coefficients and the reciprocal of the distance penalty factor are weighted and geometrically averaged to calculate the priority score of each candidate path. The priority scores of all candidate paths in the candidate path set are sorted in descending order, and a preset number of candidate paths with the highest priority scores are extracted as coupling paths to obtain the coupling path configuration.

[0071] In one specific implementation, during the path planning phase, the matching relationship matrix is ​​traversed, and the value of each element in the matrix is ​​compared with a preset matching threshold. The preset matching threshold is determined based on the application scenario of the optical waveguide chip. For high-precision optical communication chips, the threshold is set between 0.85 and 0.92; for sensor array chips, the threshold can be appropriately relaxed to 0.75 to 0.85. When the value of a matrix element is greater than the threshold, the row index i and column index j of the element in the matrix are recorded. This index pair (i, j) directly corresponds to the i-th input port and the j-th output port in the port pair set. All port pairs that meet the threshold condition are extracted, and a unique candidate path identifier is assigned to each port pair. These identifiers and their corresponding port numbers are stored in the data structure of the candidate path set.

[0072] For each candidate path in the candidate path set, the three-dimensional spatial coordinates of the starting and ending ports are read from the port topology data. The starting port coordinates are denoted as (x1, y1, z1), and the ending port coordinates are denoted as (x2, y2, z2). The port spacing d is obtained by calculating the Euclidean distance between the two points. The waveguide chip typically uses silicon-based or silicon nitride materials as the transmission medium, with absorption coefficients α at the operating wavelength of 1550 nm of approximately 0.2 dB / cm and 0.05 dB / cm, respectively. The transmission loss L is calculated by multiplying the port spacing by the absorption coefficient of the medium, in dB. Distance penalty factor P... d It is expressed in the form of an exponential function and is calculated as an exponential function value with the natural constant as the base and the negative transmission loss value as the exponent. This transformation converts the linear loss into a weighting coefficient between 0 and 1, and the penalty factor becomes smaller as the distance increases.

[0073] Obtain the pattern matching coefficient M for each candidate path in the matching relationship matrix. c Then, a weighted geometric mean is calculated using this coefficient and the reciprocal of the distance penalty factor. Specifically, the weight for pattern matching is set to w. mThe weight of the distance factor is w. d The sum of the two weights is 1. The priority score S is calculated as the sum of the pattern matching coefficients and their respective weights. m The power and the reciprocal of the distance penalty factor w d The product of powers. In practical applications, for chip regions with high port density, w m Set to 0.6, w d Setting it to 0.4 emphasizes the importance of pattern matching; for regions with sparse port distribution, w m Set to 0.4, w d The value is set to 0.6 to avoid excessive transmission loss caused by long paths.

[0074] After calculating the priority scores of all paths in the candidate path set, a quicksort algorithm is used to sort the scores in descending order. Starting from the highest digit of the sorted result, the top N candidate paths are extracted sequentially, where N is a preset number parameter, the value of which is determined based on the actual total number of ports and coupling requirements of the chip. For an optical switch chip with 32 ports, the preset number is usually set to 12 to 16 paths; for a wavelength division multiplexing chip with 64 ports, this number is set to 24 to 32 paths. These extracted N paths are the final coupling paths. The start and end port numbers, priority scores, port spacing, mode matching coefficients, and other parameters of each path are encapsulated into a coupling path configuration data structure. This configuration serves as the input for subsequent neural network prediction of waveguide spacing and coupling region length.

[0075] The multi-port coupling optimization system for optical waveguide chips based on deep learning, as described in this embodiment of the invention, includes: The data acquisition unit is used to acquire port topology data and port optical field distribution data of the optical waveguide chip; The mode analysis unit is used to perform wavefront reconstruction and mode decomposition on the port optical field distribution data, identify the fundamental mode component and higher-order mode component of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set. The matching calculation unit is used to determine the set of port pairs by waveguide path tracing, calculate the difference between the fundamental mode coupling transmission efficiency and the energy vector of higher-order modes based on the port mode feature set, generate mode matching coefficients, and construct a matching relationship matrix. The path filtering unit is used to filter port pairs whose pattern matching coefficients meet the preset matching threshold from the matching relationship matrix, establish a candidate path set, calculate the distance penalty factor based on the port spacing, calculate the priority score in combination with the pattern matching coefficient, select a preset number of coupled paths, and obtain the coupled path configuration. The parameter prediction unit is used to predict the waveguide spacing and coupling region length of each coupling path in the coupling path configuration through a neural network, thereby obtaining the structural parameter set. The optimization verification unit is used to calculate the coupling efficiency of the target port and the crosstalk level of the non-target port based on the structural parameter set. When the coupling efficiency of the target port is lower than the preset efficiency threshold or the crosstalk level of the non-target port exceeds the preset crosstalk threshold, the matching threshold is adjusted and the path planning step is returned until the threshold requirement is met, and the structural parameter set is output.

[0076] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0077] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0078] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based multi-port coupling optimization method for optical waveguide chips, characterized in that, include: Acquire port topology data and port optical field distribution data of the optical waveguide chip; Wavefront reconstruction and mode decomposition are performed on the port optical field distribution data to identify the fundamental mode component and higher-order mode component of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set. The set of port pairs is determined by waveguide path tracing. The difference between the fundamental mode coupling transmission efficiency and the energy vector of higher-order modes is calculated based on the port mode feature set. Mode matching coefficients are generated, and a matching relationship matrix is ​​constructed. From the matching relationship matrix, filter port pairs whose pattern matching coefficients meet the preset matching threshold, establish a candidate path set, calculate the distance penalty factor based on the port spacing, calculate the priority score in combination with the pattern matching coefficients, select a preset number of coupled paths, and obtain the coupled path configuration; The set of structural parameters is obtained by predicting the waveguide spacing and coupling region length of each coupling path in the coupling path configuration using a neural network. The target port coupling efficiency and non-target port crosstalk level are calculated based on the structural parameter set. When the target port coupling efficiency is lower than the preset efficiency threshold or the non-target port crosstalk level exceeds the preset crosstalk threshold, the matching threshold is adjusted and the path planning step is returned until the threshold requirements are met, and the structural parameter set is output.

2. The method according to claim 1, characterized in that, Wavefront reconstruction and mode decomposition are performed on the port optical field distribution data to identify the fundamental and higher-order mode components of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set including: Wavefront reconstruction is performed on the optical field distribution data of the ports to obtain the phase distribution and amplitude distribution of each port. The principal curvature direction of the phase distribution and the principal symmetry axis of the amplitude distribution are extracted by the adaptive principal component analysis method. The propagation direction of the fundamental mode component is determined based on the consistency between the principal curvature direction and the principal symmetry axis. A mode decomposition coordinate system is established with the propagation direction of the fundamental mode component as the axis. The eigenmodes of the light field distribution are expanded in the mode decomposition coordinate system. The expansion weights of each mode component are obtained by the singular value decomposition method. The fundamental mode component and each higher-order mode component are identified based on the numerical distribution of the expansion weights. The phase information of the fundamental mode component and each higher-order mode component during propagation is extracted. The difference between the phase information of the fundamental mode component and the phase information of each higher-order mode component is calculated as the inter-mode phase difference. The proportion of the square of the unfolded weight of each mode component to the total sum of squares of the weights is calculated to determine the energy distribution ratio. For each port, the phase difference and energy distribution ratio are bound to the spatial location information and propagation direction information of that port, respectively, to form a port mode feature set containing mode coupling path characteristics.

3. The method according to claim 2, characterized in that, The principal curvature direction of the phase distribution and the principal symmetry axis of the amplitude distribution are extracted using adaptive principal component analysis, including: The phase distribution is subjected to second-order partial derivative calculation to obtain the phase curvature tensor. Based on the spatial distribution characteristics of the phase curvature tensor, a weighted covariance matrix is ​​constructed, where the weight coefficients are determined by the phase gradient magnitude. The weighted covariance matrix is ​​subjected to eigenvalue decomposition, and the eigenvector corresponding to the largest eigenvalue is extracted as the principal curvature direction. The amplitude distribution is measured and symmetry is calculated. The direction of the principal axis of symmetry is identified by rotation invariance test, and consistency is determined by combining the principal curvature direction.

4. The method according to claim 2, characterized in that, The expansion weights of each mode component are obtained using singular value decomposition. Based on the numerical distribution of the expansion weights, the fundamental mode component and higher-order mode components are identified, including: In the mode decomposition coordinate system, the light field distribution is constructed as a light field matrix. Singular value decomposition is performed on the light field matrix to obtain a sequence of singular values ​​and a set of corresponding singular vectors. The cumulative energy contribution rate of the singular value sequence is calculated, the minimum number of singular values ​​required to make the cumulative energy contribution rate reach a preset contribution rate threshold is determined, and the number of effective mode components is determined. The singular vector set is sorted according to the magnitude of the singular values. The singular vector corresponding to the largest singular value is identified as the fundamental mode component. Other singular vectors within the range of the number of effective mode components after sorting are identified as higher-order mode components in descending order of singular values. The singular values ​​corresponding to each mode component are extracted as expansion weights.

5. The method according to claim 1, characterized in that, Port pair sets are determined by waveguide path tracing. The difference between the fundamental mode coupling transmission efficiency and the higher-order mode energy vector is calculated based on the port mode feature set. Mode matching coefficients are generated, and a matching relationship matrix is ​​constructed, including: Analyze the port topology data, construct a port spatial distribution map, perform waveguide path tracing on the port spatial distribution map, identify port pairs with direct waveguide connections or port pairs with overlapping near-field coupling regions, and form a set of port pairs; The fundamental mode component phase information of the port pair is extracted from the port mode feature set. The difference between the fundamental mode component phase information of the port pair is calculated. The difference between the phase information is compared with the preset coherence conditions to determine the fundamental mode coupling transmission efficiency of the port pair. Combined with the energy distribution ratio of the fundamental mode component of the port pair, the fundamental mode correlation index of the port pair is generated. Extract the energy distribution ratio of higher-order mode components of port pairs from the port mode feature set, construct the higher-order mode energy vector of port pairs, and calculate the Euclidean distance between the higher-order mode energy vectors as the higher-order mode energy difference index of the port pairs. The fundamental mode correlation index is used as the coupling strength weight, and the reciprocal of the higher-order mode energy difference index is used as the mode purity weight. The geometric mean is then used to generate the mode matching coefficient. Fill the pattern matching coefficients of each port pair into the corresponding row and column positions to construct the matching relationship matrix.

6. The method according to claim 5, characterized in that, Analyze the port topology data to construct a port spatial distribution map. Perform waveguide path tracing on the port spatial distribution map to identify port pairs with direct waveguide connections or port pairs with overlapping near-field coupling regions, forming a set of port pairs including: The port topology data is analyzed to extract the spatial coordinates and propagation direction information of each port. The spatial coordinates are mapped to a spatial coordinate system, and the propagation vectors of each port are marked according to the propagation direction information to construct a port spatial distribution map. In the port spatial distribution diagram, starting from the spatial coordinates of each port, a ray is projected along the direction of the propagation vector to detect the intersection of the ray with the propagation vectors of other ports, and the ports with intersections are identified as port pairs with direct waveguide connections. In the port spatial distribution map, the range of the near-field coupling region is determined based on the near-field coupling distance, with the spatial coordinates of each port as the center. It is then detected whether the near-field coupling regions of different ports have spatial overlap, and ports with overlapping near-field coupling regions are identified as port pairs with overlapping near-field coupling regions. The port pairs with direct waveguide connections are combined with the port pairs that overlap with the near-field coupling region to form a port pair set.

7. The method according to claim 1, characterized in that, From the matching matrix, port pairs whose pattern matching coefficients satisfy a preset matching threshold are selected to establish a candidate path set. A distance penalty factor is calculated based on the port spacing, and a priority score is calculated using the pattern matching coefficients. A preset number of coupled paths are then selected to obtain the coupled path configuration, which includes: Traverse the matching relationship matrix, extract matrix elements whose pattern matching coefficient is greater than a preset matching threshold, record the row and column indices corresponding to the matrix elements, extract the corresponding port pairs from the port pair set according to the row and column indices, mark them as candidate paths, and establish a candidate path set. The spatial coordinates of each candidate path in the candidate path set are extracted from the port topology data. The three-dimensional straight-line distance between ports in each candidate path is calculated as the port spacing. Based on the attenuation characteristics of electromagnetic waves in the transmission medium, the port spacing is multiplied by the medium absorption coefficient to obtain the transmission loss value. The transmission loss value is exponentially transformed to obtain the distance penalty factor. The mode matching coefficients corresponding to each candidate path are extracted. The mode matching coefficients and the reciprocal of the distance penalty factor are weighted and geometrically averaged to calculate the priority score of each candidate path. The priority scores of all candidate paths in the candidate path set are sorted in descending order, and a preset number of candidate paths with the highest priority scores are extracted as coupling paths to obtain the coupling path configuration.

8. A deep learning-based multi-port coupling optimization system for optical waveguide chips, used to implement the method of any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire port topology data and port optical field distribution data of the optical waveguide chip; The mode analysis unit is used to perform wavefront reconstruction and mode decomposition on the port optical field distribution data, identify the fundamental mode component and higher-order mode component of each port, extract the phase difference and energy distribution ratio between modes, and form a port mode feature set. The matching calculation unit is used to determine the set of port pairs by waveguide path tracing, calculate the difference between the fundamental mode coupling transmission efficiency and the energy vector of higher-order modes based on the port mode feature set, generate mode matching coefficients, and construct a matching relationship matrix. The path filtering unit is used to filter port pairs whose pattern matching coefficients meet the preset matching threshold from the matching relationship matrix, establish a candidate path set, calculate the distance penalty factor based on the port spacing, calculate the priority score in combination with the pattern matching coefficient, select a preset number of coupled paths, and obtain the coupled path configuration. The parameter prediction unit is used to predict the waveguide spacing and coupling region length of each coupling path in the coupling path configuration through a neural network, thereby obtaining the structural parameter set. The optimization verification unit is used to calculate the coupling efficiency of the target port and the crosstalk level of the non-target port based on the structural parameter set. When the coupling efficiency of the target port is lower than the preset efficiency threshold or the crosstalk level of the non-target port exceeds the preset crosstalk threshold, the matching threshold is adjusted and the path planning step is returned until the threshold requirement is met, and the structural parameter set is output.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images