XRD pattern candidate peak physical role automatic identification method and system

Abstract

The application provides a kind of XRD pattern candidate peak physical role automatic identification method and system, comprising: obtaining the discrete scattering intensity and angle data sequence output by X-ray diffractometer;Candidate peak preliminary screening and local topography window interception are carried out on angle data sequence, the discrete scattering intensity corresponding to each angle position in local topography window is subjected to logarithmic transformation, and the multi-channel feature spectrum segment matrix is obtained by splicing with the physical prior characteristics of angle data sequence;The multi-channel feature spectrum segment matrix is subjected to feature extraction, channel and position double dynamic focusing and classification using attention-enhanced convolutional neural network model, and the initial probability distribution of the physical role of each candidate peak is output, and global consistency review is carried out, and the final physical role classification result of each candidate peak is screened out. The limitations of unable to identify peaks physically, the difficulty of using nonlinear topography characteristics and the problem of weak peak identification caused by class imbalance are solved.

Description

Technical Field

[0001] This invention relates to the field of X-ray diffraction data analysis and processing technology, specifically to a method and system for automatic identification of the physical role of candidate peaks in XRD patterns. Background Technology

[0002] X-ray diffraction (XRD) is one of the core methods for characterizing the structure of semiconductor superlattice materials. In the XRD rocking curves of superlattice materials, the diffraction peaks consist of a substrate peak, a 0th-order peak representing the average composition of the epitaxial layer, and higher-order satellite peaks such as ±1st and ±2nd orders. Accurately defining "which peak is the substrate peak, which peak is the 0th-order peak, and which peak is the ±1st-order satellite peak" is a crucial prerequisite for subsequent calculations of epitaxial layer thickness, strain, and composition.

[0003] However, existing automated processing technologies face significant bottlenecks in this stage, severely hindering the high-throughput automation of semiconductor epitaxial production lines. Specifically, firstly, traditional peak detection algorithms suffer from dimensional limitations. Traditional algorithms based on pseudo-Voigt functions for full-spectrum fitting, first / second derivative peak finding, or single threshold determination essentially only address the basic detection problem of "whether there is a peak." Their output is typically just a one-dimensional "peak list," and the system cannot understand the physical hierarchy between peaks. Ultimately, it still heavily relies on R&D personnel with deep physics backgrounds for manual image interpretation and role assignment, resulting in extremely low efficiency. Secondly, physical characteristics are difficult to exhaustively represent using static rules. As a one-dimensional continuous signal, the differences between substrate peaks, satellite peaks, zero-order peaks, and noise peaks in XRD spectra mainly lie in highly nonlinear local morphologies (such as peak sharpness, asymmetric continuity of slopes on both sides, broadening degree, local background tilt, and micro-curvature). Existing hard-coded software struggles to fully describe these dynamic features using just a few static mathematical rules, leading to weak satellite peaks often being filtered out as background noise or incorrectly classified as broadened shoulder peaks of the substrate. Thirdly, there is extreme class imbalance and a common defect in AI models: in actual XRD test data, the number of background noise spikes and ordinary interference peaks often reaches hundreds or thousands, while the number of key substrate peaks and ±1-level satellite peaks is only a handful. If conventional deep learning classification models are directly applied, the network will quickly fall into "frequency bias" during training, tending to classify most ambiguous samples as the majority class (noise), completely losing the ability to capture rare weak satellite peaks.

[0004] Therefore, there is an urgent need for an automatic identification method for the physical roles of candidate peaks that can deeply integrate prior physical knowledge of superlattice XRD and possess the ability to self-constrain by physical laws. Summary of the Invention

[0005] The present invention provides a method and system for automatic identification of the physical role of candidate peaks in XRD patterns, which at least solves the problems of the inability to identify physical peaks, the difficulty in utilizing nonlinear morphological features, and the poor identification of weak peaks due to class imbalance.

[0006] According to one aspect of the present invention, an automatic identification method for the physical role of candidate peaks in XRD patterns is provided, comprising:

[0007] Obtain the discrete scattering intensity and angle data sequence output by the X-ray diffractometer;

[0008] Candidate peaks are initially screened and local morphology windows are extracted from the angle data sequence. The discrete scattering intensities corresponding to each angle position within the local morphology window are logarithmically transformed and concatenated with the physical prior features of the angle data sequence to obtain a multi-channel feature spectrum matrix.

[0009] Based on the multi-channel feature spectrum matrix, an attention-enhanced convolutional neural network model is used for feature extraction, dual dynamic focusing on channels and positions, and classification, outputting the initial probability distribution of the physical role of each candidate peak; wherein, the attention-enhanced convolutional neural network model includes a multi-scale one-dimensional convolutional layer, a dual attention layer, and a classifier;

[0010] Based on the symmetry rules of crystallography, the initial probability distribution of the physical role of each candidate peak is globally consistent and verified to select the final classification result of the physical role of each candidate peak.

[0011] As an optional approach, the preliminary screening of candidate peaks and local morphology window truncation of the angle data sequence includes:

[0012] The differential peak finding algorithm is used to determine the center angle position of the suspected candidate peak by the position where the first difference of the angle data sequence changes from positive to negative.

[0013] Using the center angle position of the suspected candidate peak as the center, a local morphology window is obtained by truncating according to a preset step size.

[0014] As an optional approach, the preliminary screening of candidate peaks and local morphology window truncation of the angle data sequence includes:

[0015] A threshold screening algorithm is used to determine the angular positions where the discrete scattering intensity is greater than a preset intensity threshold as the center angular positions of the suspected candidate peaks;

[0016] Using the center angle position of the suspected candidate peak as the center, a local morphology window is obtained by truncating according to a preset step size.

[0017] As an optional approach, the step of performing a logarithmic transformation on the discrete scattering intensities corresponding to each angular position within the local topography window and concatenating them with the physical prior features of the angular data sequence to obtain a multi-channel feature spectrum matrix includes:

[0018] Logarithmic transformation is performed on the discrete scattering intensity corresponding to each angular position within the local topography window, and the logarithmic transformed discrete scattering intensity is normalized to obtain the logarithmic scattering intensity value.

[0019] The discrete scattering intensity corresponding to each angular position within the local topography window is calculated using a difference method to obtain the first-order derivative feature and the second-order derivative feature.

[0020] The angular position corresponding to the maximum discrete scattering intensity in the full spectrum is taken as the global reference main peak angular position, and the relative position code is calculated based on the ratio of the angular offset of each angular position in the local topography window relative to the global reference main peak angular position to the scanning step size.

[0021] The logarithmic scattering intensity value, the first derivative feature, the second derivative feature, and the relative position code are concatenated along the channel dimension to obtain a multi-channel feature spectrum matrix.

[0022] As an optional approach, based on the multi-channel feature spectral matrix, an attention-enhanced convolutional neural network model is used for feature extraction, dual dynamic focusing on channels and positions, and classification, outputting the initial probability distribution of the physical role of each candidate peak, including:

[0023] The multi-channel feature spectrum matrix is ​​input into a multi-scale one-dimensional convolutional layer for multi-scale feature extraction to obtain multi-scale features; wherein, the multi-scale one-dimensional convolutional layer includes a large-scale convolutional layer, a medium-scale convolutional layer, and a small-scale convolutional layer, and the medium-scale convolutional layer and the small-scale convolutional layer are connected across layer residuals.

[0024] The output features of the mid-scale convolutional layer are superimposed onto the output features of the small-scale convolutional layer to obtain residual fusion features;

[0025] The residual fusion features are input into a dual attention layer for channel attention recalibration and spatial attention recalibration to obtain the core feature vector.

[0026] The core feature vector is input into the classifier, which outputs the initial probability distribution of the physical role of each candidate peak.

[0027] As an optional approach, the residual fusion features are input into a dual attention layer for channel attention recalibration and spatial attention recalibration to obtain the core feature vector, including:

[0028] The importance weights of each feature channel in the residual fusion feature are calculated using global average pooling and multilayer perceptron, respectively.

[0029] The residual fusion features are recalibrated channel by channel using the importance weights to obtain the calibrated features;

[0030] Calculate the normalized weights of each angular position within the local topography window;

[0031] The calibrated features are weighted using the normalized weights to obtain the core feature vector.

[0032] As an optional approach, the attention-enhanced convolutional neural network model is trained based on the total joint loss function, including:

[0033] The Focal Loss architecture is adopted, and the weight coefficients are set inversely proportional to the number of samples in each physical role category to construct the classification loss term;

[0034] Based on the superlattice dynamics diffraction theory, a physical penalty loss term is constructed according to the angular distance between the satellite peak and the main substrate peak and a preset safety distance threshold; wherein, the physical penalty loss term is configured to be activated when the angular distance is less than the preset safety distance threshold, thereby generating a gradient penalty;

[0035] An attention regularization term is introduced, and the attention regularization term, the classification loss term, and the physical penalty loss term are weighted and summed to construct a joint loss function;

[0036] Using the joint loss function as the optimization objective, the attention-enhanced convolutional neural network model is iteratively trained until it converges.

[0037] As an optional approach, the final physical role classification results of the candidate peaks include substrate peaks, 0th-order peaks, -1st-order satellite peaks, +1st-order satellite peaks, and noise peaks;

[0038] Based on crystallographic symmetry, a global consistency check is performed on the initial probability distribution of the physical role of each candidate peak to screen out the final physical role classification results of each candidate peak, including:

[0039] Extract the initial probability distribution of the physical role of each candidate peak, determine the candidate peak sets for the -1 level satellite peak, the +1 level satellite peak and the 0 level peak, and obtain the candidate -1 level satellite peak, the candidate +1 level satellite peak and the candidate 0 level peak;

[0040] Based on the angular coordinates of each candidate peak, calculate the first angular distance between the candidate -1 level satellite peak and the candidate 0 level peak, and the second angular distance between the candidate +1 level satellite peak and the candidate 0 level peak, and calculate the asymmetric distance factor based on the deviation between the first angular distance and the second angular distance.

[0041] The joint confidence score is calculated based on the initial probability distributions of the -1 level satellite peak, the +1 level satellite peak, and the 0 level peak, as well as the asymmetric distance factor.

[0042] The peak combination that maximizes the joint confidence score is identified through a global search, and the physical role of each candidate peak in the peak combination is used as the final physical role classification result.

[0043] According to another aspect of the present invention, an automatic identification system for the physical role of candidate peaks in XRD patterns is also provided, applied to a method for automatic identification of the physical role of candidate peaks in XRD patterns, comprising:

[0044] The data acquisition module is used to acquire discrete scattering intensity and angle data sequences output by the X-ray diffractometer.

[0045] The preprocessing module is used to perform candidate peak screening and local morphology window extraction on the angle data sequence, perform logarithmic transformation on the discrete scattering intensity corresponding to each angle position within the local morphology window, and concatenate it with the physical prior features of the angle data sequence to obtain a multi-channel feature spectrum matrix.

[0046] The feature extraction module is used to perform feature extraction, dual dynamic focusing on channels and positions, and classification based on the multi-channel feature spectrum matrix using an attention-enhanced convolutional neural network model, and output the initial probability distribution of the physical role of each candidate peak; wherein, the attention-enhanced convolutional neural network model includes a multi-scale one-dimensional convolutional layer, a dual attention layer, and a classifier;

[0047] The verification and output module is used to perform a global consistency verification of the initial probability distribution of the physical role of each candidate peak based on the crystallographic symmetry law, and to select the final physical role classification result of each candidate peak.

[0048] According to another aspect of the present invention, an electronic device is also provided, the electronic device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the method as described above.

[0049] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0050] The beneficial effects of this invention are as follows:

[0051] 1. This invention concatenates logarithmically transformed discrete scattering intensity data with physical prior features to form a multi-channel feature spectrum matrix. This matrix is ​​then used for feature extraction and dynamic focusing via an attention-enhanced convolutional neural network model. Furthermore, the network output is globally consistent and verified by incorporating crystallographic symmetry principles. These three elements work together to form a closed-loop technology of physical guidance, dynamic focusing, and physical verification. This enables the model to accurately output the physical role classification results of each candidate peak, achieving a technological leap from simply finding peaks to identifying peaks. This avoids the limitations of not being able to physically identify peaks and eliminates the reliance on human experience.

[0052] 2. This invention compresses the extreme dynamic range of strong substrate peaks and weak satellite peaks by performing a logarithmic transformation on the discrete scattering intensity, thus preserving the microscopic protrusion contour of the weak peaks. The synergy of multiple feature channels transforms the nonlinear peak shape features that are difficult to exhaust with static rules into structured data that can be effectively processed by neural networks, achieving high-fidelity extraction and utilization of the local morphological features of diffraction peaks.

[0053] 3. This invention introduces an attention enhancement mechanism, enabling the attention-enhanced convolutional neural network model to adaptively adjust the weight coefficients of each physical channel based on the input samples. Under strong background interference, it automatically suppresses the intensity channel and amplifies the curvature channel, which plays a crucial role in identifying weak satellite peaks, maintaining high sensitivity for rare categories of weak satellite peaks. Simultaneously, a spatial attention mechanism forces the aggregated weights to focus on physical regions with high information entropy, such as peak apex and peak shoulder. This collaborative dynamic focusing of channel attention and spatial attention allows the network to proactively combat the "frequency bias" phenomenon at the model architecture level, concentrating limited computational resources on the discrimination of weak satellite peaks. This significantly improves the model's ability to capture rare weak satellite peaks and solves the problem of poor weak peak identification caused by class imbalance. Attached Figure Description

[0054] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of an embodiment of the invention.

[0056] Figure 2 A schematic diagram of the structure of an attention-enhanced convolutional neural network model according to an embodiment of the present invention.

[0057] Figure 3 This is a schematic diagram of the system structure of an embodiment of the present invention.

[0058] Figure 4 This is a schematic diagram of the structure of the electronic device created by this invention. Detailed Implementation

[0059] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0060] Example 1

[0061] Figure 1 The diagram illustrates an automatic identification method for the physical role of candidate peaks in XRD patterns, including:

[0062] Step S11: Obtain the discrete scattering intensity and angle data sequence output by the X-ray diffractometer.

[0063] The aforementioned discrete scattering intensity refers to the X-ray photon count detected by an X-ray diffraction (XRD) instrument at each angular position, reflecting the diffraction intensity information of the crystal to X-rays.

[0064] The aforementioned angle data sequence refers to the angle data sequence during X-ray diffraction (XRD) instrument scanning, and it corresponds one-to-one with discrete scattering intensities.

[0065] It should be noted that the discrete scattering intensity and angle data sequences are raw experimental measurement data directly collected from XRD, which retain the original physical signals and angle coordinate information, providing basic physical support for subsequent analysis.

[0066] Step S12: Perform preliminary screening of candidate peaks and local morphology window extraction on the angle data sequence. Perform logarithmic transformation on the discrete scattering intensity corresponding to each angle position within the local morphology window and concatenate it with the physical prior features of the angle data sequence to obtain a multi-channel feature spectrum matrix.

[0067] The above-mentioned preliminary screening of candidate peaks quickly locates the center angle position of all suspected candidate peaks in the global spectrum, which is used to reduce the processing scale of attention-enhanced convolutional neural network models.

[0068] The aforementioned local morphology window truncation refers to truncation of fixed-length spectral segments to the left and right with a preset step size, centered on the suspected candidate peak. This is used to transform the full-spectrum analysis problem into a local window analysis problem, enabling focus on each candidate peak and its corresponding local morphology features, thus avoiding global background interference.

[0069] The aforementioned physical prior features refer to the feature channels reflecting the physical essence of diffraction peaks, including the first derivative, second derivative, and relative position encoding. The first derivative represents the slope, the second derivative represents the curvature, and the relative position encoding channel is calculated based on the ratio of the angular offset of each angular position relative to the global reference main peak to the scanning step size.

[0070] By performing a logarithmic transformation on the discrete scattering intensity, the extreme dynamic range between the strong substrate peak and the weak satellite peak is compressed by a high order of magnitude, allowing the microscopic protrusion profile of the weak peak to be preserved. The synergy between the logarithmic transformation and the physical prior feature channel transforms the nonlinear peak shape features that are difficult to exhaust with static rules into structured data that can be effectively processed by neural networks, realizing the high-fidelity extraction and utilization of the local morphological features of the diffraction peak.

[0071] Step S13: Based on the multi-channel feature spectrum matrix, an attention-enhanced convolutional neural network model is used to perform feature extraction, dual dynamic focusing on channels and positions, and classification, and output the initial probability distribution of the physical role of each candidate peak; wherein, the attention-enhanced convolutional neural network model includes a multi-scale one-dimensional convolutional layer, a dual attention layer, and a classifier.

[0072] The aforementioned Attention-Augmented Convolutional Neural Network (AA-CNN PeakRoleNet) is a deep learning architecture that combines traditional convolutional neural networks (CNNs) with attention-augmented (AA) mechanisms. Its core design concept is to organically combine convolution operators with attention-augmentation mechanisms to overcome the limitation that single-scale convolutional kernels cannot simultaneously capture diffraction peak features of different widths, thereby significantly improving the model's feature extraction capabilities in complex visual or sequence tasks.

[0073] The aforementioned multi-scale one-dimensional convolutional layer includes a large-scale one-dimensional convolutional layer, a medium-scale one-dimensional convolutional layer, and a small-scale one-dimensional convolutional layer, with a cross-layer residual path set between the medium-scale one-dimensional convolutional layer and the small-scale one-dimensional convolutional layer.

[0074] The aforementioned dual attention layer refers to a sub-module consisting of channel attention and a spatial attention sub-module, used to recalibrate and dynamically focus the features output by multi-scale one-dimensional convolutional layers.

[0075] By introducing an attention enhancement mechanism, the attention-enhanced convolutional neural network model can adaptively adjust the weight coefficients of each physical channel based on the input samples. Under strong background interference, it automatically suppresses the intensity channel and amplifies the curvature channel, which plays a crucial role in identifying weak satellite peaks, thus maintaining high sensitivity to rare categories of weak satellite peaks. Simultaneously, a spatial attention mechanism forces the aggregated weights to focus on physical regions with high information entropy, such as peak tops and shoulders. The dual attention layer, through the collaborative dynamic focusing of the channel attention submodule and the spatial attention submodule, enables the network to proactively combat the "frequency bias" phenomenon at the model architecture level. This concentrates limited computational resources on the discrimination of weak satellite peaks, significantly improving the model's ability to capture rare weak satellite peaks and solving the problem of poor weak peak identification caused by class imbalance.

[0076] Step S14: Based on the crystallographic symmetry rules, perform a global consistency check on the initial probability distribution of the physical role of each candidate peak, and select the final physical role classification result of each candidate peak.

[0077] The aforementioned crystallographic symmetry rules refer to the Bragg law and symmetrical distribution rules that diffraction peaks in XRD patterns must follow.

[0078] The aforementioned global consistency verification refers to extracting the initial probability distribution of the physical role of each candidate peak and its angular coordinates. Based on the angular symmetry between ±1-level satellite peaks and 0-level peaks, a global grid search is used to lock the peak combination that maximizes the joint confidence score. The physical role of each candidate peak in the locked peak combination is then used as the final physical role classification result.

[0079] By performing a global consistency review, the prediction bias of the model under imbalanced sample conditions was effectively corrected, ensuring that the identification results conform to physical symmetry and significantly improving the accuracy and robustness of weak satellite peak identification.

[0080] Specifically, in step S11, the discrete scattering intensity and angle data sequence output by the X-ray diffractometer is obtained. Specifically, when the X-ray diffractometer generates a new pattern file, the system directly parses the original binary database from the underlying layer and extracts the original discrete scattering intensity and angle data sequence at a 1:1 scale. The binary format includes xrdml format.

[0081] It should be noted that the raw discrete scattering intensity and angle data sequence output by the X-ray diffractometer is not subjected to any form of coordinate axis interpolation, data alignment, or pre-smoothing filtering, ensuring that the physical background information and microscopic morphology information of the local spectral bands are absolutely and accurately reproduced from the data source.

[0082] Specifically, in step S12, candidate peaks are initially screened and local morphology windows are extracted from the angle data sequence. The discrete scattering intensities corresponding to each angle position within the local morphology window are logarithmically transformed and then concatenated with the physical prior features of the angle data sequence to obtain a multi-channel feature spectrum matrix, including:

[0083] Step S121: Perform preliminary screening of candidate peaks and local morphology window extraction on the angle data sequence.

[0084] In this embodiment, the differential peak finding algorithm or the threshold screening algorithm is used to screen candidate peaks in the angle data sequence in order to locate the center angle position of all suspected candidate peaks.

[0085] Specifically, when using the differential peak-finding algorithm for initial screening of candidate peaks, the position where the first-order difference of the angle data sequence changes from positive to negative is determined as the center angle position of the suspected candidate peak; when using the threshold initial screening algorithm for initial screening of candidate peaks, the angle position where the discrete scattering intensity is greater than the preset intensity threshold is determined as the center angle position of the suspected candidate peak.

[0086] Using the center angle of the suspected candidate peak as the center, and cropping to the left and right sides according to a preset step size, a local topography window is obtained, specifically represented as follows: ; In the formula, Indicates the first A local morphology window corresponding to a suspected candidate peak; This represents a discrete scattering intensity function, used to represent the original scattering intensity value at a specific angular position; Indicates the first The center angle position of a suspected candidate peak; This indicates the width of half the window, i.e., the preset step size; This represents the candidate peak number, and its value is a positive integer.

[0087] In this embodiment, the preset step size can be dynamically set according to the scanning step size and the typical peak width to ensure that the captured window can completely cover the shape of a diffraction peak, while avoiding the introduction of too much irrelevant background.

[0088] Step S122: Perform a logarithmic transformation on the discrete scattering intensity corresponding to each angular position within the local topography window, and concatenate it with the physical prior features of the angular data sequence to obtain a multi-channel feature spectrum matrix, including:

[0089] Step S1221: Perform a logarithmic transformation on the discrete scattering intensity corresponding to each angular position within the local topography window, and normalize the discrete scattering intensity after the logarithmic transformation to obtain the logarithmic scattering intensity value.

[0090] In practical superlattice analysis, the absolute intensity of a strong substrate peak may be several orders of magnitude higher than that of a weak satellite peak. To eliminate this extreme dynamic range difference, a logarithmic transformation is performed on the original discrete scattering intensity at each angular position within the local topography window to obtain the initial logarithmic scattering intensity value, specifically expressed as follows: ; In the formula, This represents the initial logarithmic scattering intensity value; Indicates angular position Discrete scattering intensity at the location; This represents a bias term used to prevent the occurrence of extremely small positive bias terms of zero in logarithmic calculations.

[0091] The initial logarithmic scattering intensity value is normalized to obtain the logarithmic scattering intensity value.

[0092] By performing logarithmic transformation and compressing the dynamic range, even extremely weak satellite peaks can still maintain a clear and discernible microscopic protrusion profile within the local morphology window, which can be learned by the network.

[0093] Step S1222: Perform differential calculation on the discrete scattering intensity corresponding to each angular position within the local topography window to obtain the first derivative feature and the second derivative feature.

[0094] In this embodiment, the first derivative is characterized by slope; the second derivative is characterized by curvature.

[0095] Step S1223: The angle position corresponding to the maximum discrete scattering intensity in the full spectrum is taken as the global reference main peak angle position, and the relative position code is calculated based on the ratio of the angle offset of each angle position in the local topography window relative to the global reference main peak angle position to the scanning step size.

[0096] The expression for the relative position encoding is: ; In the formula, Indicates the first Within the local topography window, the first The relative position encoding of each data point; Indicates the first Within the local topography window, the first The actual angle value corresponding to each data point, i.e., the absolute coordinate; Indicates the global reference main peak angle position; This indicates the actual scanning step size of the X-ray diffractometer.

[0097] By introducing relative position encoding for directional prior, the attention-enhanced convolutional neural network model can automatically learn the spatial distance between candidate peaks and the global reference main peak. This helps the model determine whether the candidate peak is a closely overlapping 0th-order peak or a distant ±1-order satellite peak, reflecting crystallographic order. This "physical mechanism prior" design significantly reduces the model's dependence on massive amounts of labeled data, enabling it to achieve extremely fast and stable convergence even with very small experimental datasets.

[0098] Step S1224 involves concatenating the logarithmic scattering intensity values, first derivative features, second derivative features, and relative position codes along the channel dimension to obtain a multi-channel feature spectrum matrix.

[0099] In terms of channel dimension, a logarithmic scattering intensity channel is constructed using logarithmic scattering intensity values ​​and serves as the main channel for peak shape identification, preserving the basic geometric contours and local intensity gradients of candidate peaks. A first-order derivative channel is constructed using the slope characteristic of the first-order derivative to reflect the variation law of the tangent slope on both sides of the candidate peak. A second-order derivative channel is constructed using the curvature characteristic of the second-order derivative to characterize the microscopic curvature of the peak, effectively suppressing random noise points by leveraging the characteristic that real physical peaks have relatively stable negative curvature near the peak apex. A relative position encoding channel is constructed using relative position encoding to perform directional prior, enabling the network to automatically learn the spatial distance between the candidate peak and the reference main peak, thereby assisting in determining the physical role of the candidate peak.

[0100] The logarithmic scattering intensity channel, the first derivative channel, the second derivative channel, and the relative position encoding channel are concatenated to obtain a multi-channel feature spectrum matrix, specifically represented as follows: ; In the formula, Represents the multi-channel characteristic spectral segment matrix; Indicates the first Each local topography window corresponds to a logarithmic scattering intensity channel; Indicates the first The first derivative channel corresponding to each local topography window; Indicates the first The second derivative channels corresponding to each local topography window; Indicates the first The relative position encoding channel corresponding to each local shape window.

[0101] Specifically, in step S13, based on the multi-channel feature spectrum matrix, an attention-enhanced convolutional neural network model is used to perform feature extraction, dual dynamic focusing on channels and positions, and classification, and output the initial probability distribution of the physical role of each candidate peak.

[0102] Because the full width at half maximum (FWHM) and morphology of different peaks in the diffraction pattern vary greatly, this invention adopts an attention-enhanced convolutional neural network model with multi-level layers to gradually decouple the complex peak features from macroscopic to microscopic.

[0103] In this embodiment, the attention-enhanced convolutional neural network model includes a multi-scale one-dimensional convolutional layer, a dual attention layer, and a classifier. The multi-scale one-dimensional convolutional layer comprises a cascaded large-scale convolutional layer, a medium-scale convolutional layer, and a small-scale convolutional layer, with cross-layer residual connections between the medium-scale and small-scale convolutional layers. The dual attention layer comprises a channel attention layer and a spatial attention layer.

[0104] In a preferred embodiment, the kernel size of the large-scale convolutional layer is set to 7; the kernel size of the medium-scale convolutional layer is set to 5; and the kernel size of the small-scale convolutional layer is set to 3.

[0105] Specifically, an attention-enhanced convolutional neural network model is used for feature extraction, dual dynamic focusing on channels and locations, and classification, outputting the initial probability distribution of the physical role of each candidate peak, including:

[0106] Step S131: Input the multi-channel feature spectrum matrix into the multi-scale one-dimensional convolutional layer to extract multi-scale features and obtain multi-scale features.

[0107] In this embodiment, after the multi-channel feature spectral matrix enters the multi-scale one-dimensional convolutional layer, the wide receptive field of the large-scale convolutional layer is first used to capture the global peak width and long-range background slope tendency; then the mid-scale convolutional layer perceives the asymmetric distortion and continuity changes in the peak shoulder region; finally, the small-scale convolutional layer extracts the micro-morphology of the weak satellite peak and local high-frequency features with extremely high granularity, thereby obtaining multi-scale features.

[0108] Step S132: The output features of the medium-scale convolutional layer are superimposed onto the output features of the small-scale convolutional layer to obtain residual fusion features.

[0109] In this embodiment, by constructing a cross-layer residual connection between the mesoscale convolutional layer and the small-scale convolutional layer, the output features of the mesoscale convolutional layer are connected across layers and superimposed on the output features of the small-scale convolutional layer to obtain residual fusion features. This avoids the gradient vanishing phenomenon that easily occurs in the low-frequency signals of weak satellite peaks during multiple convolution operations, thus achieving physical protection of the bottom contour information of weak peaks and smoothly transmitting the high-fidelity feature flow to the downstream dual attention layer.

[0110] The expression for the residual fusion feature is as follows: ; In the formula, Indicates residual fusion characteristics; This represents the output features of a mid-scale convolutional layer; This represents the output features of a small-scale convolutional layer.

[0111] Step S133: Input the residual fusion features into the dual attention layer, perform channel attention recalibration and spatial attention recalibration to obtain the core feature vector, including:

[0112] Step S1331: Use global average pooling and multilayer perceptron to calculate the importance weights of each feature channel in the residual fusion features.

[0113] In this embodiment, firstly, global average pooling (GAP) is used to average the values ​​of all spatial points in each feature channel, compressing each feature channel into a single value to describe the overall response intensity of that feature channel. Secondly, the compressed value is input into a fully connected neural network of a multilayer perceptron (MLP), and by learning the dependencies between different feature channels, a new numerical vector is output. Finally, the sigmoid activation function is used to map the new numerical vector to a value between 0 and 1, forming the final importance weights for each feature channel.

[0114] Step S1332: The residual fusion features are recalibrated by multiplication channel by channel using importance weights to obtain the calibrated features.

[0115] The expression for the calibration feature is as follows: ; In the formula, Indicates calibration characteristics; Indicates the importance weight of the feature channels; This represents the Hadamard product, which is the product of corresponding elements in each channel.

[0116] By performing channel attention recalibration, when faced with extremely weak peaks severely contaminated by high-frequency steep slope background, the weight of the "logarithmic scattering intensity channel" can be automatically and significantly reduced, while the weight of the "second derivative channel," which is extremely sensitive to real convexities, is strongly amplified.

[0117] Step S1333: Calculate the normalized weights of each angle position within the local topography window.

[0118] In this embodiment, a spatial attention scorer is used to scan each angular position within the local topography window and outputs a score containing the probability of key information such as peaks and shoulders. After normalization, the normalized weights of each angular position within the local topography window are obtained. The spatial attention scorer is composed of a neural network submodule consisting of convolutional layers or fully connected layers.

[0119] Step S1334: The calibrated features are weighted using normalized weights to obtain the core feature vector.

[0120] The expression for the core feature vector is as follows: ; In the formula, Represents the core feature vector; Indicates angular position Normalized weights; Indicates angular position The calibration characteristics.

[0121] By using normalized weights to weight the labeled features, indiscriminate global average pooling is avoided. This forced attention-enhanced convolutional neural network model precisely locks its vision and computing power into physical regions with extremely high information entropy, such as "peaks, shoulders, or points of abrupt slope changes," thereby extracting the core feature vectors with the most decision-making value. .

[0122] Step S134: Input the core feature vector into the classifier and output the initial probability distribution of the physical role of each candidate peak.

[0123] In this embodiment, the classifier is a Softmax classifier, which outputs the initial probability distribution of the physical role of each candidate peak.

[0124] In a preferred embodiment, such as Figure 2 As shown, a multi-channel feature spectrum matrix of dimension (2m+1)×4 is input into an attention-enhanced convolutional neural network model for feature extraction, dual dynamic focusing on channels and positions, and classification, outputting the initial probability distribution of the physical role of each candidate peak. Here, 2m+1 represents the spatial length, and 4 represents the number of feature channels.

[0125] Specifically, based on the input multi-channel feature spectrum matrix with dimension (2m+1)×4, a large-scale convolutional layer (Conv1D_1) with kernel size k=7 and number of kernels c=32 is first used, combined with a batch normalization (BN) layer and a ReLU activation function, to output the output features of the large-scale convolutional layer. Its dimension is (2m+1)×32; secondly, a medium-scale convolutional layer (Conv1D_2) with kernel size k=5 and number of kernels c=64 is used, combined with batch normalization (BN) layer and ReLU activation function, to output the output features of the medium-scale convolutional layer. Its dimension is (2m+1)×64; then a small-scale convolutional layer (Conv1D_3) with kernel size k=3 and number of kernels c=64 is used, combined with batch normalization (BN) layer and ReLU activation function, to output the output features of the small-scale convolutional layer. Its dimensions are (2m+1)×64.

[0126] The output features of the mesoscale convolutional layer Output features of small-scale convolutional layers By performing element-wise addition, the residual fusion characteristics are obtained. Its dimensions are (2m+1)×64.

[0127] In the Channel Attention layer, features are first fused using residuals of dimension (2m+1)×64. Global average pooling (GAP) is performed, outputting a vector of length 64. Next, a first fully connected layer (FC) maps the 64-dimensional vector to r (r < 64), and a ReLU activation function is used to introduce non-linearity, outputting an intermediate vector of length r. Then, a second fully connected layer (FC) maps the intermediate vector of dimension r back to 64, and a Sigmoid activation function is used to restrict the output values ​​to (0,1), outputting a 1×64 channel weight vector as the importance weights. Finally, the importance weights are fused with the residual features. Perform element-wise multiplication to output a calibration feature with a dimension of (2m+1)×64.

[0128] In the Spatial Attention layer, the Tanh activation function is first used to perform spatial feature mapping on the calibration feature map of dimension (2m+1)×64, outputting a non-linear transformation feature of dimension (2m+1)×64. Next, a fully connected layer (FC) is used to map the non-linear transformation feature of dimension (2m+1)×64 to K (K<64), outputting a tensor of dimension (2m+1)×K. Then, a Softmax normalization layer is used to restrict the output value to (0,1), outputting a spatial weight vector of dimension (2m+1)×1 as the normalization weight. Finally, the normalized weight is used to perform element-wise multiplication of the calibration feature, outputting a core feature vector of dimension (2m+1)×64.

[0129] The calibrated features are added element-wise to the core feature vector to extract the 64-dimensional feature vector corresponding to the center position. Then, a fully connected layer is used to map it to 5 dimensions, and a Softmax classifier is used to calculate the initial probability distribution of the physical role of each candidate peak.

[0130] In this embodiment, the attention-enhanced convolutional neural network model is trained based on the total joint loss function, including:

[0131] Step S1341: Using the Focal Loss architecture, weight coefficients are set inversely proportional to the number of samples in each physical role category to construct a classification loss term.

[0132] By constructing classification loss terms It effectively combats the extreme class imbalance caused by the large number of noise peaks and the scarcity of satellite peak samples in the measured spectrum, and greatly improves the gradient contribution of rare satellite peaks during weight updates.

[0133] Step S1342: Based on the superlattice dynamics diffraction theory, a physical penalty loss term is constructed according to the angular distance between the satellite peak and the main substrate peak and the preset safe distance threshold.

[0134] According to the superlattice dynamics diffraction theory, due to the difference between the strain and lattice constant of the epitaxial layer, there must be a physically isolated, safe angular distance between each level of satellite peak and the main substrate peak in angular space. Therefore, based on the relationship between the angular distance and a preset safe distance threshold, a physical penalty loss term is constructed and configured to be activated when the angular distance is less than the preset safe distance threshold, generating a gradient penalty. This hard-codes crystallographic rules into the loss plane of the attention-enhanced convolutional neural network model, thereby preventing shoulder peak misjudgment from the underlying logic. The expression for the physical penalty loss term is: ;

[0135] In the formula, Indicates the physical penalty loss item; This indicates the preset safe distance threshold; This indicates the predicted angle of the satellite peak; This represents the true value of the satellite peak angle; Represents the maximum value function; This indicates the actual scanning step size of the X-ray diffractometer; Indicates the full width at half maximum (FWHM) of the substrate peak. Indicates the step size safety factor; This represents the peak width ratio coefficient.

[0136] In this embodiment, the step size safety factor ranges from 5 to 20; the peak width ratio factor ranges from 0.2 to 0.5.

[0137] It should be noted that the preset safety distance threshold is adaptively determined based on the X-ray diffractometer resolution, scan step size, and actual substrate peak width, and is used to limit the minimum allowable physical distance between the satellite peak and the substrate peak. Specifically, when the scan resolution is high, the scan step size is used to determine the safety distance; when the substrate peak is wide, a certain proportion of the substrate peak width is used to determine the safety distance, thereby ensuring that the preset safety distance threshold can adapt to different test conditions and different material systems.

[0138] In a preferred embodiment, a high-resolution X-ray diffractometer was used for testing, with a scan step size of 0.005°, a substrate peak half-width at half-maximum of approximately 0.32°, and a step size safety factor. Set to 10, peak width ratio coefficient If we take 0.4, then we can calculate the preset safe distance threshold. It is 0.128°.

[0139] Step S1343: Introduce an attention regularization term, and then sum the attention regularization term, classification loss term, and physical penalty loss term in a weighted manner to construct a joint loss function.

[0140] The expression for the joint loss function is as follows: ; In the formula, Represents the joint loss function; Represents the classification loss term; The weights of the attention regularization term are indicated. This represents an attention regularization term; This represents the weight of the physical penalty loss term.

[0141] By introducing an attention regularization term, attention-enhanced convolutional neural network models are forced to form spatially compact rather than diffuse high-dimensional feature focal points.

[0142] Step S1344: Using the joint loss function as the optimization objective, iteratively train the attention-enhanced convolutional neural network model until the attention-enhanced convolutional neural network model converges.

[0143] Specifically, in step S14, based on the crystallographic symmetry rules, the initial probability distribution of the physical role of each candidate peak is globally consistent and checked to select the final physical role classification result of each candidate peak.

[0144] In this embodiment, the final physical role classification results of the candidate peaks include substrate peak, 0th order peak, -1st order satellite peak, +1st order satellite peak, and noise peak.

[0145] Specifically, a global consistency review is performed on the initial probability distribution of physical roles for each candidate peak, and the final physical role classification results for each candidate peak are selected, including:

[0146] Step S141: Extract the initial probability distribution of the physical role of each candidate peak, determine the candidate peak sets for the -1 level satellite peak, the +1 level satellite peak and the 0 level peak, and obtain the candidate -1 level satellite peak, the candidate +1 level satellite peak and the candidate 0 level peak.

[0147] Step S142: Based on the angular coordinates of each candidate peak, calculate the first angular distance between the candidate -1 level satellite peak and the candidate 0 level peak, and the second angular distance between the candidate +1 level satellite peak and the candidate 0 level peak, and calculate the asymmetric distance factor based on the deviation between the first angular distance and the second angular distance.

[0148] The expression for the asymmetric distance factor is: ; In the formula, Indicates the asymmetric distance factor; Indicates the first The actual angular position of each candidate-1 level satellite peak; Indicates the first The actual angular positions of the candidate +1 level satellite peaks; Indicates the first The actual angular position of the candidate 0th order peak.

[0149] By introducing an asymmetric distance factor as an exponential penalty term, the deviation of the absolute distance between the candidate -1 level satellite peak and the candidate 0 level peak, and between the candidate +1 level satellite peak and the candidate 0 level peak is quantified.

[0150] Step S143: Calculate the joint confidence score based on the initial probability distributions of the -1 level satellite peak, the +1 level satellite peak, and the 0 level peak, as well as the asymmetric distance factor.

[0151] The expression for the joint confidence score is as follows: ; In the formula, Indicates the joint confidence score; Indicates the first The candidate peaks are the initial probabilities of -1 level satellite peaks; Indicates the first The candidate peak is the initial probability of a +1 order satellite peak; Indicates the first The initial probability that each candidate peak is a 0th-order peak; Indicates the spatial sensitivity control factor; , , Indicates the index of different candidate peaks.

[0152] Step S144: Locate the peak combination that maximizes the joint confidence score through global search, and use the physical role of each candidate peak in the peak combination as the final physical role classification result.

[0153] In this embodiment, a set of peak combinations that maximizes the joint confidence score is identified using a global search algorithm. This serves as the final physical role classification result.

[0154] By using an end-to-end physical role inference mechanism, the processing bottlenecks and subjective errors caused by human intervention are eliminated, enabling high-throughput unmanned analysis of material characterization, thereby greatly shortening the data feedback cycle in the R&D process of semiconductor epitaxial devices.

[0155] Example 2

[0156] In this embodiment, an automatic identification method for the physical role of candidate peaks in XRD patterns is provided, and the effectiveness of the method is verified through actual experiments.

[0157] In a high-resolution X-ray diffraction (HRXRD) test batch, the system captured a weak diffraction feature signal at approximately 31.1° to the left of the main peak of the substrate. This feature signal was deeply embedded within the high-intensity asymmetric tail slope of the main peak of the substrate.

[0158] Due to strong background interference, the traditional pseudo-Voigt full-spectrum fitting algorithm gets stuck in local optima during iterations. Calculations show that its waveform stripping confidence drops sharply to 34.1%. In conventional automated chip fabrication quality control, such low-confidence signals are easily misjudged as background noise and filtered out, leading to the loss of crucial physical information.

[0159] Weak diffraction feature signals are input into the attention-enhanced convolutional neural network model (AA-CNN PeakRoleNet), triggering the dual attention layers to perform collaborative work on channel attention recalibration and spatial attention recalibration.

[0160] Specifically, the importance weights of feature channels are... The dynamic suppression is reduced to 0.15 to decrease the interference weight of background noise. Simultaneously, based on the inherent negative curvature a priori response at the actual diffraction peak, the attention-enhanced convolutional neural network model forcibly amplifies the gain of the weight coefficients in the second derivative channel to 0.88 to highlight the peak features of weak signals from the strong background. A spatial attention scorer is used to precisely converge the peak response of the aggregated weights to the anchor point of the abrupt change in local curvature. This mechanism forces the attention allocation to the peripheral, irrelevant trailing background to zero, achieving accurate segmentation of the feature region.

[0161] The preset safety threshold was set to 0.15°, and the weight of the physical penalty loss term was set to 0.5. In the actual measurement, the angular distance between the candidate peak and the main substrate peak was 0.5°, which is greater than the preset safety distance threshold, thus avoiding the huge gradient penalty. At the same time, the candidate peak has a high-fidelity underlying curvature protrusion and exhibits a significant spatial geometric symmetry response with the +1 order correlation peak on the right.

[0162] Ultimately, the joint decision-making architecture confirmed that the distortion feature was a "-1 level satellite peak" without ignoring strong background trailing interference. The overall judgment confidence level of the system output was as high as 98.5%, which verified that the invention has excellent robustness in accurately extracting extremely weak physical signals in harsh physical backgrounds after removing human subjective intervention, and solved the technical problem that traditional fitting algorithms are prone to failure under complex working conditions.

[0163] Example 3

[0164] Based on the same design concept, such as Figure 3 As shown, this embodiment provides an automatic identification system for the physical role of candidate peaks in XRD patterns, applied to the method for automatic identification of the physical role of candidate peaks in XRD patterns, including:

[0165] The data acquisition module is used to acquire discrete scattering intensity and angle data sequences output by the X-ray diffractometer.

[0166] The preprocessing module is used to perform candidate peak screening and local morphology window extraction on the angle data sequence. It performs logarithmic transformation on the discrete scattering intensity corresponding to each angle position within the local morphology window and concatenates it with the physical prior features of the angle data sequence to obtain a multi-channel feature spectrum matrix.

[0167] The feature extraction module is used to perform feature extraction, dual dynamic focusing on channels and positions, and classification based on a multi-channel feature spectrum matrix using an attention-enhanced convolutional neural network model, and outputs the initial probability distribution of the physical role of each candidate peak; the attention-enhanced convolutional neural network model includes a multi-scale one-dimensional convolutional layer, a dual attention layer, and a classifier;

[0168] The verification and output module is used to perform a global consistency verification of the initial probability distribution of the physical role of each candidate peak based on the crystallographic symmetry law, and to select the final physical role classification result of each candidate peak.

[0169] In practical applications, the output results of the verification and output modules are physically tagged and then transmitted to the control terminal.

[0170] Example 4

[0171] Embodiments of the present invention also provide a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of the present invention.

[0172] Example 5

[0173] Embodiments of the present invention also provide a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the method of an embodiment of the present invention.

[0174] Example 6

[0175] An embodiment of the present invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of the embodiment of the present invention.

[0176] refer to Figure 4 The present invention will now describe a structural block diagram of an electronic device that can serve as an embodiment of the present invention, serving as an example of a hardware device applicable to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0177] like Figure 4 As shown, the electronic device includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the electronic device. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0178] Multiple components in the electronic device are connected to I / O interface 405, including: input unit 406, output unit 407, storage unit 408, and communication unit 409. Input unit 406 can be any type of device capable of inputting information into the electronic device. Input unit 406 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 408 may include, but is not limited to, disks and optical discs. Communication unit 409 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0179] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as computer programs tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0180] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0181] In the context of embodiments of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0182] It should be noted that the term "comprising" and its variations used in the embodiments of this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of this invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0183] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are subject to strict compliance with relevant laws, regulations, and regulatory requirements in their collection, storage, use, processing, transmission, provision, and disclosure, and adhere to the principles of legality, legitimacy, necessity, and good faith. The acquisition of relevant information and data is premised on the user's explicit consent or other legitimate reasons, and a clear and convenient authorization management approach is provided to the user, allowing the user to independently choose to consent, withdraw consent, or refuse to provide relevant information. For functions that rely on user information, if the user does not authorize or withdraws authorization, the corresponding technical function cannot be implemented, and the technical solution of this invention is not applicable in this scenario. The steps described in the method implementation embodiments provided by the embodiments of this invention can be executed in different orders and / or in parallel. Furthermore, the method implementation embodiments may include additional steps and / or omit the steps shown. The scope of protection of this invention is not limited in this respect.

[0184] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0185] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for automatically identifying the physical role of candidate peaks in XRD patterns, characterized in that, include: Obtain the discrete scattering intensity and angle data sequence output by the X-ray diffractometer; Candidate peaks are initially screened and local morphology windows are extracted from the angle data sequence. The discrete scattering intensities corresponding to each angle position within the local morphology window are logarithmically transformed and concatenated with the physical prior features of the angle data sequence to obtain a multi-channel feature spectrum matrix. Based on the multi-channel feature spectrum matrix, an attention-enhanced convolutional neural network model is used for feature extraction, dual dynamic focusing on channels and positions, and classification, outputting the initial probability distribution of the physical role of each candidate peak; wherein, the attention-enhanced convolutional neural network model includes a multi-scale one-dimensional convolutional layer, a dual attention layer, and a classifier; Based on the symmetry rules of crystallography, the initial probability distribution of the physical role of each candidate peak is globally consistent and verified to select the final classification result of the physical role of each candidate peak.

2. The method for automatic identification of the physical role of candidate peaks in XRD patterns according to claim 1, characterized in that, The preliminary screening of candidate peaks and local morphology window extraction of the angle data sequence includes: The differential peak finding algorithm is used to determine the center angle position of the suspected candidate peak by the position where the first difference of the angle data sequence changes from positive to negative. Using the center angle position of the suspected candidate peak as the center, a local morphology window is obtained by truncating according to a preset step size.

3. The method for automatic identification of the physical role of candidate peaks in XRD patterns according to claim 1, characterized in that, The preliminary screening of candidate peaks and local morphology window extraction of the angle data sequence includes: A threshold screening algorithm is used to determine the angular positions where the discrete scattering intensity is greater than a preset intensity threshold as the center angular positions of the suspected candidate peaks; Using the center angle position of the suspected candidate peak as the center, a local morphology window is obtained by truncating according to a preset step size.

4. The method for automatic identification of the physical role of candidate peaks in XRD patterns according to claim 1, characterized in that, The process of performing a logarithmic transformation on the discrete scattering intensities corresponding to each angular position within the local topography window and concatenating them with the physical prior features of the angular data sequence to obtain a multi-channel feature spectrum matrix includes: Logarithmic transformation is performed on the discrete scattering intensity corresponding to each angular position within the local topography window, and the logarithmically transformed discrete scattering intensity is normalized to obtain the logarithmic scattering intensity value. The discrete scattering intensity corresponding to each angular position within the local topography window is calculated using a difference method to obtain the first-order derivative feature and the second-order derivative feature. The angular position corresponding to the maximum discrete scattering intensity in the full spectrum is taken as the global reference main peak angular position, and the relative position code is calculated based on the ratio of the angular offset of each angular position in the local topography window relative to the global reference main peak angular position to the scanning step size. The logarithmic scattering intensity value, the first derivative feature, the second derivative feature, and the relative position code are concatenated along the channel dimension to obtain a multi-channel feature spectrum matrix.

5. The method for automatic identification of the physical role of candidate peaks in XRD patterns according to claim 1, characterized in that, Based on the multi-channel feature spectral matrix, an attention-enhanced convolutional neural network model is used for feature extraction, dual dynamic focusing on channels and positions, and classification, outputting the initial probability distribution of the physical role of each candidate peak, including: The multi-channel feature spectrum matrix is ​​input into a multi-scale one-dimensional convolutional layer for multi-scale feature extraction to obtain multi-scale features; wherein, the multi-scale one-dimensional convolutional layer includes a large-scale convolutional layer, a medium-scale convolutional layer, and a small-scale convolutional layer, and the medium-scale convolutional layer and the small-scale convolutional layer are connected across layer residuals. The output features of the mid-scale convolutional layer are superimposed onto the output features of the small-scale convolutional layer to obtain residual fusion features; The residual fusion features are input into a dual attention layer for channel attention recalibration and spatial attention recalibration to obtain the core feature vector. The core feature vector is input into the classifier, which outputs the initial probability distribution of the physical role of each candidate peak.

6. The method for automatic identification of the physical role of candidate peaks in XRD patterns according to claim 5, characterized in that, The residual fusion features are input into a dual attention layer for channel attention recalibration and spatial attention recalibration to obtain the core feature vector, including: The importance weights of each feature channel in the residual fusion feature are calculated using global average pooling and multilayer perceptron, respectively. The residual fusion features are recalibrated channel by channel using the importance weights to obtain the calibrated features; Calculate the normalized weights of each angular position within the local topography window; The calibrated features are weighted using the normalized weights to obtain the core feature vector.

7. The method for automatic identification of the physical role of candidate peaks in XRD patterns according to claim 1, characterized in that, The attention-enhanced convolutional neural network model is trained based on the total joint loss function and includes: The Focal Loss architecture is adopted, and the weight coefficients are set inversely proportional to the number of samples in each physical role category to construct the classification loss term; Based on the superlattice dynamics diffraction theory, a physical penalty loss term is constructed according to the angular distance between the satellite peak and the main substrate peak and a preset safety distance threshold; wherein, the physical penalty loss term is configured to be activated when the angular distance is less than the preset safety distance threshold, thereby generating a gradient penalty; An attention regularization term is introduced, and the attention regularization term, the classification loss term, and the physical penalty loss term are weighted and summed to construct a joint loss function; Using the joint loss function as the optimization objective, the attention-enhanced convolutional neural network model is iteratively trained until it converges.

8. The method for automatic identification of the physical role of candidate peaks in XRD patterns according to claim 1, characterized in that, The final physical role classification results of the candidate peaks include substrate peaks, 0th-order peaks, -1st-order satellite peaks, +1st-order satellite peaks, and noise peaks; Based on crystallographic symmetry, a global consistency check is performed on the initial probability distribution of the physical role of each candidate peak to screen out the final physical role classification results of each candidate peak, including: Extract the initial probability distribution of the physical role of each candidate peak, determine the candidate peak sets for the -1 level satellite peak, the +1 level satellite peak and the 0 level peak, and obtain the candidate -1 level satellite peak, the candidate +1 level satellite peak and the candidate 0 level peak; Based on the angular coordinates of each candidate peak, calculate the first angular distance between the candidate -1 level satellite peak and the candidate 0 level peak, and the second angular distance between the candidate +1 level satellite peak and the candidate 0 level peak, and calculate the asymmetric distance factor based on the deviation between the first angular distance and the second angular distance. The joint confidence score is calculated based on the initial probability distributions of the -1 level satellite peak, the +1 level satellite peak, and the 0 level peak, as well as the asymmetric distance factor. The peak combination that maximizes the joint confidence score is identified through a global search, and the physical role of each candidate peak in the peak combination is used as the final physical role classification result.

9. An automatic identification system for the physical role of candidate peaks in XRD patterns, applied to the automatic identification method for the physical role of candidate peaks in XRD patterns according to any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire discrete scattering intensity and angle data sequences output by the X-ray diffractometer. The preprocessing module is used to perform candidate peak screening and local morphology window extraction on the angle data sequence, perform logarithmic transformation on the discrete scattering intensity corresponding to each angle position within the local morphology window, and concatenate it with the physical prior features of the angle data sequence to obtain a multi-channel feature spectrum matrix. The feature extraction module is used to perform feature extraction, dual dynamic focusing on channels and positions, and classification based on the multi-channel feature spectrum matrix using an attention-enhanced convolutional neural network model, and output the initial probability distribution of the physical role of each candidate peak; wherein, the attention-enhanced convolutional neural network model includes a multi-scale one-dimensional convolutional layer, a dual attention layer, and a classifier; The verification and output module is used to perform a global consistency verification of the initial probability distribution of the physical role of each candidate peak based on the crystallographic symmetry law, and to select the final physical role classification result of each candidate peak.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.

11. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images