Carbon nanotube defect recognition classification method based on multi-modal machine learning method

Abstract

The application discloses a carbon nanotube defect identification and classification method based on a multi-modal machine learning method, and comprises the following steps: obtaining an atomic coordinate file, a microscopic image and a Raman spectrum of a carbon nanotube, forming a multi-source data set, and labeling a defect type label for defective data; pre-processing the multi-source data set; constructing a three-branch hybrid deep learning model composed of a multilayer perception machine, a gated recurrent unit and a convolutional neural network, and respectively processing atomic structure data, a Raman spectrum and a microscopic image; training and performance evaluating the model by using the data set; and identifying and classifying carbon nanotube defects by using the optimized model. The method can efficiently and accurately identify carbon nanotube defects.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross technical field of machine learning technology engineering application and high-performance nanomaterial development, and particularly relates to a carbon nanotube defect identification and classification method based on a multi-modal machine learning method. BACKGROUND

[0002] Carbon nanotubes (CNT) are widely used in material science, electronics, energy storage and composite materials due to their unique one-dimensional tubular structure and excellent mechanical, electrical and thermal properties, and have become one of the hotspots in modern nanotechnology research. The macroscopic properties of CNTs are fundamentally determined by their atomic-scale microstructure. However, it is almost impossible to obtain perfect CNT samples in the experimental preparation process, and various structural defects inevitably occur during the growth process, such as intrinsic Stone-Wales defects, vacancy defects, and doping defects caused by foreign atoms. These defects can have a significant impact on the performance of CNTs. In some cases, even a small amount of defects can fundamentally change their mechanical, electrical and thermal properties. Therefore, accurate identification and classification of CNT defects, and defect compensation and control based on this, are key links to improve material performance and optimize preparation processes.

[0003] Traditional defect analysis methods, such as image analysis based on high-resolution transmission electron microscopy (HRTEM) or scanning tunneling microscopy (STM), and characteristic peak analysis for Raman spectroscopy, largely depend on the subjective experience and manual operation of the operator. These methods are not only inefficient, but also difficult to extract stable and quantifiable defect features from complex background noise, especially unable to meet the demand for rapid and objective statistical analysis of a large amount of sample data. Machine learning technology, especially deep learning, has shown great power in image recognition, spectral analysis and structural data mining. However, existing technologies are mostly limited to a single data modality, failing to fully utilize the complementary information between multi-source characterization data, resulting in limited classification accuracy and generalization ability. Therefore, the present application proposes a systematic carbon nanotube defect classification method based on multi-modal machine learning to solve the bottlenecks of traditional methods in efficiency, accuracy and large-scale application. SUMMARY

[0004] The purpose of the present application is to provide a carbon nanotube defect identification and classification method based on a multi-modal machine learning method, which can efficiently and accurately identify carbon nanotube defects.

[0005] The technical solution adopted by the present application is:

[0006] The carbon nanotube defect identification and classification method based on the multi-modal machine learning method is implemented according to the following steps:

[0007] Step 1, obtain the atomic coordinate file of carbon nanotube, microscopic image and Raman spectrum, form a multi-source data set, and label the defect type label for the defective data;

[0008] Step 2, pre-process the multi-source data set obtained in step 1;

[0009] Step 3, construct a three-branch hybrid deep learning model composed of multilayer perception, gated recurrent unit and convolutional neural network to process atomic structure data, Raman spectrum and microscopic image respectively;

[0010] Step 4, train and evaluate the model using the data set obtained in step 2;

[0011] Step 5, use the optimized model to identify and classify the defects of carbon nanotubes.

[0012] The features of the application are:

[0013] Step 1 is specifically:

[0014] Different chiral and size carbon nanotube models are established by molecular simulation software, intrinsic defects are simulated to be introduced, and corresponding atomic coordinate files are generated. At the same time, the microscopic image and Raman spectrum data of the real carbon nanotube sample are obtained by experimental means, and the simulation data and experimental data are integrated to form a multi-source data set containing atomic structure, Raman spectrum characteristics and microscopic image characteristics, and each defective data is labeled with a defect type label.

[0015] Step 2 is specifically:

[0016] The atomic coordinate file is parsed and the geometric topological features are calculated to generate a numerical vector; the Raman spectrum is smoothed and denoised, baseline corrected and sequence normalized; the microscopic image is uniformly scaled and gray normalized.

[0017] In step 3, the three-branch hybrid deep learning model is composed of parallel structure feature processing branch, spectrum feature extraction branch, image feature extraction branch, and feature fusion and joint classification layer; wherein the structure feature processing branch uses multilayer perception to encode high-dimensional geometric topological features; the spectrum feature extraction branch uses the gated recurrent unit network to extract sequence features from the spectrum data; the image feature extraction branch uses the convolutional neural network to extract spatial features from the microscopic image; the feature fusion and joint classification layer fuses the feature vectors output by the three feature extraction branches to form a joint feature vector, which is sequentially passed through the full connection layer and the output layer to output the probability distribution of all defects.

[0018] The structure of the structure feature processing branch is:

[0019] It includes a fully connected layer 1 and a fully connected layer 2; the fully connected layer 1 contains 128 neurons, uses the ReLU activation function, and is followed by a Dropout layer; the fully connected layer 2 contains 64 neurons, uses the ReLU activation function; the final output is a structural feature vector.

[0020] The structure of the spectral feature extraction branch is as follows:

[0021] It consists of two layers of gated recurrent unit networks: the first layer GRU contains 64 neurons and returns the entire output sequence; the second layer GRU contains 32 neurons and returns only the output of the last time step; the final output is a spectral feature vector.

[0022] The structure of the image feature extraction branch is as follows:

[0023] The first layer is a convolutional layer using 32 3×3 convolutional kernels with ReLU activation function, followed by a 2×2 max pooling layer; the second layer is a convolutional layer using 64 3×3 convolutional kernels with ReLU activation function, followed by a 2×2 max pooling layer; the third layer is a convolutional layer using 128 3×3 convolutional kernels with ReLU activation function; finally, a global average pooling layer is used to convert the feature map into a feature vector.

[0024] In step 4, the dataset obtained in step 2 is divided into a training set, a validation set, and a test set. The training set is used for model training, the validation set is used for hyperparameter tuning, and the test set is used for model performance evaluation.

[0025] The model training uses the classification cross-entropy loss function and the Adam optimizer. The initial learning rate is set to 0.0004, and a learning rate scheduler is configured. When the validation set loss does not decrease for 5 consecutive epochs, the learning rate is halved. During training, the batch size is set to 32 and the training epochs are set to 120.

[0026] The model performance was evaluated using accuracy, precision, recall, F1 score, and k-fold cross-validation.

[0027] Step 5 specifically involves:

[0028] The model that meets the performance requirements can be packaged into an automated classification tool or service. After the user inputs the atomic coordinate file, Raman spectral data or microscopic image of the carbon nanotube to be tested, the system will automatically output the defect type prediction result and confidence level.

[0029] The beneficial effects of this invention are:

[0030] (1) The application innovatively fuses the atomic structure data, microscopic image features and Raman spectrum information of carbon nanotubes, accurately captures the specific representation of different defect types in microstructure, morphology texture and vibration characteristics through the complementarity of multi-modal data mining. Compared with the traditional single data modal analysis method, the hybrid deep learning model combining CNN, GRU and MLP can effectively extract multi-dimensional and deep defect features, greatly reduce the background noise interference, and the recognition accuracy of Stone-Wales defects, vacancy defects, doping defects and other types of defects is more than 95%. At the same time, through the joint training of simulation data and experimental data, the model has strong generalization ability to carbon nanotubes of different chirality and size and samples without defects, avoiding the one-sidedness of features caused by a single data source;

[0031] (2) The traditional defect analysis method based on HRTEM, STM image or Raman spectrum needs to rely on the subjective experience of professionals for manual annotation, feature extraction and judgment, which not only is low in efficiency (single sample analysis may take several hours), but also has problems of large human error and poor result repeatability. The method of the application realizes the full-process automation of carbon nanotube defect analysis through a standardized data preprocessing process, an automatic feature extraction module and an end-to-end model training and reasoning. The user only needs to input the atomic coordinate file, spectrum data or microscopic image, and the system can output the defect type prediction result and confidence within minutes, which improves the analysis efficiency by hundreds of times compared with the traditional manual method, significantly reduces the dependence on the professional skills of the operator, and provides the possibility for rapid screening and quality control of large-scale carbon nanotube samples;

[0032] (3) The technical framework of the method has high flexibility and universality, and is not limited to specific types of carbon nanotubes or defect categories. On the one hand, the classification dimension of the model can be expanded by adding simulation data and experimental data of new defect types, easily adapting to the identification needs of new defects (such as boron doping defects, composite vacancy defects, etc.); on the other hand, more characterization data sources (such as X-ray diffraction data, infrared spectrum data, etc.) can be compatible, and the feature extraction branch and fusion strategy can be added to further enrich the feature input of the model and improve the defect identification ability in complex scenarios. In addition, the framework can be migrated to the defect analysis of other nanomaterials (such as graphene, nanowires, etc.), and only needs to adjust the data preprocessing rules and feature extraction methods according to the material characteristics, so as to realize the defect classification across material systems and provide a universal solution for structure characterization and performance optimization in the field of nanomaterials;

[0033] (4) The method of this invention ensures the objectivity, repeatability, and traceability of defect analysis results by establishing unified dataset labeling specifications, feature extraction standards, and model evaluation systems (using accuracy, precision, recall, F1 score combined with k-fold cross-validation). Traditional manual analysis methods lack unified standards, making it difficult to compare analysis results from different laboratories and operators. However, the automated classification tool of this invention can output standardized analysis reports, including defect type, confidence level, key quantitative feature parameters, etc., providing a unified benchmark for the sharing and comparison of scientific research data. In addition, the model's performance evaluation results are stable and reliable, effectively avoiding experimental errors caused by subjective judgment, and ensuring the scientific rigor and soundness of carbon nanotube-related research. Attached Figure Description

[0034] Figure 1 This is a model diagram of the carbon nanotubes constructed in Embodiment 1 of the present invention;

[0035] Figure 2 These are the specific parameter settings for the carbon nanotube model constructed in Embodiment 1 of the present invention;

[0036] Figure 3 This is the double-vacancy defect carbon nanotube model constructed in Example 1 of the present invention;

[0037] Figure 4-9 All are Cartesian coordinates of atoms near double-vacancy defects;

[0038] Figure 10 This is the nitrogen-doped defective carbon nanotube model constructed in Example 1 of the present invention;

[0039] Figure 11 These are the Cartesian coordinates of the N atoms in N-doped carbon nanotubes;

[0040] Figure 12 This is the SW defect carbon nanotube model constructed in Example 1 of the present invention;

[0041] Figure 13 These are the Cartesian coordinates of atoms near the SW defect;

[0042] Figure 14 These are electron microscope images of carbon nanotubes collected in Example 1 of this invention;

[0043] Figure 15 This is the Raman spectrum of carbon nanotubes collected in Example 1 of this invention;

[0044] Figure 16 This is the accuracy curve of the model on the test set in Embodiment 1 of the present invention;

[0045] Figure 17 This is the accuracy curve of the model on the training set in Embodiment 1 of the present invention;

[0046] Figure 18 is the accuracy and iteration loss curve of the model in embodiment 1 of the present application. DETAILED DESCRIPTION

[0047] The present application will be described in detail below in conjunction with the drawings and specific embodiments.

[0048] The present application is based on a multi-modal machine learning carbon nanotube identification classification method, which constructs a hybrid deep learning model combining convolutional neural network (CNN), gated recurrent unit (GRU) and multi-layer perceptron (MLP). The specific features of the microscopic image, Raman spectrum and atomic structure data of the carbon nanotube are extracted and fused, and the implementation is carried out according to the following steps:

[0049] Step 1, multi-source data set construction:

[0050] Different chiral and size carbon nanotube models are established by molecular simulation software, specific intrinsic defects (such as Stone-Wales, vacancy or doping defects) are simulated to be introduced, and corresponding atomic coordinate files are generated. At the same time, the microscopic image and spectrum data of the real carbon nanotube sample are obtained by experimental means. The simulation data and experimental data are combined to form a multi-source data set containing atomic structure, spectral features and microscopic image features, and each defective data is labeled with a clear defect type label.

[0051] Step 2, data preprocessing:

[0052] The original multi-modal data is standardized, including: parsing the atomic coordinate file and calculating the geometric topological features to generate a numerical vector; smoothing and denoising the Raman spectrum, baseline correction and sequence normalization; uniform scaling and gray scale normalization of the microscopic image.

[0053] Step 3, model building:

[0054] A three-branch hybrid deep learning model is constructed, which consists of three parallel branch networks, i.e. structure feature processing branch, spectrum feature extraction branch, image feature extraction branch, and feature fusion and joint classification layer. Among them, the structure feature processing branch uses multi-layer perceptron to encode high-dimensional geometric topological features; the spectrum feature extraction branch uses the gated recurrent unit network to extract sequence features from the spectrum data; the image feature extraction branch uses the convolutional neural network to extract spatial features from the microscopic image; the feature fusion and joint classification layer fuses the feature vectors output by the three feature extraction branches to form a unified joint feature representation.

[0055] Step 4, model training and performance evaluation:

[0056] The data set obtained in step 2 is divided into a training set, a validation set and a test set, the training set is used for model training, the validation set is used for hyperparameter tuning to prevent overfitting; the test set is used to evaluate the performance of the trained model, and multiple indicators such as accuracy, precision, recall and F1 score are used to measure the classification ability of the model, and the k-fold cross-validation method is used to ensure the stability and reliability of the model evaluation results.

[0057] Step 5, the model with performance up to standard (such as accuracy > 95%) is packaged into an automated classification tool or service, and after the user inputs the atomic coordinate file, spectral data or microscopic image of the carbon nanotube to be tested, the system automatically outputs the defect type prediction result and confidence.

[0058] Embodiment 1:

[0059] This embodiment takes the Stone-Wales defect, single vacancy defect and nitrogen doping defect of the armchair chair-shaped nanotube as an example.

[0060] Step 1, use the Materials Studio software to establish an initial model of a carbon nanotube with a chirality parameter of (8, 8), as shown in Figure 1 , the model is composed of 32 C atoms, the diameter of the carbon nanotube is 10.85 Å, and the lattice parameters are: a = 14.195001 Å, b = 14.195001 Å, c = 2.459512 Å, α = 90.00°, β = 90.00°, = 120.00°, and the extended unit cell is 1 8 to obtain a tubular structure with sufficient length. The specific parameter settings are shown in Figure 2 .

[0061] In the software, atomic deletion, bond reconstruction, atomic replacement and other operations are used to construct carbon nanotube models containing double vacancy defects, nitrogen doping defects and S-W defects, respectively. 1500 samples are generated for each defect type, and the atomic coordinate file in cif format is exported.

[0062] As shown in Figure 3 , Figure 10 , Figure 12 , they are carbon nanotube models containing double vacancy defects, nitrogen doping defects and S-W defects, respectively. The Cartesian coordinates of the atoms near the double vacancy defect are shown in Figure 4-9 , the Cartesian coordinates of the N atoms in the N-doped carbon nanotube are shown in Figure 11 , and the Cartesian coordinates of the atoms near the S-W defect are shown in Figure 13 .

[0063] HRTEM images, STM images and Raman spectrum data of real carbon nanotube samples are collected, and defect types are labeled according to image and spectrum characteristics and atomic coordinate files. Finally, the defect types are divided into the following three categories:

[0064] Category 1: Stone-Wales defects

[0065] Category 2: Double vacancy defects

[0066] Category 3: Nitrogen-doped defects

[0067] As shown in Figure 14 , the collected carbon nanotube electron microscope images, Figure 15 , and the carbon nanotube Raman spectrum images.

[0068] Integrate simulation data and experimental data to establish a total data set, and randomly divide it into training set, validation set and test set according to the ratio of 7:2:1.

[0069] Step 2, standardize the original multi-modal data.

[0070] For atomic coordinate data, parse it through Python's ASE library, calculate C-C bond length distribution statistics, C-C-C bond angle distribution statistics and topological ring statistical characteristics, and generate numerical feature vectors.

[0071] For Raman spectrum data, use Savitzky-Golay filter for smoothing and denoising, use asymmetric least squares method for baseline correction, intercept the feature interval of 1000cm - to 1800cm - and normalize it to a fixed standard sequence of 1000 data points.

[0072] For microscopic image data, uniformly scale to 224x224 pixels and perform grayscale normalization to obtain a standard image matrix.

[0073] Step 3, construct a multi-branch hybrid deep learning model, which consists of structure feature processing branch, spectrum feature extraction branch, image feature extraction branch, feature fusion and joint classification layer.

[0074] (1) Structure feature processing branch

[0075] High-dimensional coding is performed on the geometric and topological features designed by artificial, and the complex mapping relationship between feature combination and defects is learned. The input of this branch is the geometric and topological feature vector extracted from the atomic coordinate file, including bond length, bond angle distribution statistics and topological ring statistics.

[0076] The structure of this branch is:

[0077] including fully connected layer 1 and fully connected layer 2:

[0078] Fully connected layer 1: 128 neurons, ReLU activation function, followed by a Dropout layer;

[0079] Fully connected layer 2: 64 neurons, ReLU activation function;

[0080] Finally output a 64-dimensional structure of high-level feature vector.

[0081] (2) Spectral feature extraction branch

[0082] For processing Raman spectrum sequence data, capturing its peak position, peak shape and other sequence-dependent relationships. The input of this branch is the pre-processed one-dimensional spectrum sequence with a length of 1000, which consists of two layers of gated recurrent unit network: the first layer GRU contains 64 neurons, returning the entire output sequence; the second layer GRU contains 32 neurons, only returning the output of the last time step, finally outputting a 32-dimensional spectral high-level feature vector.

[0083] (3) Image feature extraction branch

[0084] Learn spatial features related to defects from the original microscopic image, extract spatial features such as contrast changes, texture abnormalities and structural distortions caused by defects in the image. The input of this branch is the pre-processed standard image matrix, and the network structure is:

[0085] First layer: convolutional layer, using 32 3x3 convolutional kernels, activation function ReLU; followed by a 2x2 max pooling layer.

[0086] Second layer: convolutional layer, using 64 3x3 convolutional kernels, activation function ReLU; followed by a 2x2 max pooling layer.

[0087] Third layer: convolutional layer, using 128 3x3 convolutional kernels, activation function ReLU.

[0088] Finally, use the global average pooling layer to convert the feature map to a 128-dimensional feature vector.

[0089] (4) Feature fusion and joint classification layer

[0090] Concatenate the feature vectors output by the above three branches to form a 224-dimensional joint feature vector. This joint feature vector passes through a fully connected layer (128 neurons, ReLU activation, Dropout) and an output layer (3 neurons, Softmax activation) in turn, outputting the probability distribution of the three types of defects.

[0091] Step 4, model training and performance evaluation:

[0092] Model training adopts classification cross-entropy loss function and Adam optimizer, the initial learning rate is set to 0.0004, and the learning rate scheduler is selected, such as when the validation set loss does not decrease for 5 consecutive epochs, the learning rate is halved. The batch size is set to 32, and the training period is set to 120. The training set is used for end-to-end training, and the model performance is evaluated on the validation set after each epoch to monitor the training loss and validation loss to prevent overfitting.

[0093] The reserved test set (10% of the whole set) is used to evaluate the performance of the trained model, and multiple indicators such as accuracy, precision, recall, and F1 score are used to comprehensively measure the classification ability of the model. Through k-fold cross-validation method to ensure the stability and reliability of the model evaluation results. As shown in Figure 16 , the model achieves an accuracy of 98% on the test set. As shown in Figure 17 , the training prediction accuracy reaches 97.81%. As shown in Figure 18 , the learning curve shows that the model converges well during training, with rapid increase in accuracy and rapid decrease in loss function, and no obvious overfitting phenomenon, proving the efficiency and reliability of the method.

[0094] Step 5, encapsulate the performance qualified model as an automated classification tool or service. Users can input the atomic coordinate file, spectral data or microstructure image of the carbon nanotube to be tested through the interface, and the system automatically performs feature extraction and classification, quickly outputs the prediction results and corresponding confidence of the defect type, and provides instant feedback for material quality control and process optimization.

[0095] Example 2:

[0096] This embodiment is a carbon nanotube defect identification and classification method based on multi-modal machine learning method, which is implemented according to the following steps:

[0097] Step 1, obtain the atomic coordinate file, microscopic image and Raman spectrum of the carbon nanotube, form a multi-source data set, and label the defect type label for the defective data;

[0098] Step 2, pre-process the multi-source data set obtained in step 1;

[0099] Step 3, construct a three-branch hybrid deep learning model composed of multi-layer perception, gated recurrent unit and convolutional neural network to process atomic structure data, Raman spectrum and microscopic image respectively;

[0100] Step 4, train and evaluate the model using the data set obtained in step 2;

[0101] Step 5, use the optimized model to identify and classify carbon nanotube defects.

[0102] Example 3:

[0103] On the basis of example 2, step 1 is specifically:

[0104] Different chiral and size carbon nanotube models are established by molecular simulation software, intrinsic defects are simulated to be introduced, and corresponding atomic coordinate files are generated. At the same time, the microscopic image and Raman spectrum data of the real carbon nanotube sample are obtained by experimental means. The simulation data and experimental data are integrated to form a multi-source data set containing atomic structure, Raman spectrum characteristics and microscopic image characteristics, and each defective data is labeled with a defect type label.

[0105] Example 4:

[0106] On the basis of example 3, step 2 is specifically:

[0107] The atomic coordinate file is parsed and the geometric topological features are calculated to generate a numerical vector; the Raman spectrum is smoothed and denoised, baseline corrected and sequence normalized; the microscopic image is uniformly scaled and gray normalized.

[0108] Example 5:

[0109] In step 3 of example 4, the three-branch hybrid deep learning model is composed of a parallel structure feature processing branch, a spectrum feature extraction branch, an image feature extraction branch, and a feature fusion and joint classification layer. Among them, the structure feature processing branch uses a multilayer perceptron to encode the geometric topological features in a high dimension; the spectrum feature extraction branch uses a gated recurrent unit network to extract sequence features from the spectrum data; the image feature extraction branch uses a convolutional neural network to extract spatial features from the microscopic image; the feature fusion and joint classification layer fuses the feature vectors output by the three feature extraction branches to form a joint feature vector. The joint feature vector passes through the full connection layer and the output layer in turn to output the probability distribution of all defects.

[0110] Example 6:

[0111] On the basis of example 5, the structure of the structure feature processing branch is:

[0112] It includes full connection layer 1 and full connection layer 2; full connection layer 1 contains 128 neurons, uses ReLU activation function, and is followed by a Dropout layer; full connection layer 2 contains 64 neurons, uses ReLU activation function; and finally outputs the structure feature vector.

[0113] The structure of the spectrum feature extraction branch is:

[0114] The two-layer gated recurrent unit network consists of: the first layer GRU contains 64 neurons, returning the entire output sequence; the second layer GRU contains 32 neurons, returning only the output of the last time step; and the final output is the spectral feature vector.

Claims

1. A method for defect identification and classification of carbon nanotubes based on multimodal machine learning, characterized in that, The specific steps are as follows: Step 1: Obtain the atomic coordinate files, microscopic images, and Raman spectra of carbon nanotubes to form a multi-source dataset, and label the defective data with defect type labels; Step 2: Preprocess the multi-source dataset obtained in Step 1; Step 3: Construct a three-branch hybrid deep learning model consisting of a multilayer perceptron, a gated recurrent unit, and a convolutional neural network to process atomic structure data, Raman spectra, and microscopic images respectively; Step 4: Use the dataset obtained in Step 2 to train the model and evaluate its performance; Step 5: Use the optimized model to identify and classify defects in carbon nanotubes; In step 3, the three-branch hybrid deep learning model consists of a parallel structural feature processing branch, a spectral feature extraction branch, an image feature extraction branch, and a feature fusion and joint classification layer. The structural feature processing branch uses a multilayer perceptron to encode geometric topological features in high dimension. The spectral feature extraction branch uses a gated recurrent unit network to extract sequence features from spectral data. The image feature extraction branch uses a convolutional neural network to extract spatial features from microscopic images. The feature fusion and joint classification layer fuses the feature vectors output by the three feature extraction branches to form a joint feature vector. This joint feature vector passes through a fully connected layer and an output layer in sequence to output the probability distribution of all defects.

2. The carbon nanotube defect identification and classification method based on multimodal machine learning as described in claim 1, characterized in that, Step 1 is as follows: Models of carbon nanotubes with different chirities and sizes were established using molecular simulation software. Intrinsic defects were simulated and corresponding atomic coordinate files were generated. Simultaneously, microscopic images and Raman spectral data of real carbon nanotube samples were obtained through experimental methods. The simulated data and experimental data were integrated to form a multi-source dataset containing atomic structure, Raman spectral features, and microscopic image features. Defect type labels were also assigned to each defective data point.

3. The carbon nanotube defect identification and classification method based on multimodal machine learning as described in claim 1, characterized in that, Step 2 is as follows: The atomic coordinate file is parsed and geometric topological features are calculated to generate numerical vectors; Raman spectra are smoothed, denoised, baseline corrected, and sequence normalized; and microscopic images are uniformly scaled and grayscale normalized.

4. The carbon nanotube defect identification and classification method based on multimodal machine learning as described in claim 1, characterized in that, The structure of the structural feature processing branch is as follows: It includes a fully connected layer 1 and a fully connected layer 2; the fully connected layer 1 contains 128 neurons, uses the ReLU activation function, and is followed by a Dropout layer; the fully connected layer 2 contains 64 neurons, uses the ReLU activation function; the final output is a structural feature vector.

5. The carbon nanotube defect identification and classification method based on multimodal machine learning according to claim 1, characterized in that, The structure of the spectral feature extraction branch is as follows: It consists of two layers of gated recurrent unit networks: the first layer GRU contains 64 neurons and returns the entire output sequence; the second layer GRU contains 32 neurons and returns only the output of the last time step; the final output is a spectral feature vector.

6. The carbon nanotube defect identification and classification method based on multimodal machine learning according to claim 1, characterized in that, The structure of the image feature extraction branch is as follows: The first layer is a convolutional layer using 32 3×3 convolutional kernels with ReLU activation function, followed by a 2×2 max pooling layer; the second layer is a convolutional layer using 64 3×3 convolutional kernels with ReLU activation function, followed by a 2×2 max pooling layer; the third layer is a convolutional layer using 128 3×3 convolutional kernels with ReLU activation function; finally, a global average pooling layer is used to convert the feature map into a feature vector.

7. The carbon nanotube defect identification and classification method based on multimodal machine learning according to claim 1, characterized in that, In step 4, the dataset obtained in step 2 is divided into a training set, a validation set, and a test set. The training set is used for model training, the validation set is used for hyperparameter tuning, and the test set is used for model performance evaluation.

8. The carbon nanotube defect identification and classification method based on multimodal machine learning according to claim 7, characterized in that, The model training uses the classification cross-entropy loss function and the Adam optimizer. The initial learning rate is set to 0.0004, and a learning rate scheduler is configured. When the validation set loss does not decrease for 5 consecutive epochs, the learning rate is halved. During training, the batch size is set to 32 and the training epochs are set to 120. The model performance was evaluated using accuracy, precision, recall, F1 score, and k-fold cross-validation.

9. The carbon nanotube defect identification and classification method based on multimodal machine learning according to claim 1, characterized in that, Step 5 specifically involves: The model that meets the performance requirements can be packaged into an automated classification tool or service. After the user inputs the atomic coordinate file, Raman spectral data and microscopic image of the carbon nanotube to be tested, the system will automatically output the defect type prediction result and confidence level.

Images