CVD diamond lattice size prediction method and device, equipment and storage medium

Abstract

The invention discloses a CVD diamond lattice size prediction method, device and equipment and a storage medium, and the method comprises the steps: taking information data of various CVD diamonds with known lattice sizes, and generating an ultraviolet Raman spectrum database according to the information data; by taking the information data as training data, training through a convolutional neural network to obtain multi-dimensional features of the ultraviolet Raman spectrum on Raman peaks affecting the lattice size of the CVD diamond; after multi-dimensional feature extraction, according to different linetypes, carrying out peak-splitting fitting on Raman peaks in each ultraviolet Raman spectrum through a set fitting function, and judging a main peak influencing the lattice size; calculating the lattice size of the CVD diamond by adopting a phonon confinement model; judging whether a judgment decision coefficient R2 of the model is greater than a preset value, and if not, adjusting a preset parameter of the fitting function; and outputting a predicted value of the lattice size of the to-be-tested CVD diamond through the model. According to the method, the accuracy of predicting the lattice size of the CVD diamond can be improved.

Description

Technical Field

[0001] The present invention relates to the field of chemical technology, and in particular to a CVD diamond lattice size prediction method, device, equipment and storage medium. Background Art

[0002] Diamond, with its excellent broadband semiconductor properties, is expected to become a candidate material for next-generation power devices. Advances in chemical vapor deposition (CVD) diamond film growth technology are enabling the composition and quality requirements of diamond films to be further refined. The lattice size of CVD diamond is directly related to its structure as a semiconductor material, its voltage / power range, its epitaxial growth, and its ease of doping. Furthermore, varying lattice mismatches directly correspond to dislocation densities, which in turn affect diamond's potential performance as a broadband semiconductor device. Therefore, obtaining high-precision CVD diamond lattice size data is crucial.

[0003] In the prior art, the lattice size of CVD diamond is usually measured using a scanning electron microscope (SEM) or a transmission electron microscope (TEM). In this way, the lattice size data is determined by measuring the diagonal length of the grains in the electron microscope photograph and taking the average value.

[0004] After research, the inventors found that the existing technology has at least the following defects:

[0005] The measurement methods in the prior art often result in inaccurate lattice dimensions due to errors in the selection of electron microscope photographs and human measurement.

[0006] The information disclosed in this background technology section is only intended to increase understanding of the overall background of the present invention and should not be regarded as an admission or any form of suggestion that the information constitutes prior art already known to those skilled in the art. Summary of the Invention

[0007] The purpose of the present invention is to improve the measurement accuracy of CVD diamond lattice size.

[0008] The present invention provides a method for predicting CVD diamond lattice size, comprising the steps of:

[0009] S11. Acquire information data containing multiple CVD diamonds of known lattice sizes, and generate an ultraviolet Raman spectrum database based on the information data; the information data includes a spectrum of the ultraviolet Raman spectrum of each CVD diamond, as well as label information of its corresponding lattice size and peak line shape convolution;

[0010] S12, using the information data as training data, training a convolutional neural network to obtain multi-dimensional features of the ultraviolet Raman spectrum for Raman peaks that affect the lattice size of CVD diamond;

[0011] S13, after the multi-dimensional features are extracted, the Raman peaks in each of the ultraviolet Raman spectra are subjected to peak fitting by a set fitting function according to different line shapes, and the main peak that affects the lattice size is determined;

[0012] S14, using a phonon confinement model, calculating the lattice size of the CVD diamond based on the feedback of the main peak peak displacement, half-peak width and lattice size on the peak signal strength;

[0013] S15, obtaining the accuracy of the model for lattice size regression of CVD diamond, and determining whether the coefficient of determination R2 is greater than a preset value. If not, backpropagating the result to step S13, and then adjusting the preset parameters of the fitting function;

[0014] S16. Using the ultraviolet Raman spectrum data of the CVD diamond to be tested as input, the model trained in steps S12-S15 is used to output a predicted value of the lattice size of the CVD diamond to be tested.

[0015] In another aspect of the present invention, a CVD diamond lattice size prediction device is provided, comprising:

[0016] The method comprises: a model building unit and a prediction unit; wherein the model building unit comprises:

[0017] A database module is configured to acquire information data of CVD diamonds of various known lattice sizes and generate an ultraviolet Raman spectrum database based on the information data; the information data includes a spectrum of the ultraviolet Raman spectrum of each CVD diamond, as well as label information of its corresponding lattice size and peak line shape convolution;

[0018] A multi-dimensional feature acquisition module is used to use the information data as training data to train a convolutional neural network to obtain multi-dimensional features of the ultraviolet Raman spectrum for Raman peaks that affect the lattice size of CVD diamond;

[0019] A main peak determination module is used to perform peak fitting on the Raman peaks in each of the ultraviolet Raman spectra according to different line shapes using a set fitting function after multi-dimensional feature extraction, and to determine the main peak that affects the lattice size;

[0020] a lattice size calculation module for calculating the lattice size of the CVD diamond using a phonon confinement model based on the main peak displacement, half-peak width, and feedback of the lattice size on the peak signal strength;

[0021] A model correction module is used to obtain the accuracy of the phonon confinement model for the lattice size regression of CVD diamond and determine whether the determination coefficient R2 is greater than a preset value. If not, the result is back-propagated to the main peak judgment module to adjust the preset parameters of the fitting function;

[0022] The prediction unit is used to take the ultraviolet Raman spectrum data of the CVD diamond to be tested as input, and output a predicted value of the lattice size of the CVD diamond to be tested through the model trained by the model building unit.

[0023] In another aspect of the present invention, there is also provided a CVD diamond lattice size prediction device, comprising:

[0024] Memory for storing computer programs;

[0025] A processor is used to call and execute the computer program to implement each step of the CVD diamond lattice size prediction method as described in any one of the above items.

[0026] In another aspect of the embodiment of the present invention, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, each step of the CVD diamond lattice size prediction method as described above is implemented.

[0027] The CVD diamond lattice size prediction device includes a computer program stored on a medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the methods described in the above aspects and achieves the same technical effects.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] To address the limitations of existing techniques for determining CVD diamond lattice size, such as human error and measurement difficulties, this paper develops a method for predicting CVD diamond lattice size based on a convolutional neural network. Because UV Raman spectroscopy is sensitive for characterizing carbon materials, its patterns contain a wealth of information, facilitating the extraction of multidimensional features using convolutional neural networks. Consequently, compared to methods like SEM and XRD, which require human intervention or suffer from poor signal strength, this method offers greater accuracy and versatility.

[0030] Specifically, the present invention establishes an ultraviolet Raman spectrum database containing CVD diamond, and trains a convolutional neural network on this basis. By utilizing the powerful computing power of computers for image recognition, the multidimensional features of the Raman peaks in the spectrum can be fully extracted, solving the problem of peak interference in the ultraviolet Raman spectrum.

[0031] In the present invention, during the establishment of the convolutional neural network, multiple convolution kernels are used to divide the convolution layers, which is different from the existing method of using a single convolution kernel. The advantage of this method is that it cleverly utilizes the characteristics of the ultraviolet spectrum to pool a large number of pixels unrelated to the Raman peak. At the same time, convolution kernels of different dimensions are used to convolve the Raman peak to extract the multidimensional characteristics of the main peak that affects the lattice size. This can greatly simplify the number of parameters of the convolutional neural network and avoid the loss of spectral information caused by dimensionality reduction.

[0032] Finally, the present invention utilizes the phonon confinement model in the calculation method of the lattice size, which can be better used in the parameter adjustment process of linear fitting, so that the peak separation results obtained by the convolutional neural network can be visualized, thereby improving the accuracy of predicting the CVD diamond lattice size.

[0033] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and to implement it according to the contents of the specification, and to make the above and other purposes, technical features and advantages of the present invention easier to understand, one or more preferred embodiments are listed below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0035] Figure 1 is a step diagram of the CVD diamond lattice size prediction method of the present invention;

[0036] Figure 2 Schematic diagram of the structure of the CVD diamond lattice size prediction device of the present invention;

[0037] Figure 3 It is a schematic diagram of the structure of the CVD diamond lattice size prediction equipment described in the present invention. DETAILED DESCRIPTION

[0038] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

[0039] Unless expressly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising” will be understood to include the stated elements or components but not to exclude other elements or components.

[0040] In this document, the terms "first", "second", etc. are used to distinguish two different elements or parts, and are not used to limit specific positions or relative relationships. In other words, in some embodiments, the terms "first", "second", etc. can also be interchangeable with each other.

[0041] Example 1

[0042] In order to improve the measurement accuracy of CVD diamond lattice size, such as Figure 1 As shown, in an embodiment of the present invention, a method for predicting CVD diamond lattice size is provided, comprising the steps of:

[0043] S11. Establishing a UV Raman spectrum database containing CVD diamonds, including the UV Raman spectrum of each type of CVD diamond, and label information of its corresponding lattice size and peak line shape convolution;

[0044] In an embodiment of the present invention, the UV Raman spectrum database stores spectrum results of CVD diamonds of known lattice sizes under UV light irradiation of different wavelengths, that is, a collection of UV Raman spectra of multiple CVD diamond samples under UV light irradiation of different wavelengths; in this collection, each spectrum corresponds to label information of lattice size and peak line shape convolution;

[0045] In an embodiment of the present invention, the label information of the lattice size can be the average value of the lattice size or the collection of the distribution of the lattice size; for the label information of the peak line shape convolution, since each peak in the ultraviolet Raman spectrum can essentially be regarded as the result of a convolution operation of different peak line shapes, the peak line shape can be Gaussian, Lorentzian and Voigt line shapes, which is also the classification basis of the convolutional neural network in an embodiment of the present invention;

[0046] S12. Obtaining spectrum data from a database, inputting it into a convolutional neural network and training it, including setting hyperparameters and arranging convolutional layers, pooling layers, and fully connected layers, to obtain multidimensional features of the UV Raman spectrum for Raman peaks that affect the lattice size of CVD diamonds;

[0047] In an embodiment of the present invention, all spectral data in the database need to be input into a convolutional neural network for training to obtain multidimensional features related to the diamond lattice size. Through the arrangement of convolutional layers, pooling layers and fully connected layers, the hyperparameters can be effectively iterated to obtain more targeted feature variables.

[0048] In the embodiment of the present invention, Figure 2 As shown, a schematic diagram of the distribution of the convolutional layer, pooling layer and fully connected layer of the convolutional neural network is provided, including the steps of establishing and training the convolutional neural network:

[0049] S21, establishing a convolution layer C1, using a 3×3 convolution kernel to extract preliminary multidimensional features of the Raman peaks in the input CVD diamond UV Raman spectrum;

[0050] In practical applications, a diamond UV Raman spectrum can be viewed as a matrix containing N×M pixels, where N×M represents the pixel resolution of the spectrum. Because UV Raman spectra lack color, each pixel can be represented using a 256-level grayscale representation, i.e., [0, 255], where 0 represents complete black and 255 represents complete white. To better represent this within a convolutional neural network, in this embodiment of the present invention, the 256-level grayscale representation is converted into a digital signal in the interval [-1, 1], where -1 represents complete white and 1 represents complete black.

[0051] In practical applications, a 3×3 convolution kernel is used. The initial shape of the convolution kernel is determined by the hyperparameters. Then, during the iteration process, the shape of the convolution kernel will change according to the iteration results by the computer modifying the hyperparameters automatically without human intervention.

[0052] In practical applications, since the pixels in the UV Raman spectrum are all represented in grayscale, in the embodiment of the present invention, the number of input channels is 1, and therefore the number of channels of the convolution kernel is also 1;

[0053] In practical applications, the number of convolution kernels in the convolution layer C1 depends on the number of output channels, which mainly depends on the complexity of the UV Raman spectrum and the computing power of the computer. Therefore, it is not limited here.

[0054] In this embodiment of the present invention, all UV Raman spectrum samples in the database are input into the convolutional layer C1 of the convolutional neural network to extract preliminary multidimensional features of Raman peaks in the UV Raman spectrum of CVD diamond. The size of the output image of the convolutional layer C1 can be expressed as:

[0055]

[0056]

[0057] Where W represents the image width of the UV Raman spectrum, H represents the image height of the UV Raman spectrum, S represents the step size of the convolution kernel, P represents the number of boundary pixel layers required to fill the image boundary with zeros, input represents input, output represents output, and filter represents the hyperparameters of the convolution kernel.

[0058] In this embodiment of the present invention, all UV Raman spectrum samples in the database are input into the convolutional layer C1 of the convolutional neural network to extract preliminary multidimensional features of Raman peaks in the UV Raman spectrum of CVD diamond. The output of the convolutional layer C1 is expressed as:

[0059]

[0060] Among them, X c1 represents the output of convolutional layer C1, X1 represents the single channel input of convolutional layer C1, and w i represents the weight of the i-th convolution kernel, and b represents the bias term.

[0061] S22, using the activation function to modify the preliminary multidimensional features extracted by the convolution layer C1;

[0062] In practical applications, it is necessary to use an activation function to correct the output result of the convolution layer C1. The activation function used can be one of Sigmoid, Softmax, Tanh, ReLU, LeakyReLU, ELU, SELU, and Softplus functions. In the embodiment of the present invention, preferably, the ReLU function is used, and its specific equation is as follows:

[0063] ReLU(x c1 )=max(x c1 ,0);

[0064] In an embodiment of the present invention, since the input ultraviolet Raman spectrum contains a large number of blank pixels other than the spectrum signal, that is, there are a large number of pixels whose digital signals are -1. These pixels are not helpful for the feature extraction of the Raman peak. Therefore, it is necessary to simplify the result through an activation function, which will help improve the response speed of subsequent steps.

[0065] S23. Based on the correction results of the preliminary multidimensional features obtained in S22, a convolution layer C2 is established; according to the feature distribution, convolution operations are performed on the groups using convolution kernels of different sizes to further extract the multidimensional features of the Raman peaks in the input CVD diamond UV Raman spectrum;

[0066] In practical applications, the input to convolutional layer C2 is the multidimensional features extracted by convolutional layer C1, so the relative position of the Raman peak can be roughly determined. However, if the Raman peak has a large span in the image, a convolution kernel that is too small can easily lose the receptive field. Therefore, continuing to use the same 3×3 convolution kernel is not reasonable. However, using a large number of larger convolution kernels would result in an excessively large amount of hyperparameter data, which is not conducive to calculation.

[0067] Therefore, in step S22, due to the action of the activation function, a large number of pixel points are processed into digital signals less than 0, which reduces the amount of calculation to a certain extent (using the preferred ReLU function, a large number of pixel points can be processed into digital signals equal to 0, and since 0 multiplied by any number is 0, the amount of calculation is lower than that of other activation functions). The embodiment of the present invention also needs to set a reasonable grouping to plan the areas responsible for different-sized convolutional kernels.

[0068] In the embodiment of the present invention, grouping can be set to achieve the convolutional calculation of different-sized convolutional kernels in the convolutional layer C2, and further extract multi-dimensional features in CVD diamond; specifically, 4 grouping (group) forms can be set, namely (3×3), (3×3, 5×5), (3×3, 5×5, 7×7), (3×3, 5×5, 7×7, 9×9), that is, the computer will autonomously select the combination method of different-sized convolutional kernels according to the iteration result, that is, it can only use the 3×3 convolutional kernel for convolution, or use three convolutional kernels of 3×3, 5×5, and 7×7 for convolution, and no specific limitation is made here.

[0069] In the embodiment of the present invention, the output of the convolutional layer C2 is expressed as:

[0070]

[0071] S24. Establish a pooling layer P1 to reduce the dimension of the data features extracted in the convolutional layer C2;

[0072] In the embodiment of the present invention, since the number of convolutional kernels depends on the complexity of the Raman spectrum diagram, after two layers of convolution of C1 and C2, the amount of data may be large. And according to the previous steps, there are pixel points in these data that are not very helpful for the extraction of Raman peak features. Therefore, it is necessary to reduce the dimension of these pixels through the pooling layer P1 to ensure the smooth operation of the convolutional neural network.

[0073] In the embodiment of the present invention, the pooling layer P1 adopts the max pooling (Max Pooling) method, that is, for a pooling area of N×N, remember the index position of the maximum value, and assign the maximum value to the area reduced to 1×1 (1 < N) after pooling. The advantage of doing this is that during backpropagation, the eigenvalue can be filled to the index position of the original maximum value. In this way, for the area where the digital signal itself is 0, the area after pooling can be greatly reduced in dimension; and for the area containing Raman peak signals, the spectral information will not be lost due to the dimension reduction of pooling.

[0074] S25. Based on the data features after dimensionality reduction by the pooling layer P1, convolution layers C3 and C4 are established, using 1×N and N×1 convolution kernels respectively to extract the multidimensional features of the Raman peak in the input CVD diamond UV Raman spectrum in terms of peak height and half-maximum width;

[0075] In practical applications, since the information of Raman peak peak height and half-maximum width has an important influence on the judgment of diamond structure and lattice size, it is necessary to set up two additional convolution layers to extract the multi-dimensional features of Raman peak peak height and half-maximum width.

[0076] In an embodiment of the present invention, since the peak height and peak width have large changes on the horizontal and vertical axes compared to the entire spectrum, it is necessary to additionally use convolution kernels of 1×N and N×1 sizes for calculation, where the 1×N convolution kernel is mainly used to discover wide features, and the N×1 convolution kernel is mainly used to discover high features.

[0077] In this embodiment of the present invention, the size of the output image of the convolutional layer C3 can be expressed as:

[0078]

[0079] H output =1

[0080] In this embodiment of the present invention, the size of the output image of the convolutional layer C4 can be expressed as:

[0081] W output =1

[0082]

[0083] In this embodiment of the present invention, the output of the convolutional layer C3 is expressed as:

[0084]

[0085] In this embodiment of the present invention, the output of the convolutional layer C4 is expressed as:

[0086]

[0087] S26, using an activation function to correct the multi-dimensional features of peak height and half-peak width extracted by convolutional layers C3 and C4;

[0088] In practical applications, an activation function needs to be used to correct the output results of convolutional layers C3 and C4. The activation function used can be one of the Sigmoid, Softmax, Tanh, ReLU, LeakyReLU, ELU, SELU, and Softplus functions. In the embodiments of the present invention, preferably, the ReLU function is used, and its specific equation is as follows:

[0089] ReLU(x c3 ) = max(x c3 , 0);

[0090] ReLU(x c4 ) = max(x c4 , 0);

[0091] S27. Establish a pooling layer P2 to reduce the dimensionality of the data features extracted from convolutional layer C4;

[0092] In the embodiments of the present invention, since the number of convolutional kernels depends on the complexity of the Raman spectrum, after two convolutional layers C3 and C4, the amount of data may be large. Therefore, it is necessary to reduce the dimensionality of these pixels through pooling layer P2 to ensure the smooth operation of the convolutional neural network.

[0093] In the embodiments of the present invention, the pooling layer P2 adopts the stochastic pooling method. That is, for a pooling region of N×N, it is randomly selected probabilistically according to the magnitude of its eigenvalue. That is, the larger the eigenvalue, the greater the probability of being selected, and the selected value is assigned to the region whose dimensionality is reduced to 1×1 (1 < N) after pooling. The advantage of this is that stochastic pooling does not select those pixel points with eigenvalue 0 that are not helpful for extracting the multi-dimensional features of Raman peaks; at the same time, the generalization ability of this method is stronger by selecting according to probability, which can better prevent overfitting; in addition, during the backpropagation process, similar to the max pooling method, it can well record the position and value of the selected nodes and ensure the consistency of features.

[0094] S28. Establish a fully connected layer A1, adopt an activation function, and perform a full permutation classification on the eigenvalues input into the fully connected layer A1 to determine the probabilities of the peak types of different peaks.

[0095] In the embodiments of the present invention, there can be several fully connected layers. Since the number of parameters doubles for each additional fully connected layer, the number of fully connected layers is determined by the computing power of the specific computer and is not limited here;

[0096] In this embodiment of the present invention, the fully connected layer is used to integrate the multidimensional features obtained in the previous steps. That is, according to different weight values, the feature values ​​are multiplied and summed to predict the probability of the peak line shape of each Raman peak. Therefore, the final output result is the classification result of the peak line shape with the highest probability.

[0097] In the embodiment of the present invention, the activation function may be one of Sigmoid, Softmax, Tanh, ReLU, LeakyReLU, ELU, SELU, and Softplus functions. In the embodiment of the present invention, preferably, the Softmax function is used, and its specific equation is as follows:

[0098]

[0099] Among them, x ji Indicates the probability that the j-th Raman peak in the spectrum output by the fully connected layer of the last layer belongs to the i-th line type, Softmax(x ji ) represents the probability that the j-th Raman peak in the spectrum obtained by the last fully connected layer after the activation function softmax processing belongs to the i-th line type.

[0100] S29. Obtain the accuracy of the model for line type classification through the loss function, and determine whether the F1 value is greater than the preset value; if not, modify the convolution kernel and its corresponding hyperparameters used in steps S21-S28 through backpropagation; if yes, the result with the maximum peak type probability is used as the output result of the training.

[0101] In the embodiment of the present invention, the calculation standard of the loss function can be calculated using mean square error (MSE) or cross entropy. In the embodiment of the present invention, preferably, cross entropy is used for calculation, and its specific equation is as follows:

[0102]

[0103] in, represents the predicted probability that the j-th Raman peak in the spectrum belongs to the i-th line type, y ji It represents the probability that the j-th Raman peak in the spectrum belongs to the i-th line type.

[0104] In the embodiment of the present invention, a confusion matrix is ​​used to judge the accuracy of line type classification. By comparing the classification results and the actual predicted values, the prediction ability of the model for a certain category can be intuitively reflected. The specific expression is as follows:

[0105] Predicted value = (G, L, V) line type Predicted value = non-(G,L,V) linear True value = (G, L, V) line type TP FN True value = non-(G,L,V) linear FP TN

[0106] Among them, TP refers to the number of true samples that are correctly classified, that is, they are predicted to be true and are actually true; FP refers to the number of false samples that are incorrectly marked as true samples, that is, they are actually false but predicted to be true; TN refers to the number of false samples that are correctly classified, that is, they are predicted to be false and are actually false; FN refers to the number of true samples that are incorrectly marked as false samples, that is, they are actually true but predicted to be false.

[0107] Specifically, when predicting the Gauss line type, when the actual value is the Gauss line type and the predicted value is also the Gauss line type, it is classified as TP; when the actual value is the Gauss line type and the predicted value is the Lorentz line type or the Voigt line type, it is classified as FN; when the actual value is the Lorentz line type or the Voigt line type and the predicted value is the Gauss line type, it is classified as FP; when the actual value is the Lorentz line type or the Voigt line type and the predicted value is the Lorentz line type or the Voigt line type, it is classified as TN; it is worth noting that the classification of TN depends on the currently specified category: for example, when the actual value is the Lorentz line type and the predicted value is the Voigt line type, it is also classified as TN, that is, although the model predicts the actual classification incorrectly, from the perspective of whether it is the Gauss line type, the model's judgment is not wrong.

[0108] In the embodiment of the present invention, the F1 value is an important indicator for measuring classification problems and a key parameter for judging the quality of training results. It is the harmonic mean of precision (Pre) and recall (Recall). The closer it is to 1, the better the prediction effect, and the closer it is to 0, the worse the prediction effect. Precision (Pre) is the proportion of correctly classified true samples in the predicted true samples, and recall (Recall) is the proportion of correctly classified true samples in the actual true samples. Their specific calculation formulas are as follows:

[0109]

[0110]

[0111]

[0112] In the embodiment of the present invention, the preset value of the F1 value is generally not less than 0.9, preferably 0.95;

[0113] S13. After multi-dimensional feature extraction, the Raman peaks in each spectrum are separated and fitted according to different line shapes, and the main peak that affects the lattice size is determined;

[0114] In the embodiment of the present invention, by training the convolutional neural network, the peak separation results of each Raman peak in the spectrum can be obtained, and the peak line shape of each Raman peak can be obtained at the same time. Different operations will be taken for fitting for different line shapes.

[0115] When the peak line shape is a Gaussian line shape, the curve_fit function in the scipy library in Python language can be used for fitting. The specific formula of the Gaussian line shape is as follows:

[0116]

[0117] When the peak line shape is a Lorentz line shape, the polyfit function in the numpy library in Python language can be used for fitting. The specific formula of the Lorentz line shape is as follows:

[0118]

[0119] When the peak line shape is a Voigt line shape, since the Voigt line shape can actually be regarded as the result of the convolution of the Gaussian line shape and the Lorentz line shape, the convolve function in the numpy library in the Python language can be used for convolution calculation, or the wofz function in the scipy library in the Python language can be used for fitting, and the final line shape result can be fitted.

[0120] Since the above-mentioned line fitting method belongs to the prior art, it is not described in detail in the embodiment of the present invention.

[0121] In the embodiment of the present invention, the basis for judging the main peak that affects the lattice size is mainly to look at the shift of the peak position. Generally, when the displacement of the Raman peak is 1333 cm -1 When , this peak can be considered as the main peak affecting the lattice size.

[0122] S14. Using the phonon confinement model, the lattice size of the diamond is calculated using the feedback of the main peak displacement, half-peak width and lattice size on the peak signal strength;

[0123] In the embodiment of the present invention, the phonon confinement model is used to describe the simple multiplication of phonons in a plane wave and a Gaussian function. Therefore, the integrate.quad function in the scipy library in the Python language can be used to solve L in the integrand, where L is the lattice size of the diamond. The specific formula is as follows:

[0124]

[0125] ω(q)=1193.75+139.25cos(qπ)

[0126] Γ=α+β / L

[0127] Where I(ω) represents the signal intensity of the UV Raman spectrum, ω(q) represents the approximate one-dimensional phonon dispersion curve, and Γ represents the half-maximum width of the main peak. α and β are coefficients related to the UV Raman spectroscopy equipment, where α is related to the resolution (linewidth) of the Raman spectroscopy equipment and β is related to the material used for the Raman spectroscopy detection equipment. Both of the above can be calculated using the instruction manual provided with the UV Raman spectroscopy equipment.

[0128] S15. Obtain the accuracy of the model for lattice size regression of CVD diamond and determine the coefficient of determination R 2 Is it greater than a preset value? If not, the result is back-propagated to step S13 to adjust the relevant parameters of the fitting function;

[0129] In the embodiment of the present invention, the coefficient of determination R 2 It is the key parameter to judge the effectiveness of training, and the determination coefficient R 2 The closer it is to 1, the better the prediction effect, and the closer it is to 0, the worse the prediction effect. The specific calculation formula is as follows:

[0130]

[0131] in represents the predicted value of the CVD diamond lattice size calculated after training, y i represents the actual value of the CVD diamond lattice size in the database, and n represents the number of UV Raman spectra in the database.

[0132] In practical applications, the coefficient of determination R 2 Should be no less than 0.8, it is considered that the prediction effect of the model is good. In the embodiment of the present invention, the determination coefficient R 2 The default value is 0.85.

[0133] S16. Inputting the UV Raman spectrum data of the CVD diamond to be tested into the model trained through steps S12-S15, and outputting the predicted value of the lattice size of the CVD diamond;

[0134] In summary, to address the shortcomings of existing techniques for obtaining CVD diamond lattice size, such as human error and measurement difficulties, the present invention establishes a method for predicting CVD diamond lattice size based on a convolutional neural network. Because UV Raman spectroscopy is sensitive for characterizing carbon materials, its spectra contain a wealth of information, which facilitates the extraction of multidimensional features using convolutional neural networks. Therefore, compared with methods such as SEM and XRD, which require human intervention or have poor signal strength, this method offers greater accuracy and versatility.

[0135] Specifically, the embodiment of the present invention establishes an ultraviolet Raman spectrum database containing CVD diamond, and trains a convolutional neural network on this basis. By utilizing the powerful computing power of computers for image recognition, the multidimensional features of the Raman peaks in the spectrum can be fully extracted, solving the problem of peak interference in the ultraviolet Raman spectrum.

[0136] In the embodiment of the present invention, during the establishment of the convolutional neural network, multiple convolution kernels are used to divide the convolution layers, which is different from the existing method of using a single convolution kernel. The advantage of this method is that it cleverly utilizes the characteristics of the ultraviolet spectrum to pool a large number of pixels unrelated to the Raman peak. At the same time, convolution kernels of different dimensions are used to convolve the Raman peak to extract the multidimensional characteristics of the main peak that affects the lattice size. This can greatly simplify the number of parameters of the convolutional neural network and avoid the loss of spectral information caused by dimensionality reduction.

[0137] Finally, in the calculation method of the lattice size, the embodiment of the present invention utilizes the phonon confinement model, which can be better used in the parameter adjustment process of linear fitting, so that the peak separation results obtained by the convolutional neural network can be visualized, thereby improving the accuracy of predicting the CVD diamond lattice size.

[0138] Example 2

[0139] Corresponding to the method embodiment, another aspect of the present invention further provides a CVD diamond lattice size prediction device. Figure 3 The schematic diagram of the structure of the CVD diamond lattice size prediction device provided by the embodiment of the present invention is shown. Figure 1 The device corresponding to the CVD diamond lattice size prediction method described in the corresponding embodiment, that is, is implemented by means of a virtual device Figure 1 In the corresponding embodiment of the CVD diamond lattice size prediction method, each virtual module constituting the CVD diamond lattice size prediction device can be executed by an electronic device, such as a network device, a terminal device, or a server. Specifically, the CVD diamond lattice size prediction device in the embodiment of the present invention includes:

[0140] Model building unit 01 and prediction unit 02; wherein, model building unit 01 includes:

[0141] The database module 11 is configured to acquire information data of CVD diamonds of various known lattice sizes and generate a UV Raman spectrum database based on the information data; the information data includes a UV Raman spectrum of each CVD diamond, as well as label information of its corresponding lattice size and peak line shape convolution;

[0142] A multi-dimensional feature acquisition module 12 is used to use the information data as training data to train a convolutional neural network to obtain multi-dimensional features of the ultraviolet Raman spectrum for Raman peaks that affect the lattice size of CVD diamond;

[0143] The main peak judgment module 13 is used to perform peak fitting on the Raman peaks in each of the ultraviolet Raman spectra according to different line shapes using a set fitting function after multi-dimensional feature extraction, and to judge the main peak that affects the lattice size;

[0144] a lattice size calculation module 14 for calculating the lattice size of the CVD diamond using a phonon confinement model based on the main peak displacement, half-peak width, and feedback of the lattice size on the peak signal strength;

[0145] The model correction module 15 is used to obtain the accuracy of the phonon confinement model for the lattice size regression of CVD diamond and determine whether the determination coefficient R2 is greater than a preset value. If not, the result is back-propagated to the main peak judgment module to adjust the preset parameters of the fitting function.

[0146] The prediction unit 02 is used to take the ultraviolet Raman spectrum data of the CVD diamond to be tested as input, and output a predicted value of the lattice size of the CVD diamond to be tested through the model trained by the model building unit.

[0147] It should be noted that the specific implementation method and technical effects of the CVD diamond lattice size prediction device in the embodiment of the present invention can be referred to Figure 1 The corresponding CVD diamond lattice size prediction method will not be described here in detail.

[0148] Example 3

[0149] Corresponding to the method embodiments, embodiments of the present invention also provide a CVD diamond lattice size prediction device, such as a terminal or server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be, but is not limited to, a smartphone, tablet computer, laptop computer, or desktop computer.

[0150] An example of a hardware structure block diagram of a CVD diamond lattice size prediction device provided in an embodiment of the present application is shown in FIG. Figure 3 As shown, this may include:

[0151] Processor 1, communication interface 2, memory 3 and communication bus 4;

[0152] The processor 1, the communication interface 2, and the memory 3 communicate with each other via the communication bus 4;

[0153] Optionally, the communication interface 2 may be an interface of a communication module, such as an interface of a GSM module;

[0154] The processor 1 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0155] The memory 3 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0156] The processor 1 is specifically configured to execute the computer program stored in the memory 3 to perform the following steps:

[0157] S11. Acquire information data containing multiple CVD diamonds of known lattice sizes, and generate an ultraviolet Raman spectrum database based on the information data; the information data includes a spectrum of the ultraviolet Raman spectrum of each CVD diamond, as well as label information of its corresponding lattice size and peak line shape convolution;

[0158] S12, using the information data as training data, training a convolutional neural network to obtain multi-dimensional features of the ultraviolet Raman spectrum for Raman peaks that affect the lattice size of CVD diamond;

[0159] S13, after the multi-dimensional features are extracted, the Raman peaks in each of the ultraviolet Raman spectra are subjected to peak fitting by a set fitting function according to different line shapes, and the main peak that affects the lattice size is determined;

[0160] S14, using a phonon confinement model, calculating the lattice size of the CVD diamond based on the feedback of the main peak peak displacement, half-peak width and lattice size on the peak signal strength;

[0161] S15, obtaining the accuracy of the phonon confinement model for the lattice size regression of CVD diamond, and determining whether the determination coefficient R2 is greater than a preset value. If not, backpropagating the result to step S13, and then adjusting the preset parameters of the fitting function;

[0162] S16. Using the ultraviolet Raman spectrum data of the CVD diamond to be tested as input, the model trained in steps S12-S15 is used to output a predicted value of the lattice size of the CVD diamond to be tested.

[0163] The above-mentioned product can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the CVD diamond lattice size prediction method provided by the embodiment of the present invention.

[0164] Example 4

[0165] In an embodiment of the present invention, a storage medium is further provided. The storage medium may store a program suitable for execution by a processor, wherein the program is used to:

[0166] S11. Acquire information data containing multiple CVD diamonds of known lattice sizes, and generate an ultraviolet Raman spectrum database based on the information data; the information data includes a spectrum of the ultraviolet Raman spectrum of each CVD diamond, as well as label information of its corresponding lattice size and peak line shape convolution;

[0167] S12, using the information data as training data, training a convolutional neural network to obtain multi-dimensional features of the ultraviolet Raman spectrum for Raman peaks that affect the lattice size of CVD diamond;

[0168] S13, after the multi-dimensional features are extracted, the Raman peaks in each of the ultraviolet Raman spectra are subjected to peak fitting by a set fitting function according to different line shapes, and the main peak that affects the lattice size is determined;

[0169] S14, using a phonon confinement model, calculating the lattice size of the CVD diamond based on the feedback of the main peak peak displacement, half-peak width and lattice size on the peak signal strength;

[0170] S15, obtaining the accuracy of the phonon confinement model for the lattice size regression of CVD diamond, and determining whether the determination coefficient R2 is greater than a preset value. If not, backpropagating the result to step S13, and then adjusting the preset parameters of the fitting function;

[0171] S16. Using the ultraviolet Raman spectrum data of the CVD diamond to be tested as input, the model trained in steps S12-S15 is used to output a predicted value of the lattice size of the CVD diamond to be tested.

[0172] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0173] The above-mentioned product can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the methods provided by other embodiments of the present invention.

[0174] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0175] In the several embodiments provided herein, it should be understood that the disclosed systems, devices, and methods may be implemented in other ways. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interface, device, or unit, and may be electrical, mechanical, or other means.

[0176] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0177] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0178] It should be understood that in the embodiments of the present application, the various embodiments and features can be combined with each other to solve the aforementioned technical problems.

[0179] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0180] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting CVD diamond lattice size, characterized in that: Including steps: S11. Acquire information data containing multiple CVD diamonds of known lattice sizes, and generate an ultraviolet Raman spectrum database based on the information data; the information data includes a spectrum of the ultraviolet Raman spectrum of each CVD diamond, as well as label information of its corresponding lattice size and peak line shape convolution; S12, using the information data as training data, training a convolutional neural network to obtain multi-dimensional features of the ultraviolet Raman spectrum for Raman peaks that affect the lattice size of CVD diamond; S13, after the multi-dimensional features are extracted, the Raman peaks in each of the ultraviolet Raman spectra are subjected to peak fitting by a set fitting function according to different line shapes, and the main peak that affects the lattice size is determined; S14, using a phonon confinement model, calculating the lattice size of the CVD diamond based on the feedback of the main peak peak displacement, half-peak width and lattice size on the peak signal strength; S15, obtaining the accuracy of the phonon confinement model for the lattice size regression of CVD diamond, and determining whether the determination coefficient R2 is greater than a preset value. If not, backpropagating the result to step S13, and then adjusting the preset parameters of the fitting function; S16. Using the ultraviolet Raman spectrum data of the CVD diamond to be tested as input, the model trained in steps S12-S15 is used to output a predicted value of the lattice size of the CVD diamond to be tested.

2. The CVD diamond lattice size prediction method according to claim 1, characterized in that: The label information of the lattice size includes: The average value of the lattice size, or the collection of distribution data of the lattice size.

3. The CVD diamond lattice size prediction method according to claim 2, characterized in that: The method of using the information data as training data and performing training through a convolutional neural network includes: S21, establishing a convolution layer C1, using a 3×3 convolution kernel to extract preliminary multidimensional features of the Raman peak in the input spectrum; S22, modifying the preliminary multidimensional features using an activation function; S23. Based on the correction result of the preliminary multidimensional features, a convolution layer C2 is established; and according to the feature distribution, convolution operations are performed in groups using convolution kernels of different sizes to extract the multidimensional features of the Raman peaks in the input spectrum; S24, establish the pooling layer P1, and reduce the dimension of the data features extracted by the convolution layer C2; S25. Based on the data features after dimensionality reduction by the pooling layer P1, a convolution layer C3 and a convolution layer C4 are established, and 1×N and N×1 convolution kernels are used respectively to extract the multidimensional features of the peak height and half-maximum width of the Raman peak in the input spectrum; S26, using an activation function to correct the multi-dimensional features of peak height and half-maximum width extracted by convolutional layer C3 and convolutional layer C4; S27, establish a pooling layer P2 to reduce the dimension of the data features extracted from the convolutional layers C3 and C4; S28. Establish a fully connected layer A1, and use an activation function to perform full permutation classification on the feature values ​​input into the fully connected layer A1 to determine the probabilities of peak types of different peaks; S29. Obtain the accuracy of the model for line type classification through the loss function, and determine whether the F1 value is greater than the preset value; if not, modify the convolution kernels and their corresponding hyperparameters adopted in steps S21 - S28 through backpropagation; if so, use the result with the maximum peak type probability as the output result of training.

4. The CVD diamond lattice size prediction method according to claim 3, characterized in that: The extraction of the preliminary multi - dimensional features of Raman peaks in the input spectrum includes: The spectrogram is input into the convolution layer C1 of the convolutional neural network to extract the preliminary multi - dimensional features of Raman peaks in the CVD diamond ultraviolet Raman spectrum. The size of the output image of the convolution layer C1 can be expressed as: Where, W represents the image width of the ultraviolet Raman spectrum, H represents the image height of the ultraviolet Raman spectrum, S represents the stride of the convolution kernel, P represents the number of boundary pixel layers required for padding 0 to the image boundary, input represents the input, output represents the output, and filter represents the hyperparameter of the convolution kernel; The output of the convolution layer C1 is expressed as: X c1 =∑ i w i X1+b; Among them, X c1 represents the output of convolutional layer C1, X1 represents the single channel input of convolutional layer C1, and w i represents the weight of the i-th convolution kernel, and b represents the bias term.

5. The CVD diamond lattice size prediction method according to claim 4, characterized in that: The use of an activation function to correct the preliminary multi - dimensional features includes: The activation functions used include Sigmoid, Softmax, Tanh, ReLU, LeakyReLU, ELU, SELU or Softplus functions; Further, preferably, using the ReLU function to correct the preliminary multi - dimensional features includes: The equation of the ReLU function includes: ReLU(x c1 )=max(x c1 ,0).

6. The CVD diamond lattice size prediction method according to claim 5, characterized in that: The extraction of the multi - dimensional features of Raman peaks in the input spectrum includes: Determine the grouping method of convolution kernels with different sizes according to the results of iteration; the grouping methods of the convolution kernels include: group1: (3×3), group2: (3×3, 5×5), group3: (3×3, 5×5, 7×7), group4: (3×3, 5×5, 7×7, 9×9); The output of the convolution layer C2 is expressed as:

7. The CVD diamond lattice size prediction method according to claim 6, characterized in that: Establish a pooling layer P1 and perform dimensionality reduction on the data features extracted in the convolution layer C2, including: The pooling layer P1 adopts the maximum pooling (Max Pooling) method, including: for a pooling region of N×N, remember the index position of the maximum value and assign this maximum value to the region after pooling and dimensionality reduction to 1×1 (1 < N).

8. The CVD diamond lattice size prediction method according to claim 7, characterized in that: The extraction of the multi - dimensional features of Raman peaks in the peak height and full - width at half - maximum of the CVD diamond ultraviolet Raman spectrum input includes: Use convolution kernels with sizes of 1×N and N×1 for calculation, where the 1×N convolution kernel is used to discover wide features, and the N×1 convolution kernel is used to discover high features; The size of the output image of the convolution layer C3 is expressed as: H output ＝1； The size of the output image of the convolution layer C4 is expressed as: W output ＝1； The output of the convolution layer C3 is expressed as: X c3 =∑ i w i (1×N) X3+b; The output of the convolution layer C4 is expressed as: X c4 =∑ i w i (N×1) X4+b。 9. The CVD diamond lattice size prediction method according to claim 8, characterized in that: The use of an activation function to correct the multi - dimensional features of peak height and full - width at half - maximum extracted in the convolution layers C3 and C4 includes: The activation function used is the ReLU function, and the specific equation is as follows: ReLU(x c3 )=max(x c3 ,0); ReLU(x c4 )=max(x c4 ,0).

10. The CVD diamond lattice size prediction method according to claim 9, characterized in that: Establish a pooling layer P2 to perform dimensionality reduction on the data features extracted in the convolution layer C4, including: The pooling layer P2 adopts the Stochastic Pooling method, which includes: for a pooling region of N×N, randomly select according to the probability of the magnitude of its feature values. The larger the feature value, the greater the probability of being selected, and the selected value is assigned to the region with dimensionality reduction to 1×1 (1 < N) after pooling.

11. The method for predicting CVD diamond lattice size according to claim 10, wherein: The establishment of the fully connected layer A1 uses an activation function to perform a full permutation classification on the feature values input into the fully connected layer A1 to determine the probabilities of the peak types of different peaks, including: The activation function includes one of the Sigmoid, Softmax, Tanh, ReLU, LeakyReLU, ELU, SELU, Softplus functions; Furthermore, preferably, the activation function uses the Softmax function, and its specific equation is as follows: Among them, x ji Indicates the probability that the j-th Raman peak in the spectrum output by the fully connected layer of the last layer belongs to the i-th line type, Softmax(x ji ) represents the probability that the j-th Raman peak in the spectrum obtained by the last fully connected layer after the activation function softmax processing belongs to the i-th line type.

12. The CVD diamond lattice size prediction method according to claim 11, characterized in that: The obtaining of the accuracy of the model for linear classification through the loss function includes: The calculation standard of the loss function is calculated using the mean squared error (MSE) or cross-entropy (Crossentropy); the specific equation for calculating the cross-entropy is as follows: in, represents the predicted probability that the j-th Raman peak in the spectrum belongs to the i-th line type, y ji It represents the probability that the j-th Raman peak in the spectrum belongs to the i-th line type.

13. The CVD diamond lattice size prediction method according to claim 12, characterized in that: The S29. Obtain the accuracy of the model for linear classification through the loss function, and judge whether the F1 value is greater than the preset value; if not, modify the convolution kernels and their corresponding hyperparameters adopted in steps S21 - S28 through backpropagation; if so, use the result with the largest peak type probability as the output result of the training. In the embodiments of the present invention, the calculation standard of the loss function can be calculated using the mean squared error (MSE) or cross-entropy (Crossentropy). In the embodiments of the present invention, preferably, cross-entropy is used for calculation, and its specific equation is as follows: in, represents the predicted probability that the j-th Raman peak in the spectrum belongs to the i-th line type, y ji It represents the probability that the j-th Raman peak in the spectrum belongs to the i-th line type.

14. A CVD diamond lattice size prediction device, characterized in that: including: a model establishment unit and a prediction unit; wherein, the model establishment unit includes: a database module for obtaining information data containing CVD diamonds with various known lattice size sizes, and generating an ultraviolet Raman spectroscopy database according to the information data; the information data includes the spectral maps of the ultraviolet Raman spectra of each CVD diamond, as well as the label information of its corresponding lattice size and peak line type convolution; a multi-dimensional feature acquisition module for using the information data as training data and training through a convolutional neural network to obtain the multi-dimensional features of the Raman peaks of the ultraviolet Raman spectroscopy that affect the lattice size of CVD diamonds; a main peak judgment module for, after multi-dimensional feature extraction, fitting the Raman peaks in each of the ultraviolet Raman spectra through a set fitting function according to different line types, and judging the main peak that affects the lattice size; a lattice size calculation module for using the phonon confinement model to calculate the lattice size of the CVD diamond according to the feedback of the main peak peak shift, full width at half maximum, and lattice size on the peak signal strength; a model correction module for obtaining the accuracy of the phonon confinement model for the lattice size regression of CVD diamonds, and judging whether the coefficient of determination R2 is greater than the preset value. If not, backpropagate the result to the main peak judgment module to adjust the preset parameters of the fitting function. The prediction unit is used to take the ultraviolet Raman spectrum data of the CVD diamond to be tested as input, and output a predicted value of the lattice size of the CVD diamond to be tested through the model trained by the model building unit.

15. A CVD diamond lattice size prediction device, characterized in that: include: memory for storing computer programs; A processor is configured to call and execute the computer program to implement the steps of the CVD diamond lattice size prediction method according to any one of claims 1 to 13.

16. A storage medium, characterized in that The method comprises a software program adapted to execute, by a processor, the steps of the CVD diamond lattice size prediction method according to any one of claims 1 to 13.

Images