A parameter adaptive material reflectivity curve generation method and system

Abstract

The application discloses a kind of parameter self-adapting material reflectivity curve generation method and system, comprising: S1, constructing discrete sample set;S2, form conditional sample set;S3, cost coupling parallel point selection is executed using improved parallel bayesian optimization algorithm;S4, input improved DeepONet model, introduce conditional kernel generation mechanism in integral kernel operator layer, output predicted reflectivity curve;S5, training operator parameters are generated using Kramers-Kronig algorithm to execute Hilbert transform;S6, improved parallel bayesian optimization algorithm is used to generate update operator parameters;S7, physical boundary constraint is executed, and material reflectivity curve is output.The present application realizes the high-precision prediction and self-adaptive update of material reflectivity curve, improves the efficiency and amplitude and phase consistency of measurement point selection, enhances the stability and physical rationality of reflectivity curve generation under different working conditions.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic parameter testing and material electromagnetic property modeling technology, and in particular to a method and system for generating material reflectivity curves with adaptive parameters. Background Technology

[0002] With the increasing demand for electromagnetic stealth materials, absorbing structures, and broadband electromagnetic compatibility design, the technology for obtaining and predicting reflectivity curves of materials under different incident angles, polarization modes, and structural parameters has attracted widespread attention. Existing material reflectivity testing methods mainly rely on fixed-frequency scanning and point-by-point measurements to obtain the amplitude and phase of the reflection coefficient at discrete frequency points, and then use simple interpolation to form a reflectivity curve. However, these methods generally suffer from the following problems in practical applications: The material reflectivity testing process is significantly affected by factors such as thickness deviation, electrical center offset, and attitude error. The measured data exhibits discrete fluctuations and phase jumps on the frequency axis. Traditional curve fitting methods struggle to maintain the consistency and continuity of amplitude and phase under multi-parameter coupling conditions, resulting in non-physical oscillations or local distortions in the generated reflectivity curves. The number of combinations of test frequencies, incident angles, and polarization modes is enormous, making full-scan testing costly and inefficient. Existing measurement point selection strategies are mostly based on experience or a single information gain criterion, failing to consider the cost of incident angle changes and test constraints, easily leading to redundant measurements or constraint conflicts. For frequency-related nonlinear reflection characteristics, traditional modeling methods based on polynomial regression or piecewise interpolation struggle to simultaneously characterize the intrinsic causal relationship between amplitude dispersion and phase dispersion, lacking a modeling mechanism for the coordinated constraint of reflection coefficient amplitude and phase. This results in insufficient stability of prediction results under small sample conditions, making it difficult to meet the accuracy and consistency requirements for parameter-adaptive reflectivity curve generation.

[0003] Therefore, how to provide a method and system for generating material reflectivity curves with adaptive parameters is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a parameter-adaptive method and system for generating material reflectivity curves. This invention utilizes a material reflectivity testing system, an improved parallel Bayesian optimization algorithm, and an improved DeepONet model. It constructs discrete and conditional sample sets around the test frequency, incident angle, and polarization mode. Through cost-coupled parallel point selection, it generates a batch of measurement points. It introduces a conditional kernel function and frequency domain integration to generate the predicted reflection coefficient amplitude and phase. Furthermore, it combines the Kramers-Kronig algorithm and Hilbert transform to construct a causal phase benchmark, completing the generation of training and updating operator parameters. Finally, it forms a material reflectivity curve that satisfies amplitude boundary constraints and phase continuity constraints. This method possesses advantages such as high measurement point selection efficiency, good amplitude and phase consistency, strong parameter adaptability, and high reflectivity curve prediction accuracy.

[0005] A method for generating a material reflectance curve with adaptive parameters according to an embodiment of the present invention includes the following steps: S1. Set the test frequency, incident angle and polarization mode in the material reflectivity testing system, perform reflection coefficient measurement on the material under test, and obtain the reflection coefficient amplitude and reflection coefficient phase at discrete frequency points; simultaneously acquire the thickness ratio, electrical center offset and attitude error of the material under test, and construct a discrete sample set. S2. Construct a working condition vector based on the discrete sample set, and generate a phase cue quantity based on the test frequency, thickness ratio, incident angle, electrical center offset and attitude error to form a condition sample set; S3. Based on the conditional sample set, an improved parallel Bayesian optimization algorithm is used to construct the information collection volume and execution cost. The execution cost is coupled with parallel point selection to obtain a batch of measurement points. S4. Input the operating condition vector, phase cue quantity and test frequency into the improved DeepONet model. The improved DeepONet model introduces a conditional kernel generation mechanism in the integral kernel sub-layer. Based on the operating condition vector, a conditional kernel function is generated. Frequency domain integration is performed on the frequency domain basis function to generate the predicted reflection coefficient amplitude and the predicted reflection coefficient phase, and the predicted reflectivity curve is output. S5. The improved DeepONet model is trained based on the reflection coefficient amplitude and reflection coefficient phase corresponding to the discrete sample set and the batch measurement point set. During the training process, the Kramers-Kronig algorithm is used to perform Hilbert transform on the predicted reflection coefficient amplitude to generate a causal phase reference. The predicted reflection coefficient phase is constrained based on the causal phase reference to generate training operator parameters. S6. Obtain the target test frequency, target incident angle and target polarization mode. Based on the training operator parameters and conditional sample set, use an improved parallel Bayesian optimization algorithm to obtain the target batch measurement point set. Obtain the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch measurement point set. Perform small sample adaptive update to generate update operator parameters. S7. Based on the updated operator parameters, generate the material reflectivity curve under the target working condition, execute physical boundary constraints, and output the material reflectivity curve.

[0006] Optionally, the step of setting the test frequency, incident angle, and polarization mode in the material reflectivity testing system, and performing a reflection coefficient measurement on the material under test to obtain the reflection coefficient amplitude and reflection coefficient phase at discrete frequency points, specifically involves: In the material reflectivity testing system, the frequency scanning range and frequency step size are set, and the test frequency signal is output point by point according to the frequency step size. Adjust the drive turntable to the set incident angle position and lock the turntable angle; Set the polarization mode and switch the transmit and receive polarization states; Reflected signals are acquired at each test frequency, incident angle, and polarization combination. The amplitude and phase of the reflection coefficient are calculated based on the acquired reflected signals. The amplitude and phase of the reflection coefficient are recorded according to the correspondence between the test frequency, incident angle, and polarization.

[0007] Optionally, the simultaneous acquisition of the thickness ratio, electrical center offset, and attitude error of the material under test to construct a discrete sample set specifically involves: The thickness parameters of the material to be tested, the geometric coordinates of the sample mounting position, and the turntable attitude data are obtained; the thickness ratio is calculated based on the test frequency, the electrical center offset is calculated based on the geometric coordinates of the sample mounting position, and the attitude error is calculated based on the turntable attitude data. The test frequency, incident angle, polarization mode, thickness ratio, electrical center offset, attitude error, and corresponding reflection coefficient amplitude and reflection coefficient phase are associated and numbered for storage to construct a discrete sample set.

[0008] Optionally, S2 specifically includes: Based on the incident angle, polarization mode, thickness ratio, electrical center offset and attitude error, the working condition vector is generated by splicing the sample numbers. The test frequency, thickness ratio, incident angle, electrical center offset and attitude error are read from the discrete sample set. A frequency-thickness term is generated based on the test frequency and thickness ratio, a frequency-angle term is generated based on the test frequency and incident angle, and an error phase term is generated based on the electrical center offset and attitude error. The frequency-thickness term, frequency-angle term and error phase term are combined to generate a phase indication value. Align the operating condition vector, phase indication quantity, and corresponding test frequency with the sample, and write the reflection coefficient amplitude and reflection coefficient phase corresponding to the test frequency into the alignment result to form a condition sample set.

[0009] Optionally, the improved parallel Bayesian optimization algorithm is specifically as follows: A candidate measurement point set is constructed based on the conditional sample set. The candidate measurement point set contains candidate combinations of test frequency, incident angle and polarization mode. The information collection quantity is calculated based on the conditional sample set for the candidate measurement point set, and the information collection quantity corresponds to the candidate measurement point set. The execution cost is calculated based on the candidate test point set. The execution cost includes the incident angle change cost, polarization switching cost, and risk cost. The incident angle change cost is determined by the difference between the candidate incident angle and the reference incident angle. The polarization switching cost is determined by the difference between the candidate polarization mode and the reference polarization mode. The risk cost is calculated and determined by the test constraints corresponding to the candidate incident angle and the candidate test frequency. A cost-coupled scoring system is constructed based on the amount of information collected and the execution cost. Parallel point selection is then performed on the candidate measurement point set based on the cost-coupled scoring system, and a batch measurement point set is output. If the incident angle corresponding to the candidate test point set does not meet the turntable limit constraint, the corresponding candidate test point is removed from the candidate test point set. If the incident angle corresponding to the candidate test point set does not meet the anti-collision constraint, the corresponding candidate test point is removed from the candidate test point set. If the incident angle and test frequency corresponding to the candidate test point set do not meet the quiet zone constraint, the corresponding candidate test point is removed from the candidate test point set.

[0010] Optionally, the improved DeepONet model specifically includes a conditional coding layer, a frequency domain remapping layer, a basis function generation layer, an integral kernel operator layer, and a kernel normalization layer; The conditional coding layer forms a one-dimensional numerical sequence by horizontally arranging the working condition vector in column order. The numerical value at each position in the one-dimensional numerical sequence is multiplied by the corresponding mapping coefficient and then summed to obtain the corresponding position value. All position values ​​are arranged in column order to form a conditional feature vector. The frequency domain remapping layer multiplies the value at each position in the conditional feature vector by the test frequency and then sums them to obtain the frequency offset. It also squares the value at each position in the conditional feature vector, sums them, and then takes the square root to obtain the frequency scaling factor. Finally, it adds the frequency offset to the test frequency and multiplies it by the frequency scaling factor to obtain the remapping frequency. The remapping frequency, squared remapping frequency, cubic remapping frequency, sine remapping frequency, and cosine remapping frequency generated by the basis function layer are arranged horizontally in column order to form frequency domain basis functions. The integral kernel sublayer arranges the conditional feature vector and the phase cue quantity horizontally in column order to form a spliced ​​vector. The value at each position in the spliced ​​vector is multiplied by the corresponding kernel coefficient and then summed to obtain the conditional kernel function. The value at each position in the conditional kernel function is multiplied by the corresponding value in the frequency domain basis function, and then summed in the test frequency range and multiplied by the frequency step size to generate the predicted reflection coefficient amplitude. The kernel normalization layer divides the conditional kernel function by the norm obtained by the square root of the sum of the squares of the values ​​at each position in the conditional kernel function to obtain the normalized conditional kernel function. Based on the normalized conditional kernel function and the frequency domain basis function, the phase of the predicted reflection coefficient is generated. The test frequency, the amplitude of the predicted reflection coefficient, and the phase of the predicted reflection coefficient are associated with the same test frequency index to form a frequency point record. The frequency points are sorted in ascending order of test frequency; the sorted predicted reflection coefficient amplitudes are arranged in order of test frequency to form a predicted reflection coefficient amplitude sequence; the sorted predicted reflection coefficient phases are arranged in order of test frequency to form a predicted reflection coefficient phase sequence. The predicted reflectance amplitude sequence and the predicted reflectance phase sequence are arranged correspondingly on the test frequency axis to form the predicted reflectance curve.

[0011] Optionally, S5 specifically includes: The test frequency, operating condition vector, phase indication, reflection coefficient amplitude and reflection coefficient phase are read from the discrete sample set and the batch measurement point set to form training sample pairs; Input the operating condition vector, phase cue, and test frequency into the improved DeepONet model to generate the predicted reflection coefficient amplitude and the predicted reflection coefficient phase. The amplitudes of the predicted reflection coefficients are arranged in ascending order of the test frequency to form an amplitude sequence. The Hilbert transform is then performed on the amplitude sequence to obtain the causal phase reference sequence. Calculate the phase deviation sequence between the predicted reflection coefficient phase and the causal phase reference sequence; Based on the amplitude error between the amplitude of the reflection coefficient and the amplitude of the predicted reflection coefficient, and the phase error between the phase of the reflection coefficient and the phase of the predicted reflection coefficient, the phase deviation sequence forms the training target term; The improved DeepONet model is iteratively updated based on the training objective terms to generate training operator parameters.

[0012] Optionally, S6 specifically includes: Obtain the target test frequency, target incident angle and target polarization mode, and construct a set of candidate test points for the target; The information collection volume is calculated for the target candidate measurement point set based on the conditional sample set, and the execution cost is calculated. A cost coupling score is constructed based on the information collection volume and the execution cost. Parallel point selection is performed on the target candidate measurement point set to obtain the target batch measurement point set. In the material reflectivity testing system, the reflection coefficient is measured based on a target batch of measurement points to obtain the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch of measurement points. The test frequency, incident angle and polarization mode corresponding to the target batch measurement point set are associated with the reflection coefficient amplitude and reflection coefficient phase data to construct a target small sample set; Based on a small target sample set, a target operating condition vector and a target phase cue quantity are constructed. The target operating condition vector, the target phase cue quantity, and the test frequency are then input into an improved DeepONet model. The trained operator parameters are updated based on the amplitude error between the amplitude of the reflection coefficient and the amplitude of the predicted reflection coefficient, as well as the phase error between the phase of the reflection coefficient and the phase of the predicted reflection coefficient, to generate updated operator parameters.

[0013] Optionally, S7 specifically includes: The target operating condition vector, target phase indication, and target test frequency are input into the improved DeepONet model, and the predicted reflection coefficient amplitude and predicted reflection coefficient phase are generated based on the updated operator parameters. Arrange the predicted reflection coefficient amplitude and predicted reflection coefficient phase in order of the target test frequency to form the material reflectivity curve; Apply amplitude boundary constraints and phase continuity constraints to the material reflectivity curve, and output the material reflectivity curve.

[0014] Optionally, a parameter-adaptive material reflectivity profile generation system includes the following modules: The test acquisition module is used to set the test frequency, incident angle and polarization mode in the material reflectivity test system, perform reflection coefficient measurement on the material under test, and obtain the reflection coefficient amplitude and reflection coefficient phase at discrete frequency points; simultaneously acquire the thickness ratio, electrical center offset and attitude error of the material under test, and construct a discrete sample set. The condition construction module is used to construct a working condition vector based on a discrete sample set, and generate a phase cue quantity based on the test frequency, thickness ratio, incident angle, electrical center offset and attitude error to form a condition sample set; The active point selection module is used to construct the information collection volume and execution cost based on the conditional sample set using an improved parallel Bayesian optimization algorithm. The execution cost is coupled with parallel point selection to obtain a batch of measurement points. The operator prediction module is used to input the operating condition vector, phase cue and test frequency into the improved DeepONet model. The improved DeepONet model introduces a condition kernel generation mechanism in the integral kernel operator layer, generates a condition kernel function based on the operating condition vector, performs frequency domain integration on the frequency domain basis function, generates the predicted reflection coefficient amplitude and the predicted reflection coefficient phase, and outputs the predicted reflectivity curve. The causal training module is used to train the improved DeepONet model based on the reflection coefficient amplitude and reflection coefficient phase corresponding to the discrete sample set and the batch measurement point set. During the training process, the Kramers-Kronig algorithm is used to perform Hilbert transform on the predicted reflection coefficient amplitude to generate a causal phase reference. Based on the causal phase reference, the predicted reflection coefficient phase is constrained to generate training operator parameters. The adaptive update module is used to obtain the target test frequency, target incident angle and target polarization mode, and to obtain the target batch measurement point set by using an improved parallel Bayesian optimization algorithm based on the training operator parameters and conditional sample set. It also obtains the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch measurement point set, performs small sample adaptive update, and generates update operator parameters. The curve generation module is used to generate the material reflectivity curve under the target working condition based on the updated operator parameters, execute physical boundary constraints, and output the material reflectivity curve.

[0015] The beneficial effects of this invention are: An improved parallel Bayesian optimization algorithm is used to construct a parallel point selection method that couples information acquisition volume with execution cost and execution cost. At the candidate measurement point set level, the test frequency, incident angle and polarization mode are constrained simultaneously to reduce redundant measurement and the probability of constraint conflict. This improves the selection efficiency of batch measurement point sets while ensuring information coverage. By introducing a conditional kernel function and performing frequency domain integration in the integral kernel sub-layer of the improved DeepONet model, the working condition vector and frequency domain basis function are coupled to generate the amplitude and phase of the predicted reflection coefficient, thereby improving the collaborative modeling capability of amplitude dispersion and phase dispersion and enhancing the expression accuracy of the material reflectivity curve. The Kramers–Kronig algorithm is used to perform Hilbert transform on the predicted reflection coefficient amplitude to generate a causal phase reference. The phase of the predicted reflection coefficient is then constrained based on the causal phase reference, which improves the causal consistency and physical rationality between the predicted reflection coefficient amplitude and the predicted reflection coefficient phase. By constructing a target small sample set from a target batch of measurement points and performing small sample adaptive updates, update operator parameters are generated, enabling the material reflectivity curve to adaptively adjust with changes in target test frequency, target incident angle, and target polarization mode, thereby improving the stability and accuracy of reflectivity curve prediction under different working conditions. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a parameter-adaptive material reflectivity curve generation method proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved parallel Bayesian optimization algorithm proposed in this invention; Figure 3 This is a schematic diagram of the structure of the improved DeepONet model proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A method for generating material reflectivity curves with adaptive parameters, comprising the following steps: S1. Set the test frequency, incident angle and polarization mode in the material reflectivity testing system, perform reflection coefficient measurement on the material under test, and obtain the reflection coefficient amplitude and reflection coefficient phase at discrete frequency points; simultaneously acquire the thickness ratio, electrical center offset and attitude error of the material under test, and construct a discrete sample set. S2. Construct a working condition vector based on a discrete sample set, and generate phase cue quantities based on test frequency, thickness ratio, incident angle, electrical center offset and attitude error to form a condition sample set; S3. Based on the conditional sample set, an improved parallel Bayesian optimization algorithm is used to construct the information collection volume and execution cost. The execution cost is coupled with parallel point selection to obtain a batch of measurement points. S4. Input the operating condition vector, phase cue quantity and test frequency into the improved DeepONet model. The improved DeepONet model introduces a conditional kernel generation mechanism in the integral kernel sub-layer. It generates a conditional kernel function based on the operating condition vector, performs frequency domain integration on the frequency domain basis function, generates the predicted reflection coefficient amplitude and the predicted reflection coefficient phase, and outputs the predicted reflectivity curve. S5. The improved DeepONet model is trained based on the reflection coefficient amplitude and reflection coefficient phase corresponding to the discrete sample set and the batch measurement point set. During the training process, the Kramers-Kronig algorithm is used to perform Hilbert transformation on the predicted reflection coefficient amplitude to generate a causal phase reference. The reflection coefficient phase is then constrained based on the causal phase reference to generate training operator parameters. S6. Obtain the target test frequency, target incident angle and target polarization mode. Based on the training operator parameters and conditional sample set, use an improved parallel Bayesian optimization algorithm to obtain the target batch measurement point set. Obtain the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch measurement point set. Perform small sample adaptive update to generate update operator parameters. S7. Based on the updated operator parameters, generate the material reflectivity curve under the target working condition, execute physical boundary constraints, and output the material reflectivity curve.

[0019] In this embodiment, the test frequency, incident angle, and polarization mode are set in the material reflectivity testing system, and the reflection coefficient is measured on the material under test to obtain the reflection coefficient amplitude and reflection coefficient phase at discrete frequency points, specifically: In the material reflectivity testing system, set the frequency scan start value and frequency scan end value, and set the frequency step size. The test frequency satisfies the condition that the test frequency is equal to the frequency scan start value plus the frequency step size multiplied by the frequency sequence number. The frequency sequence number is an integer greater than or equal to 0 and the corresponding test frequency does not exceed the frequency scan end value. Output continuous wave signals point by point according to the test frequency sequence. The amplitude of the continuous wave signal remains constant, and the transmission power remains within the rated range of the equipment. The turntable is driven to adjust to the set incident angle, and the actual angle of the turntable is read in real time through the angle encoder. When the absolute value of the difference between the actual angle of the turntable and the set incident angle is less than the angle tolerance, the turntable angle is locked. The polarization mode is set to horizontal polarization or vertical polarization. The polarization switching component changes the electric field direction of the transmitter and receiver so that the electric field direction of the transmitter is consistent with the electric field direction of the receiver. The reflected signal is acquired under each combination of test frequency, incident angle and polarization mode. The reflected signal is a complex signal composed of in-phase component and quadrature component. The amplitude of the reflection coefficient is equal to the square root of the sum of the squares of the in-phase component and the squares of the quadrature component. The phase of the reflection coefficient is equal to the arctangent of the quadrature component divided by the in-phase component. According to the correspondence between test frequency, incident angle and polarization mode, the test frequency, incident angle, polarization mode, reflection coefficient amplitude and reflection coefficient phase are arranged in the order of the fields to form data rows, and stored in order of test frequency from small to large. When simultaneously acquiring thickness ratio, electrical center offset, and attitude error, the thickness ratio is equal to the actual thickness of the material under test divided by the free space wavelength at the corresponding test frequency, and the free space wavelength is equal to the speed of light divided by the test frequency; the electrical center offset is equal to the distance difference between the geometric center position of the sample and the electromagnetic reference center position; the attitude error is equal to the difference between the actual angle of the turntable and the set incident angle. The test frequency, incident angle, polarization mode, thickness ratio, electrical center offset, attitude error, reflection coefficient amplitude and reflection coefficient phase are aligned in the row direction according to the same sample number and arranged in the column direction to form a data matrix, with each row corresponding to a discrete frequency point sample. The data matrix is ​​sorted in ascending order of test frequency and then stored to form a discrete sample set.

[0020] In this embodiment, the thickness ratio, electrical center offset, and attitude error of the material under test are acquired simultaneously to construct a discrete sample set, specifically as follows: The actual thickness of the material to be tested is obtained by measuring the actual thickness with a micrometer and taking the arithmetic mean as the thickness parameter. The thickness ratio is calculated based on the free space wavelength corresponding to the test frequency. The free space wavelength is equal to the speed of light divided by the test frequency, and the thickness ratio is equal to the actual thickness divided by the free space wavelength. Obtain the geometric coordinates of the sample installation position. The geometric coordinates are given by the three-dimensional measuring device, which shows the spatial position of the sample's geometric center in the test coordinate system. The electromagnetic reference center is the spatial position of the antenna phase center in the test coordinate system. The electric center offset is equal to the distance difference between the sample's geometric center position and the electromagnetic reference center position projected in the incident direction. The distance difference is equal to the difference of the dot products of the two position coordinates on the unit vector in the incident direction. Acquire turntable attitude data, which includes the actual incident angle and the zero-position calibration angle of the turntable; the attitude error is equal to the actual incident angle of the turntable minus the set incident angle, plus the zero-position calibration angle correction. The test frequency, incident angle, polarization mode, thickness ratio, electrical center offset, attitude error and the corresponding reflection coefficient amplitude and reflection coefficient phase are correlated with the data according to the sample number. The data correlation adopts row alignment, and the data with the same sample number are arranged in the same row. The fields are arranged in a fixed column order of test frequency, incident angle, polarization mode, thickness ratio, electrical center offset, attitude error, reflection coefficient amplitude and reflection coefficient phase to form a data matrix. The data matrix is ​​sorted in ascending order of test frequency, and each row is assigned a unique number to form a discrete sample set.

[0021] In this embodiment, S2 specifically refers to: The incident angle, polarization mode, thickness ratio, electrical center offset, and attitude error corresponding to the same number are read from the discrete sample set. The numbering is kept consistent in the row direction, and the incident angle, polarization mode, thickness ratio, electrical center offset, and attitude error are arranged in a fixed order in the column direction to form a one-dimensional numerical sequence as the working condition vector. The splicing method is horizontal splicing, that is, the incident angle value, polarization mode value, thickness ratio value, electrical center offset value, and attitude error value are arranged in a predetermined column order to form the same row, without performing superposition or weighting processing. The test frequency and thickness ratio are read from the discrete sample set, and the frequency-thickness term is equal to the test frequency multiplied by the thickness ratio; the test frequency and incident angle are read from the discrete sample set, and the frequency-angle term is equal to the test frequency multiplied by the sine of the incident angle; the electrical center offset and attitude error are read from the discrete sample set, and the error phase term is equal to the sum of the electrical center offset and attitude error. The phase indication quantity is formed by horizontally arranging the frequency thickness term, frequency angle term, and error phase term in a fixed column order to form a three-dimensional numerical sequence. The first column is the frequency thickness term, the second column is the frequency angle term, and the third column is the error phase term. The operating condition vector, phase indication quantity and corresponding test frequency are aligned according to the number, and data with the same number are arranged in the same row. In the aligned data row, the reflection coefficient amplitude and reflection coefficient phase corresponding to the test frequency are written to the end of the row. The fields are ordered as follows: test frequency, incident angle, polarization mode, thickness ratio, electrical center offset, attitude error, frequency-thickness term, frequency-angle term, error phase term, reflection coefficient amplitude, and reflection coefficient phase. All data rows are sorted in ascending order of test frequency to form a matrix-structured data set, which constitutes the conditional sample set.

[0022] In this embodiment, the improved parallel Bayesian optimization algorithm is specifically as follows: Based on the conditional sample set, the discrete value set of test frequency, the discrete value set of incident angle, and the discrete value set of polarization mode are read. A candidate test point set is generated by using the Cartesian combination method. The Cartesian combination method involves pairing the test frequency value, incident angle value, and polarization mode value one by one to form a triple. The triple field order is test frequency, incident angle, and polarization mode. The candidate test point set is sorted by test frequency from small to large, and under the same test frequency, it is sorted by incident angle from small to large. The information collection quantity is calculated based on the conditional sample set for the candidate test point set. The information collection quantity consists of a predicted mean term and an uncertainty term. The predicted mean term is a binary quantity consisting of the weighted average of the amplitude of the reflection coefficients of the candidate test point's neighborhood samples and the weighted average of the phase of the reflection coefficients of the candidate test point's neighborhood samples. The uncertainty term is a binary quantity consisting of the weighted square root of the variance of the amplitude of the reflection coefficients of the candidate test point's neighborhood samples and the weighted square root of the variance of the phase of the reflection coefficients of the candidate test point's neighborhood samples. The neighborhood sample selection meets the following conditions: the absolute value of the difference between the neighborhood samples and the candidate test points does not exceed the frequency window width, the absolute value of the difference in the incident angle does not exceed the angle window width, and the polarization mode is consistent. The weight is determined by the normalized distance between the candidate test point and the neighborhood samples in the test frequency and incident angle directions. The normalized distance is equal to the square root of the sum of the squares of the absolute value of the difference in the test frequency divided by the frequency window width and the absolute value of the difference in the incident angle divided by the angle window width. The weight is equal to 1 divided by 1 plus the normalized distance. The information collection quantity is equal to the predicted mean term plus the exploration coefficient multiplied by the uncertainty term. The execution cost is calculated based on the candidate measurement point set. The execution cost is obtained by weighted summation of the incident angle change cost, polarization switching cost, and risk cost. The incident angle change cost is equal to the absolute value of the difference between the candidate incident angle and the reference incident angle. The reference incident angle is taken from the incident angle of the previous measurement point or from the preset reference incident angle. The polarization switching cost is 0 if the candidate polarization mode is the same as the reference polarization mode and 1 if they are different. The reference polarization mode is taken from the polarization mode of the previous measurement point or from the preset reference polarization mode. The risk cost is obtained by weighted summation of the turntable limit margin cost, anti-collision margin cost, and quiet zone deviation cost. The cost of the turntable limit margin is determined by the margin between the incident angle and the minimum and maximum allowable angles of the turntable. The margin is equal to the incident angle minus the minimum value between the minimum and maximum allowable angles of the turntable and the incident angle minus the minimum allowable angle of the turntable. If the margin is less than the angle safety margin, the cost of the turntable limit margin is 1; if the margin is greater than or equal to the angle safety margin, the cost of the turntable limit margin is 0. The collision avoidance margin cost is determined by the minimum distance between the outer contour of the structure of the turntable at the candidate incident angle position and the fixed structure around the turntable in the darkroom. The minimum distance is obtained by looking up the pre-calibrated incident angle and minimum distance lookup table or by calculating the turntable pose and the fixed structure pose. If the minimum distance is less than the safety distance, the collision avoidance margin cost is 1, and if the minimum distance is greater than or equal to the safety distance, the collision avoidance margin cost is 0. The quiet zone deviation cost is determined by the coverage of the irradiation path corresponding to the candidate incident angle and candidate test frequency within the quiet zone. The coverage is determined by the proportion of the area of ​​the effective area of ​​the sample falling into the quiet zone window. If the area proportion is less than the quiet zone proportion threshold, the quiet zone deviation cost is 1. If the area proportion is greater than or equal to the quiet zone proportion threshold, the quiet zone deviation cost is 0. A cost-coupled score is constructed based on the amount of information collected and the execution cost. The cost-coupled score is equal to the product of the amount of information collected and the information weight minus the product of the execution cost and the cost weight. The information weight and the cost weight are preset non-negative constants. Parallel point selection is performed on the candidate test point set based on the cost coupling score. The point selection method is to sort the candidate test points from largest to smallest according to the cost coupling score and select a preset number of candidate test points in sequence to form a batch test point set. Any two test points in the batch test point set must satisfy the condition that the absolute value of the difference between the test frequencies is not less than the frequency separation threshold or the absolute value of the difference between the incident angles is not less than the angle separation threshold. If the separation condition is not met, the corresponding candidate test point is skipped and the selection continues to the next step. If the incident angle of a candidate measurement point does not meet the constraint between the minimum allowable angle and the maximum allowable angle of the turntable, the candidate measurement point is removed from the candidate measurement point set. If the minimum spacing corresponding to a candidate measurement point is less than the safe distance, the candidate measurement point is removed from the candidate measurement point set. If the area ratio of the quiet zone corresponding to a candidate measurement point is less than the quiet zone ratio threshold, the candidate measurement point is removed from the candidate measurement point set.

[0023] In this embodiment, the improved parallel Bayesian optimization algorithm introduces the working condition vector and phase cue as conditional inputs during the information acquisition calculation process, thus associating the information acquisition with material working condition parameters and improving the sensitivity of measurement point selection to changes in the target reflectivity curve. During the execution cost construction process, it introduces the incident angle change cost, polarization switching cost, and risk cost to form a unified execution cost, which is then coupled with the information acquisition cost for scoring. This ensures that measurement point selection simultaneously considers information gain and test constraints, reducing the number of turntable switches and the probability of constraint conflicts. During the parallel point selection process, it outputs a batch of measurement points based on the cost-coupled scoring ranking results, making the distribution of measurement points more balanced on the test frequency axis, improving the convergence stability and reflectivity curve prediction accuracy in the small-sample adaptive update stage.

[0024] In this embodiment, the improved DeepONet model specifically includes a conditional coding layer, a frequency domain remapping layer, a basis function generation layer, an integral kernel sublayer, and a kernel normalization layer; The conditional coding layer forms a one-dimensional numerical sequence by horizontally arranging the working condition vector in column order. The numerical value at each position in the one-dimensional numerical sequence is multiplied by the corresponding mapping coefficient and then summed to obtain the corresponding position value. All position values ​​are arranged in column order to form a conditional feature vector. The frequency domain remapping layer multiplies the value at each position in the conditional feature vector by the test frequency and then sums them to obtain the frequency offset. It also squares the value at each position in the conditional feature vector, sums them, and then takes the square root to obtain the frequency scaling factor. Finally, it adds the frequency offset to the test frequency and multiplies it by the frequency scaling factor to obtain the remapped frequency. The remapping frequencies, squared remapping frequencies, cubic remapping frequencies, sine remapping frequencies, and cosine remapping frequencies of the basis function generation layer are arranged horizontally in column order to form frequency domain basis functions. The integral kernel sublayer arranges the conditional feature vector and the phase cue quantity horizontally in column order to form a spliced ​​vector. The value at each position in the spliced ​​vector is multiplied by the corresponding kernel coefficient and then summed to obtain the conditional kernel function. The value at each position in the conditional kernel function is multiplied by the corresponding value in the frequency domain basis function, and then summed in the test frequency range and multiplied by the frequency step size to generate the predicted reflection coefficient amplitude. The kernel normalization layer divides the conditional kernel function by the norm obtained by the square root of the sum of the squares of the values ​​at each position in the conditional kernel function to obtain the normalized conditional kernel function. Based on the normalized conditional kernel function and the frequency domain basis function, the phase of the predicted reflection coefficient is generated. The test frequency, the amplitude of the predicted reflection coefficient, and the phase of the predicted reflection coefficient are associated with the same test frequency index to form a frequency point record. The frequency points are sorted in ascending order of test frequency; the sorted predicted reflection coefficient amplitudes are arranged in order of test frequency to form a predicted reflection coefficient amplitude sequence; the sorted predicted reflection coefficient phases are arranged in order of test frequency to form a predicted reflection coefficient phase sequence. The predicted reflectance amplitude sequence and the predicted reflectance phase sequence are arranged correspondingly on the test frequency axis to form the predicted reflectance curve.

[0025] In this embodiment, the improved DeepONet model uses a weighted summation of mapping coefficients to generate conditional feature vectors at the conditional coding layer, ensuring a stable representation of the working condition vectors in the numerical domain and reducing the discrete perturbations of thickness ratio, electrical center offset, and attitude error on the model output. At the frequency domain remapping layer, frequency offset and frequency scaling coefficients are introduced to remap the test frequency, ensuring the frequency domain basis functions remain consistent with changes in material working conditions and improving the adaptability of the frequency response to these changes. At the basis function generation layer, a frequency domain basis function containing remapped frequency multi-order terms and trigonometric function terms is constructed, enhancing the ability to express nonlinear dispersion characteristics. At the integral kernel layer, a conditional kernel function is introduced and frequency domain integration is performed, allowing the reflection coefficient amplitude to be coupled between the conditional kernel function and the frequency domain basis function, improving the accuracy of the amplitude spectrum shape. At the kernel normalization layer, norm normalization is performed on the conditional kernel function, ensuring that the predicted reflection coefficient phase and predicted reflection coefficient amplitude maintain scale consistency, improving the numerical stability and causal consistency of the reflectivity curve.

[0026] In this embodiment, S5 specifically refers to: The test frequency, operating condition vector, phase indication, reflection coefficient amplitude and reflection coefficient phase are read from the discrete sample set and the batch measurement point set to form training sample pairs; Input the operating condition vector, phase cue, and test frequency into the improved DeepONet model to generate the predicted reflection coefficient amplitude and the predicted reflection coefficient phase. The amplitudes of the predicted reflection coefficients are arranged in ascending order of the test frequency to form an amplitude sequence. The Hilbert transform is then performed on the amplitude sequence to obtain the causal phase reference sequence. Calculate the phase deviation sequence between the predicted reflection coefficient phase and the causal phase reference sequence; Based on the amplitude error between the amplitude of the reflection coefficient and the amplitude of the predicted reflection coefficient, and the phase error between the phase of the reflection coefficient and the phase of the predicted reflection coefficient, the phase deviation sequence forms the training target term; The improved DeepONet model is iteratively updated based on the training objective terms to generate training operator parameters.

[0027] In this embodiment, S6 specifically refers to: Obtain the target test frequency, target incident angle and target polarization mode, and construct a set of candidate test points for the target; The information collection volume is calculated for the target candidate measurement point set based on the conditional sample set, and the execution cost is calculated. A cost coupling score is constructed based on the information collection volume and the execution cost. Parallel point selection is performed on the target candidate measurement point set to obtain the target batch measurement point set. In the material reflectivity testing system, the reflection coefficient is measured based on a target batch of measurement points to obtain the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch of measurement points. The test frequency, incident angle and polarization mode corresponding to the target batch measurement point set are associated with the reflection coefficient amplitude and reflection coefficient phase data to construct a target small sample set; Based on a small target sample set, a target operating condition vector and a target phase cue quantity are constructed. The target operating condition vector, the target phase cue quantity, and the test frequency are then input into an improved DeepONet model. The trained operator parameters are updated based on the amplitude error between the amplitude of the reflection coefficient and the amplitude of the predicted reflection coefficient, as well as the phase error between the phase of the reflection coefficient and the phase of the predicted reflection coefficient, to generate updated operator parameters.

[0028] In this embodiment, S7 specifically refers to: The target operating condition vector, target phase indication, and target test frequency are input into the improved DeepONet model, and the predicted reflection coefficient amplitude and predicted reflection coefficient phase are generated based on the updated operator parameters. Arrange the target test frequencies in ascending order to form a target test frequency sequence; The predicted reflection coefficient amplitudes are rearranged according to the corresponding indices of the target test frequency sequence to form a predicted reflection coefficient amplitude sequence; The predicted reflection coefficient phases are rearranged according to the corresponding indices of the target test frequency sequence to form the predicted reflection coefficient phase sequence; The predicted reflection coefficient amplitude sequence and the predicted reflection coefficient phase sequence are mapped one-to-one with the same index to construct a complex reflection coefficient sequence. The complex reflection coefficient is equal to the predicted reflection coefficient amplitude multiplied by an exponential function, and the exponent of the exponential function is equal to the imaginary unit multiplied by the predicted reflection coefficient phase. The material reflectivity curve is formed by using the target test frequency sequence as the horizontal axis and the complex reflection coefficient sequence as the vertical axis. Physical boundary constraints include amplitude boundary constraints and phase continuity constraints; Amplitude boundary constraints are applied to the predicted reflection coefficient amplitude sequence. The amplitude boundary constraints include assigning a value of 0 to the predicted reflection coefficient amplitude that is less than 0 and assigning a value of 1 to the predicted reflection coefficient amplitude that is greater than 1. A phase continuity constraint is applied to the phase sequence of predicted reflection coefficients. The phase continuity constraint includes calculating the phase difference between the predicted reflection coefficients at adjacent indices. When the phase difference is greater than π, 2π is subtracted from the subsequent phase. When the phase difference is less than negative π, 2π is added to the subsequent phase. Output the material reflectivity curve after amplitude boundary constraints and phase continuity constraints.

[0029] In this embodiment, a parameter-adaptive material reflectivity curve generation system includes the following modules: The test acquisition module is used to set the test frequency, incident angle and polarization mode in the material reflectivity test system, perform reflection coefficient measurement on the material under test, and obtain the reflection coefficient amplitude and reflection coefficient phase at discrete frequency points; simultaneously acquire the thickness ratio, electrical center offset and attitude error of the material under test, and construct a discrete sample set. The condition construction module is used to construct working condition vectors based on discrete sample sets, and generate phase cue quantities based on test frequency, thickness ratio, incident angle, electrical center offset and attitude error to form a condition sample set; The active point selection module is used to construct the information collection volume and execution cost based on the conditional sample set using an improved parallel Bayesian optimization algorithm. The execution cost is coupled with parallel point selection to obtain a batch of measurement points. The operator prediction module is used to input the operating condition vector, phase cue and test frequency into the improved DeepONet model. The improved DeepONet model introduces a conditional kernel generation mechanism in the integral kernel operator layer, generates a conditional kernel function based on the operating condition vector, performs frequency domain integration on the frequency domain basis function, generates the predicted reflection coefficient amplitude and the predicted reflection coefficient phase, and outputs the predicted reflectivity curve. The causal training module is used to train the improved DeepONet model based on the reflection coefficient amplitude and reflection coefficient phase corresponding to the discrete sample set and the batch measurement point set. During the training process, the Kramers-Kronig algorithm is used to perform Hilbert transform on the predicted reflection coefficient amplitude to generate a causal phase reference. Based on the causal phase reference, the reflection coefficient phase is constrained and the training operator parameters are generated. The adaptive update module is used to obtain the target test frequency, target incident angle and target polarization mode. Based on the training operator parameters and conditional sample set, an improved parallel Bayesian optimization algorithm is used to obtain the target batch measurement point set, obtain the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch measurement point set, perform small sample adaptive update, and generate update operator parameters. The curve generation module is used to generate the material reflectivity curve under the target working condition based on the updated operator parameters, execute physical boundary constraints, and output the material reflectivity curve.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to the generation and consistency evaluation of reflectivity curves for composite material samples used in the radome of a certain type of UAV. The on-site testing environment was an electromagnetic anechoic chamber. The material reflectivity testing system consisted of a vector network analyzer, a broadband horn antenna, a turntable, and a polarization switch. The frequency scanning range was set to 8.0 GHz to 12.0 GHz, the frequency step size was set to 0.01 GHz, and the incident angles were 0°, 30°, and 45°. The polarization methods included co-polarization and cross-polarization. Traditional methods often involve point-by-point measurements under a complete frequency grid and multiple incident angle combinations. This results in long measurement times, and even slight drift in the turntable's attitude can introduce phase jumps, leading to significant differences in the curves of the same material under different installation states, making subsequent modeling and acceptance conclusions unstable.

[0031] In this scenario, after the reflection coefficient measurement is completed, the sample thickness parameters, geometric coordinates of the sample installation position, and turntable attitude data are recorded simultaneously. The thickness ratio is obtained from the test frequency, the electrical center offset is obtained from the installation coordinates, and the attitude error is obtained from the attitude data, forming a discrete sample set. The discrete sample set is further used to generate a working condition vector, and a phase cue is generated from the test frequency, thickness ratio, incident angle, electrical center offset, and attitude error, resulting in a conditional sample set. Subsequently, information acquisition quantity and execution cost are constructed in the candidate measurement point set. The execution cost is coupled with parallel point selection to obtain a batch measurement point set. The batch measurement point set completes the measurement point acquisition in the same round of testing, reducing the additional time consumption caused by frequent switching of incident angle and polarization. The working condition vector, phase cue, and test frequency are input into the improved DeepONet model. The integral kernel sublayer generates a conditional kernel function and performs frequency domain integration on the frequency domain basis function, outputting the predicted reflection coefficient amplitude and predicted reflection coefficient phase, which are then arranged in order of test frequency to form the predicted reflectivity curve. During the training phase, the true reflection coefficient amplitude and phase of the discrete sample set and the batch measurement point set are used to construct the training target term. A Hilbert transform is performed on the predicted reflection coefficient amplitude to obtain a causal phase reference sequence, which constrains the predicted reflection coefficient phase, thus obtaining the training operator parameters. In actual deployment, the target test frequency, target incident angle, and target polarization mode are input to generate a batch measurement point set. After completing reflection coefficient measurements on a small number of measurement points, a small-sample adaptive update is performed to obtain the updated operator parameters. Finally, the material reflectivity curve that satisfies the physical boundary constraints is output.

[0032] Table 1 Comparison of Reflectivity Curve Generation Accuracy and Testing Efficiency

[0033] Table 1 shows that, under the same frequency scanning range and frequency step size, the traditional method requires point-by-point measurement of all 401 frequency points. This invention, through an improved parallel Bayesian optimization algorithm, obtains a batch measurement point set, compressing the number of measurement points for a single operating condition combination to between 108 and 132. The total time consumption is reduced from 18.6 min to 28.7 min to 6.4 min to 9.4 min, a reduction of approximately 60%. In terms of accuracy, the root mean square error of amplitude remains between 0.21 dB and 0.37 dB, and the root mean square error of phase remains between 3.8° and 5.8°. The errors increase slightly under conditions of increased incident angle and cross-polarization but remain stable, indicating that the conditional sample set and phase indications effectively characterize changes in operating conditions, and the consistency between the predicted reflection coefficient amplitude and phase on the frequency axis is better.

[0034] Table 2. Comparison of Consistency and Physical Constraints Before and After Small Sample Adaptive Update

[0035] Table 2 focuses on the inconsistency issues in curves caused by installation deviations commonly encountered in real-world testing. Before the update, when the electrical center offset and attitude error were superimposed, the maximum amplitude difference reached 3.08 dB, and the number of phase jumps reached 9, indicating abrupt phase changes between adjacent test frequency points, affecting subsequent curve interpretation. After introducing a target batch of test points and completing small-sample adaptive updates, the maximum amplitude difference converged to 0.42 dB to 1.12 dB, and the number of phase jumps decreased to 0 to 2. This shows that the updated operator parameters can absorb the influence of installation disturbances on the amplitude and phase of the reflection coefficient. At the same time, physical boundary constraints and phase continuity constraints make the curve smoother on the test frequency axis, and the extrapolation segment of the curve no longer exhibits amplitude overshoot and phase discontinuity, improving the consistency and reproducibility of the material reflectivity curve in the acceptance and simulation input stages.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for generating material reflectivity curves with adaptive parameters, characterized in that, Includes the following steps: S1. Set the test frequency, incident angle and polarization mode in the material reflectivity test system, perform reflection coefficient measurement on the material to be tested, and obtain the reflection coefficient amplitude and reflection coefficient phase at discrete frequency points. Simultaneously acquire the thickness ratio, electrical center offset, and attitude error of the material under test to construct a discrete sample set; S2. Construct a working condition vector based on the discrete sample set, and generate a phase cue quantity based on the test frequency, thickness ratio, incident angle, electrical center offset and attitude error to form a condition sample set; S3. Based on the conditional sample set, an improved parallel Bayesian optimization algorithm is used to construct the information collection volume and execution cost. The execution cost is coupled with parallel point selection to obtain a batch of measurement points. S4. Input the operating condition vector, phase cue quantity and test frequency into the improved DeepONet model. The improved DeepONet model introduces a conditional kernel generation mechanism in the integral kernel sub-layer. Based on the operating condition vector, a conditional kernel function is generated. Frequency domain integration is performed on the frequency domain basis function to generate the predicted reflection coefficient amplitude and the predicted reflection coefficient phase, and the predicted reflectivity curve is output. S5. The improved DeepONet model is trained based on the reflection coefficient amplitude and reflection coefficient phase corresponding to the discrete sample set and the batch measurement point set. During the training process, the Kramers-Kronig algorithm is used to perform Hilbert transform on the predicted reflection coefficient amplitude to generate a causal phase reference. The predicted reflection coefficient phase is constrained based on the causal phase reference to generate training operator parameters. S6. Obtain the target test frequency, target incident angle and target polarization mode. Based on the training operator parameters and conditional sample set, use an improved parallel Bayesian optimization algorithm to obtain the target batch measurement point set. Obtain the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch measurement point set. Perform small sample adaptive update to generate update operator parameters. S7. Based on the updated operator parameters, generate the material reflectivity curve under the target working condition, execute physical boundary constraints, and output the material reflectivity curve.

2. The method for generating a parameter-adaptive material reflectivity curve according to claim 1, characterized in that, The process involves setting the test frequency, incident angle, and polarization mode in the material reflectivity testing system, performing a reflection coefficient measurement on the material under test, and obtaining the reflection coefficient amplitude and phase at discrete frequency points. Specifically: In the material reflectivity testing system, the frequency scanning range and frequency step size are set, and the test frequency signal is output point by point according to the frequency step size. Adjust the drive turntable to the set incident angle position and lock the turntable angle; Set the polarization mode and switch the transmit and receive polarization states; The reflected signal is collected under each test frequency, incident angle, and polarization combination, and the amplitude and phase of the reflection coefficient are calculated based on the collected reflected signal. Record the amplitude and phase of the reflection coefficient according to the correspondence between the test frequency, incident angle and polarization mode.

3. The method for generating a parameter-adaptive material reflectivity curve according to claim 1, characterized in that, The process of simultaneously acquiring the thickness ratio, electrical center offset, and attitude error of the material under test, and constructing a discrete sample set, specifically involves: The thickness parameters of the material to be tested, the geometric coordinates of the sample mounting position, and the turntable attitude data are obtained; the thickness ratio is calculated based on the test frequency, the electrical center offset is calculated based on the geometric coordinates of the sample mounting position, and the attitude error is calculated based on the turntable attitude data. The test frequency, incident angle, polarization mode, thickness ratio, electrical center offset, attitude error, and corresponding reflection coefficient amplitude and reflection coefficient phase are associated and numbered for storage to construct a discrete sample set.

4. The method for generating a material reflectivity curve with adaptive parameters according to claim 1, characterized in that, Specifically, S2 is: Based on the incident angle, polarization mode, thickness ratio, electrical center offset and attitude error, the working condition vector is generated by splicing the sample numbers. The test frequency, thickness ratio, incident angle, electrical center offset and attitude error are read from the discrete sample set. A frequency-thickness term is generated based on the test frequency and thickness ratio, a frequency-angle term is generated based on the test frequency and incident angle, and an error phase term is generated based on the electrical center offset and attitude error. The frequency-thickness term, frequency-angle term and error phase term are combined to generate a phase indication value. Align the operating condition vector, phase indication quantity, and corresponding test frequency with the sample, and write the reflection coefficient amplitude and reflection coefficient phase corresponding to the test frequency into the alignment result to form a condition sample set.

5. The method for generating a material reflectivity curve with adaptive parameters according to claim 1, characterized in that, The improved parallel Bayesian optimization algorithm is specifically as follows: A candidate measurement point set is constructed based on the conditional sample set. The candidate measurement point set contains candidate combinations of test frequency, incident angle and polarization mode. The information collection quantity is calculated based on the conditional sample set for the candidate measurement point set, and the information collection quantity corresponds to the candidate measurement point set. The execution cost is calculated based on the candidate test point set. The execution cost includes the incident angle change cost, polarization switching cost, and risk cost. The incident angle change cost is determined by the difference between the candidate incident angle and the reference incident angle. The polarization switching cost is determined by the difference between the candidate polarization mode and the reference polarization mode. The risk cost is calculated and determined by the test constraints corresponding to the candidate incident angle and the candidate test frequency. A cost-coupled scoring system is constructed based on the amount of information collected and the execution cost. Parallel point selection is then performed on the candidate measurement point set based on the cost-coupled scoring system, and a batch measurement point set is output. If the incident angle corresponding to the candidate test point set does not meet the turntable limit constraint, the corresponding candidate test point is removed from the candidate test point set. If the incident angle corresponding to the candidate test point set does not meet the anti-collision constraint, the corresponding candidate test point is removed from the candidate test point set. If the incident angle and test frequency corresponding to the candidate test point set do not meet the quiet zone constraint, the corresponding candidate test point is removed from the candidate test point set.

6. The method for generating a material reflectivity curve with adaptive parameters according to claim 1, characterized in that, The improved DeepONet model specifically includes a conditional coding layer, a frequency domain remapping layer, a basis function generation layer, an integral kernel operator layer, and a kernel normalization layer. The conditional coding layer forms a one-dimensional numerical sequence by horizontally arranging the working condition vector in column order. The numerical value at each position in the one-dimensional numerical sequence is multiplied by the corresponding mapping coefficient and then summed to obtain the corresponding position value. All position values ​​are arranged in column order to form a conditional feature vector. The frequency domain remapping layer multiplies the value at each position in the conditional feature vector by the test frequency and then sums them to obtain the frequency offset. It also squares the value at each position in the conditional feature vector, sums them, and then takes the square root to obtain the frequency scaling factor. Finally, it adds the frequency offset to the test frequency and multiplies it by the frequency scaling factor to obtain the remapping frequency. The remapping frequency, squared remapping frequency, cubic remapping frequency, sine remapping frequency, and cosine remapping frequency generated by the basis function layer are arranged horizontally in column order to form frequency domain basis functions. The integral kernel sublayer arranges the conditional feature vector and the phase cue quantity horizontally in column order to form a spliced ​​vector. The value at each position in the spliced ​​vector is multiplied by the corresponding kernel coefficient and then summed to obtain the conditional kernel function. The value at each position in the conditional kernel function is multiplied by the corresponding value in the frequency domain basis function, and then summed in the test frequency range and multiplied by the frequency step size to generate the predicted reflection coefficient amplitude. The kernel normalization layer divides the conditional kernel function by the norm obtained by the square root of the sum of the squares of the values ​​at each position in the conditional kernel function to obtain the normalized conditional kernel function. Based on the normalized conditional kernel function and the frequency domain basis function, the phase of the predicted reflection coefficient is generated. The test frequency, the amplitude of the predicted reflection coefficient, and the phase of the predicted reflection coefficient are associated with the same test frequency index to form a frequency point record. The frequency points are sorted in ascending order of test frequency; the sorted predicted reflection coefficient amplitudes are arranged in order of test frequency to form a predicted reflection coefficient amplitude sequence; the sorted predicted reflection coefficient phases are arranged in order of test frequency to form a predicted reflection coefficient phase sequence. The predicted reflectance amplitude sequence and the predicted reflectance phase sequence are arranged correspondingly on the test frequency axis to form the predicted reflectance curve.

7. The method for generating a material reflectivity curve with adaptive parameters according to claim 1, characterized in that, Specifically, S5 is: The test frequency, operating condition vector, phase indication, reflection coefficient amplitude and reflection coefficient phase are read from the discrete sample set and the batch measurement point set to form training sample pairs; Input the operating condition vector, phase cue, and test frequency into the improved DeepONet model to generate the predicted reflection coefficient amplitude and the predicted reflection coefficient phase. The amplitudes of the predicted reflection coefficients are arranged in ascending order of the test frequency to form an amplitude sequence. The Hilbert transform is then performed on the amplitude sequence to obtain the causal phase reference sequence. Calculate the phase deviation sequence between the predicted reflection coefficient phase and the causal phase reference sequence; Based on the amplitude error between the amplitude of the reflection coefficient and the amplitude of the predicted reflection coefficient, and the phase error between the phase of the reflection coefficient and the phase of the predicted reflection coefficient, the phase deviation sequence forms the training target term; The improved DeepONet model is iteratively updated based on the training objective terms to generate training operator parameters.

8. The method for generating a material reflectivity curve with adaptive parameters according to claim 1, characterized in that, Specifically, S6 is: Obtain the target test frequency, target incident angle and target polarization mode, and construct a set of candidate test points for the target; The information collection volume is calculated for the target candidate measurement point set based on the conditional sample set, and the execution cost is calculated. A cost coupling score is constructed based on the information collection volume and the execution cost. Parallel point selection is performed on the target candidate measurement point set to obtain the target batch measurement point set. In the material reflectivity testing system, the reflection coefficient is measured based on a target batch of measurement points to obtain the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch of measurement points. The test frequency, incident angle and polarization mode corresponding to the target batch measurement point set are associated with the reflection coefficient amplitude and reflection coefficient phase data to construct a target small sample set; Based on a small target sample set, a target operating condition vector and a target phase cue quantity are constructed. The target operating condition vector, the target phase cue quantity, and the test frequency are then input into an improved DeepONet model. The trained operator parameters are updated based on the amplitude error between the amplitude of the reflection coefficient and the amplitude of the predicted reflection coefficient, as well as the phase error between the phase of the reflection coefficient and the phase of the predicted reflection coefficient, to generate updated operator parameters.

9. The method for generating a material reflectivity curve with adaptive parameters according to claim 1, characterized in that, Specifically, S7 is: The target operating condition vector, target phase indication, and target test frequency are input into the improved DeepONet model, and the predicted reflection coefficient amplitude and predicted reflection coefficient phase are generated based on the updated operator parameters. Arrange the predicted reflection coefficient amplitude and predicted reflection coefficient phase in order of the target test frequency to form the material reflectivity curve; Apply amplitude boundary constraints and phase continuity constraints to the material reflectivity curve, and output the material reflectivity curve.

10. A parameter-adaptive material reflectance curve generation system, comprising executing the parameter-adaptive material reflectance curve generation method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The test acquisition module is used to set the test frequency, incident angle and polarization mode in the material reflectivity test system, perform reflection coefficient measurement on the material under test, and obtain the reflection coefficient amplitude and reflection coefficient phase at discrete frequency points. Simultaneously acquire the thickness ratio, electrical center offset, and attitude error of the material under test to construct a discrete sample set; The condition construction module is used to construct a working condition vector based on a discrete sample set, and generate a phase cue quantity based on the test frequency, thickness ratio, incident angle, electrical center offset and attitude error to form a condition sample set; The active point selection module is used to construct the information collection volume and execution cost based on the conditional sample set using an improved parallel Bayesian optimization algorithm. The execution cost is coupled with parallel point selection to obtain a batch of measurement points. The operator prediction module is used to input the operating condition vector, phase cue and test frequency into the improved DeepONet model. The improved DeepONet model introduces a condition kernel generation mechanism in the integral kernel operator layer, generates a condition kernel function based on the operating condition vector, performs frequency domain integration on the frequency domain basis function, generates the predicted reflection coefficient amplitude and the predicted reflection coefficient phase, and outputs the predicted reflectivity curve. The causal training module is used to train the improved DeepONet model based on the reflection coefficient amplitude and reflection coefficient phase corresponding to the discrete sample set and the batch measurement point set. During the training process, the Kramers-Kronig algorithm is used to perform Hilbert transform on the predicted reflection coefficient amplitude to generate a causal phase reference. Based on the causal phase reference, the predicted reflection coefficient phase is constrained to generate training operator parameters. The adaptive update module is used to obtain the target test frequency, target incident angle and target polarization mode, and to obtain the target batch measurement point set by using an improved parallel Bayesian optimization algorithm based on the training operator parameters and conditional sample set. It also obtains the reflection coefficient amplitude and reflection coefficient phase corresponding to the target batch measurement point set, performs small sample adaptive update, and generates update operator parameters. The curve generation module is used to generate the material reflectivity curve under the target working condition based on the updated operator parameters, execute physical boundary constraints, and output the material reflectivity curve.

Images