Method and device for measuring thickness of two-dimensional material, electronic equipment and storage medium

Abstract

The application provides a method and device for measuring the thickness of a two-dimensional material, electronic equipment and a storage medium, the method comprising: obtaining a target image of a CVD-synthesized two-dimensional material to be processed; removing the background of the target image and dividing the target image into a plurality of to-be-processed sub-images, each to-be-processed sub-image having the same color; and performing the following processing on each to-be-processed sub-image: inputting the to-be-processed sub-image into a trained EfficientNet-B0 neural network to obtain the corresponding thickness of the to-be-processed sub-image. The method can measure the thickness of a CVD-synthesized two-dimensional material.

Description

Technical Field

[0001] This invention relates to the field of chemical vapor deposition technology, and more particularly to a method, apparatus, electronic device, and storage medium for measuring the thickness of two-dimensional materials. Background Technology

[0002] Two-dimensional materials possess atomically thin structures and excellent comprehensive electrical, optical, and mechanical properties, making them highly promising for applications in cutting-edge fields such as next-generation semiconductor devices, flexible electronic devices, and low-power integrated circuits. They are currently a key research focus in the field of micro- and nanomaterials. Chemical vapor deposition (CVD), with its outstanding advantages such as large-area fabrication and controllable film layer number and growth process, has become the mainstream process for the large-scale preparation of two-dimensional materials.

[0003] However, in the actual process of growing two-dimensional materials using CVD technology, multiple factors such as uneven temperature field distribution in the furnace, disturbance of the reaction gas flow field, and differences in the surface morphology of the substrate can cause significant non-uniformity in the number of layers and thickness distribution of the resulting two-dimensional material film. Since the film thickness directly determines the core optoelectronic properties of the two-dimensional material, achieving accurate and full-domain characterization of the thickness of two-dimensional materials is an indispensable key step in the material preparation and device development process.

[0004] Current methods for thickness measurement of CVD-synthesized two-dimensional materials still have many shortcomings: the purchase and maintenance costs of precision detection equipment such as atomic force microscopes and Raman spectroscopy are high, and they can only complete small-area, single-point detection at a time, resulting in low detection efficiency; traditional optical visual judgment methods rely heavily on the personal experience of operators, resulting in poor consistency of detection results and difficulty in forming a unified judgment standard; none of the above-mentioned existing technologies can complete automated and standardized rapid detection of the thickness of large-area thin films, making it difficult to meet the high-throughput characterization needs of batch material samples.

[0005] In conclusion, how to measure the thickness of CVD-synthesized two-dimensional materials has become an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a method, apparatus, electronic device, and storage medium for measuring the thickness of two-dimensional materials.

[0007] To achieve one of the above-mentioned objectives, an embodiment of the present invention provides a method for measuring the thickness of a CVD-synthesized two-dimensional material, comprising the following steps: acquiring a target image of the CVD-synthesized two-dimensional material to be processed; removing the background of the target image and dividing the target image into several sub-images to be processed, each sub-image having the same color; and performing the following processing on each sub-image to be processed: inputting the sub-image to be processed into a trained EfficientNet-B0 neural network to obtain the corresponding thickness of the sub-image to be processed.

[0008] As a further improvement of one embodiment of the present invention, the removal of the background of the target image specifically includes: processing the target image based on a trained U-Net image segmentation framework to obtain a black and white mask, and removing the background of the target image based on the black and white mask. The U-Net image segmentation framework uses a ResNet34 residual network as an encoder.

[0009] As a further improvement of one embodiment of the present invention, the step of dividing the target image into several sub-images to be processed, each sub-image having the same color, specifically includes: dividing the target image into several sub-regions using the SLIC superpixel segmentation algorithm, and clustering the several sub-regions using the Mean Shift clustering algorithm to obtain several sub-images to be processed, each sub-image having the same color.

[0010] This invention also provides a device for measuring the thickness of a CVD-synthesized two-dimensional material, comprising the following modules: an image acquisition module for acquiring a target image of the CVD-synthesized two-dimensional material to be processed; an image preprocessing module for removing the background of the target image and dividing the target image into several sub-images to be processed, each sub-image having the same color; and a thickness measurement module for performing the following processing on each sub-image to be processed: inputting the sub-image to be processed into a trained EfficientNet-B0 neural network to obtain the corresponding thickness of the sub-image to be processed.

[0011] As a further improvement of one embodiment of the present invention, the image preprocessing module is further configured to: process the target image based on the trained U-Net image segmentation framework to obtain a black and white mask, and remove the background of the target image based on the black and white mask, wherein the U-Net image segmentation framework uses a ResNet34 residual network as an encoder.

[0012] As a further improvement of one embodiment of the present invention, the image preprocessing module is further configured to: divide the target image into several sub-regions using the SLIC superpixel segmentation algorithm, and perform clustering processing on the several sub-regions using the Mean Shift clustering algorithm to obtain several sub-images to be processed, wherein each sub-image to be processed has the same color.

[0013] This invention also provides an electronic device, including: a memory for storing executable instructions; and a processor for implementing the above-described measurement method when executing the executable instructions stored in the memory.

[0014] This invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described measurement method.

[0015] Compared to existing technologies, the technical advantages of this invention are as follows: This invention provides a method, apparatus, electronic device, and storage medium for measuring the thickness of two-dimensional materials. The measurement method includes: acquiring a target image of a CVD-synthesized two-dimensional material to be processed; removing the background of the target image and dividing the target image into several sub-images to be processed, each sub-image having the same color; and performing the following processing on each sub-image: inputting the sub-image to be processed into a trained EfficientNet-B0 neural network to obtain the corresponding thickness of the sub-image. This measurement method can measure the thickness of CVD-synthesized two-dimensional materials. Attached Figure Description

[0016] Figure 1 and Figure 2 This is a schematic diagram of the structure of the image acquisition tool in an embodiment of the present invention; Figure 3 This is a schematic diagram of the target image in an embodiment of the present invention; Figure 4 and Figure 5 This is a schematic diagram of the measurement method in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of the EfficientNet-B0 neural network in an embodiment of the present invention; Figure 7 This is a flowchart illustrating the measurement method in an embodiment of the present invention. Detailed Implementation

[0017] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some embodiments may include or substitute parts and features of other embodiments. The scope of the embodiments herein encompasses the entire scope of the claims and all available equivalents thereof. Throughout this document, the terms “first,” “second,” etc., are used only to distinguish one element from another without requiring or implying any actual relationship or order between the elements. Indeed, a first element can also be referred to as a second element, and vice versa. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-parallel inclusion, such that a structure, apparatus, or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a structure, apparatus, or device. Without further limitation, an element defined by the phrase “comprising one…” does not preclude the presence of other identical elements in the structure, apparatus, or device that includes said element. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0018] The terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer" used in this document to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings and are used only for the convenience of describing this document and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. In the description herein, unless otherwise specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two elements, or direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0019] Embodiment 1 of the present invention provides a method for measuring the thickness of a CVD-synthesized two-dimensional material, such as... Figure 7 As shown, it includes the following steps: Step 701: Obtain the target image of the 2D material to be processed by CVD synthesis.

[0020] here, Figure 1 and Figure 2A device for acquiring target images is shown, comprising: 1. an omnidirectional adjustable stage capable of rotation, translation, and vertical adjustment; 2. an integrated controllable light source module that, in conjunction with the stage, strictly controls illumination conditions according to predetermined standards to minimize chromatic aberration; and 3. an optical microscope equipped with x5, x10, and x20 multi-magnification objectives and a high-resolution camera, which can optimize imaging for different sample characteristics to acquire clear and detailed images.

[0021] Step 702: Remove the background of the target image and divide the target image into several sub-images to be processed, each sub-image having the same color.

[0022] Here, the advantage of first removing the background is that it eliminates substrate interference. The original image contains background pixels such as silicon substrate, dust, and oxide spots. Background colors and textures can interfere with the neural network's extraction of color features from 2D material films. After removing the background, only the pixels of the 2D material to be measured are retained, and irrelevant noise is completely isolated, significantly reducing thickness prediction errors. This reduces unnecessary computation, as background pixels do not participate in thickness calculations. Removing them reduces the amount of data input to subsequent networks, speeds up EfficientNet-B0 inference, and improves the overall batch detection efficiency.

[0023] Then, the core advantage of dividing independent sub-images by the same color is that two-dimensional CVD thin films have a key characteristic: under optical microscopy, the film thickness corresponds one-to-one with the image color, the color is uniform and consistent in areas of the same thickness, and the color difference is obvious in areas of different thicknesses.

[0024] Based on this characteristic, block segmentation has multiple advantages: 1. Region division conforms to the physical law of thickness: Same color = same thickness. It naturally splits the entire image into independent blocks of uniform thickness. The thickness within each sub-image has no gradient change, avoiding feature confusion caused by a single image containing multiple thicknesses. 2. Reduces the difficulty of neural network recognition and improves prediction accuracy: If the entire mixed multi-color image is directly input, the network needs to distinguish multiple thickness features simultaneously, resulting in a heavy recognition burden and easy confusion. A single sub-image only contains color features corresponding to a single thickness. EfficientNet-B0 only needs to learn a single feature mapping relationship, significantly improving prediction accuracy. 3. Achieves accurate thickness measurement by region and outputs spatially layered thickness results: After splitting, the thickness of each monochrome region can be calculated separately, and the final output is partitioned data of "position + corresponding thickness", rather than the entire image. Figure 1 4. The average thickness meets the needs of observing the local film thickness distribution in scientific research / production. It is compatible with data cleaning and standardization preprocessing. Monochrome sub-blocks have clear boundaries and simple features, facilitating subsequent processing.

[0025] Step 703: Perform the following processing on each sub-image to be processed: input the sub-image to be processed into the trained EfficientNet-B0 neural network to obtain the corresponding thickness of the sub-image to be processed.

[0026] Here, CVD (Chemical Vapor Deposition) produces thin-film interference colors on the substrate, and different thicknesses result in different visual features in the image. EfficientNet-B0 is a high-performance, lightweight convolutional network that can automatically learn the mapping relationship from "image texture / color / contour" to "thin-film nanometer thickness," and it is decomposed into four layers: The underlying physical premise is that thickness directly determines the visual characteristics of a microscopic image (the root of model learning). The two-dimensional material on the silicon oxide substrate is a nanoscale transparent film, subject to optical thin-film interference: light passes through air → two-dimensional material → SiO2 substrate, and the reflected light from multiple interfaces interferes with each other. Changes in the material's nanometer thickness synchronously alter the wavelength and intensity of the interfering light, resulting in completely different colors, saturations, and brightness distributions under the microscope. Besides color, uneven thickness leads to surface texture undulations, gradual changes in edge contours, and differences in the area and shape of regions. In summary, the color, texture, and contour in each material image inherently contain thickness information; however, the human eye can only roughly distinguish the number of layers and cannot quantify nanometer values, while convolutional networks can precisely capture this.

[0027] The structural advantages of the EfficientNet-B0 network make it suitable for extracting thickness-related visual features. The core innovation of EfficientNet is balanced scaling (simultaneous optimization of width, depth, and resolution). B0 is the lightest baseline version, which is more suitable for this scenario than ordinary CNNs (VGG, basic ResNet): MBConv inverse residual module (with depthwise separable convolution + SE attention). Depthwise separable convolution can extract spatial texture and color channel features separately, accurately capturing subtle color differences and small texture undulations in thin films (weak color differences in single-layer / double-layer materials); the SE attention mechanism enables the network to automatically focus on the color channel regions most sensitive to thickness, ignoring irrelevant noise such as dust and minor scratches; multi-scale 3×3 / 5×5 convolution kernel stacking can: simultaneously extract fine-grained edges (small convolution) and color distribution of the whole area (large convolution), taking into account both local subtle thickness differences and the overall color display rules of global blocks. The network process matches the thickness detection input (224×224 material slice image). The first layer uses 3×3 convolution downsampling to extract basic brightness and darkness and edges. Multiple layers of MBConv extract deep features step by step: from simple color → complex texture → thickness-related latent features. At the end, global average pooling compresses the entire image into a one-dimensional feature vector, where each value represents a visual pattern related to thickness. The lightweight advantage is that the number of parameters is much smaller than that of large networks, and it can perform real-time inference on ordinary industrial control computers / terminals, meeting the needs of online rapid thickness measurement of equipment.

[0028] Its training logic for learning "thickness" is as follows: Supervised regression task. EfficientNet-B0 is not inherently capable of measuring thickness; it establishes a mapping through training on labeled datasets. Training dataset: A large number of material micrographs + corresponding ground truth thickness labels (nanometer-thickness precisely measured by AFM / Raman); Task type: Regression task (outputting continuous thickness values, not classification layers); Network training objective: Continuously adjust convolutional weights to reduce the error between "network-predicted thickness" and "actual thickness"; After training, the network's internal convolutional kernels have solidified a set of rules: Input a background-exposed color image of the material → extract multi-dimensional visual feature vectors → vector values ​​correspond to a fixed thickness range. To compensate for its shortcomings, 14-dimensional artificial feature engineering is used. The features extracted by EfficientNet alone have limited distinguishing power for extremely thin materials. Therefore, the network output depth features are combined with 14-dimensional physical features (such as color, area, and compactness) designed by the network, and these dual features are input into the regression head to further improve the accuracy of thickness prediction.

[0029] The thickness of two-dimensional materials is transformed into unique color, texture, and contour features of the image through thin film interference. EfficientNet-B0 relies on multi-scale lightweight convolution and attention mechanism to efficiently extract these subtle visual features. Then, through training with labeled thickness data, it learns the mathematical mapping relationship between image features and nanometer thickness, and finally realizes the direct quantitative prediction of thin film thickness from the image.

[0030] like Figure 6 As shown, the left figure (a) shows the complete forward inference network architecture of EfficientNet-B0, which is used to automatically extract deep visual features of images; the right figure (b) shows the dimension radar distribution map of 14-dimensional features constructed by humans, which intuitively displays all manually input feature items such as geometry, color, saturation, and centroid.

[0031] Module a shows the EfficientNet-B0 deep network automatically extracting deep visual features from images. The network input is the original cropped two-dimensional material region image with a uniform size of 224×223 pixels.

[0032] The network backbone hierarchy (with corresponding flowcharts for each layer) is as follows: 1. First layer convolution, Conv3×3, stride 2, downsampling the original image to initially compress the image size and extract basic edge textures; 2. Multi-layer MBConv module stacking (EfficientNet core lightweight inverse residual module): MBConv1 (3×3 convolution kernel), MBConv6×2 (3×3), MBConv6×3 (5×5), MBConv6×2 (5×5), MBConv6×4 (5×5), MBConv6×1 (3×3), MBConv fuses depthwise separable convolution and SE attention mechanism to extract multi-scale features with extremely low parameter count: fine surface texture, material edge contours, and global color distribution; the last 1×1 convolution Conv1×1, channel dimension feature compression and fusion; Global Average Pooling: compresses the two-dimensional image feature map into a one-dimensional depth output feature vector, automatically learning the underlying mapping relationship between the image and thickness. Function: Purely data-driven, requiring no manual rule design, it automatically captures subtle color and texture differences that are difficult for the human eye to distinguish, and establishes a basic correlation between image visual information and material thickness.

[0033] Module b shows a 14-dimensional artificial feature engineering supplement (radar feature map on the right). Relying solely on deep network features is insufficient for distinguishing nanoscale thin film thicknesses. Therefore, a 14-dimensional quantized feature vector is artificially constructed and fused with EfficientNet deep features. Each axis on the radar map corresponds to a feature dimension, divided into four categories: 1. Spatial geometry: total area of ​​the block, logarithmic area (Log Area), compactness, centroid X coordinate, centroid Y coordinate, and proportion coefficients A / B / L; principle: there are regular differences in the growth area and contour regularity of two-dimensional materials of different thicknesses. 2. LAB color space: L (brightness), A (red-green channel), B (yellow-blue channel); principle: the interference color of a two-dimensional material on a silicon substrate is directly determined by its thickness, and the LAB space can accurately quantify the color rendering of the thin film. 3. Optical chromaticity: Light Intensity, Color Magnitude, and Luminance Ratio; 4. Color saturation: Saturation; All features are concatenated to form a fixed 14-dimensional artificial feature vector, which makes up for the shortcomings of convolutional networks in numerical representation of thin film interference colors and enhances the ability to distinguish small thickness differences between thin and thick layers.

[0034] The data optimization strategy (a core means to improve model stability) involves a two-step standardization transformation of the total feature set after fusing EfficientNet deep features with 14-dimensional artificial features. Simultaneously, the dataset is cleaned: RobustScaler robust standardization, unlike ordinary mean-variance normalization, uses median and quartile scaling to significantly reduce feature value shifts caused by abnormal pixels such as image scratches, dust, and localized bright spots, resulting in stronger noise resistance; Logarithmic transformation of thickness labels reflects the natural long-tailed distribution of two-dimensional material samples: a large number of ultra-thin single-layer samples and a small number of thick multi-layer samples; logarithmic transformation compresses the numerical range of thick samples and stretches the range of thin samples, balancing the weights of thick and thin samples and preventing the model from only learning the patterns of thin layers and making inaccurate predictions of thick layers; data cleaning and filtering automatically removes invalid image blocks without thickness ground truth or with invalid labels during the training phase, reducing noise samples from interfering with training convergence.

[0035] By integrating "automatic deep feature extraction using lightweight convolutional networks + supplementation of geometric color features from artificial thin-film interference + robust data equalization preprocessing", a color-thickness regression algorithm is constructed to solve two major pain points of traditional methods: manual interpretation of interference colors by the naked eye has large subjective errors and cannot quantitatively output nanometer-level thickness; a single CNN network is insufficient to distinguish subtle color differences at the nanometer level and the prediction accuracy of thick and thin samples is uneven; ultimately, a fully automatic, stable, and high-precision quantitative thickness analysis output of two-dimensional material microscopic images is achieved.

[0036] In this embodiment, as Figure 4 As shown, removing the background of the target image specifically includes: The target image is processed based on the trained U-Net image segmentation framework to obtain a black and white mask. The background of the target image is removed based on the black and white mask. The U-Net image segmentation framework uses a ResNet34 residual network as the encoder.

[0037] Here, the substrate region of the original CVD two-dimensional material micrograph contains a large number of interfering elements: substrate color difference, dust particles, scratches, blank substrate spots, and other cluttered backgrounds. If the entire original image is directly fed into the subsequent EfficientNet-B0 thickness prediction model, background-irrelevant colors and textures will severely contaminate the feature extraction process, resulting in large thickness prediction errors and poor model stability. Therefore, it is necessary to first separate the foreground / background pixels and remove all invalid background pixels.

[0038] The standard encoder-decoder symmetric structure of the U-Net basic architecture, plus skip connections, is the basic framework for achieving pixel-level semantic segmentation. ResNet34 serves as the encoder backbone, replacing the simple stacked convolution downsampling branches of the native U-Net. It uses residual modules to solve the gradient vanishing problem in deep networks, enabling the extraction of deeper and more subtle thin film color and contour features, and accurately distinguishing the weak interference color difference between ultrathin two-dimensional materials and substrates.

[0039] The encoder (downsampling branch, ResNet34) progressively downsamples and compresses the input microscopic image while increasing the number of feature channels, extracting multi-scale features layer by layer: the bottom layer extracts edges and fine color patches, while the top layer extracts the overall material region and large-area color distribution, generating multiple sets of high- and low-resolution feature map caches, which are then passed to the decoder. The residual structure advantage allows the network to deepen without experiencing gradient vanishing during training, resulting in stronger segmentation and recognition capabilities for single-layer, small-sized two-dimensional materials. The decoder (upsampling branch) progressively enlarges the feature map size layer by layer through transposed convolutions, gradually restoring it to the exact same resolution as the original input image for pixel-by-pixel classification. A key structure is the skip connection, which directly concatenates the high-resolution shallow detail features cached at each layer of the encoder to the corresponding layer of the decoder. Its function is to compensate for the loss of edge details caused by downsampling, fully preserving the contours of tiny material fragments and thin film edges, and avoiding the problem of missing material edges and small thin sheets.

[0040] The model output and background removal process result in a single-channel black-and-white binary mask image that corresponds one-to-one with the original image size: white pixels are identified as foreground targets (effective areas of 2D material); black pixels are identified as interfering backgrounds (substrate, dust, scratches, etc.); post-mask processing involves multiplying the black-and-white mask pixel-by-pixel with the original color micrograph; pixels in the black background area are directly cleared to zero, retaining only the color pixels of the 2D material covered by the white mask; output result: a clean image without substrate, blemishes, or dust interference, retaining only the color information of the 2D material, which serves as the standard input for the downstream thickness prediction module.

[0041] In this embodiment, as Figure 5 As shown, dividing the target image into several sub-images to be processed, each sub-image having the same color, specifically includes: The target image is divided into several sub-regions using the SLIC superpixel segmentation algorithm. The mean shift clustering algorithm is then used to cluster these sub-regions, resulting in several sub-images to be processed, each with the same color.

[0042] Perform secondary fine partitioning on a single piece of material: divide continuous pixels with consistent color and uniform thickness into independent blocks, each block corresponding to a single thickness, and send them separately to the model for thickness measurement to achieve precise quantitative analysis in a localized manner.

[0043] SLIC (Simple Linear Iterative Clustering) superpixel segmentation (coarse-grained, fine-grained block segmentation) algorithm principle: SLIC is based on a dual-dimensional clustering approach using pixel spatial coordinates and LAB color values, dividing the complete material image into numerous small, compact superpixel blocks. Segmentation constraints include: spatial continuity (adjacent pixels are assigned to the same superpixel, preventing cross-regional discrete pixels); color similarity (pixels with similar colors are preferentially grouped together); and controllable block size (the basic superpixel size is manually set, resulting in numerous uniformly fragmented small color blocks that precisely match the subtle color gradient boundaries of the material). Its function is to decompose the image into basic micro-units, solving the problem of single-pixel noise and minute color differences preventing clustering, and providing the smallest processing unit for subsequent large-scale color block merging.

[0044] Mean Shift Clustering (Merging of Same-Color Blocks): The algorithm uses each superpixel output by SLIC as a sample, extracting the LAB average color feature of each superpixel. Unsupervised clustering is then performed within the color space: the algorithm automatically finds the peak color distribution density and merges adjacent superpixels with minimal color differences and belonging to the same thickness range into a large block. Key advantages (adaptable to 2D material scenarios): No need to pre-set the number of clusters; it automatically outputs corresponding blocks based on the actual color types of the image, eliminating the need for manual specification of the number of thickness partitions; it only merges regions with similar visual and interference color, preserving the boundaries of regions with significant color differences and varying thicknesses. Output: The image is segmented into several independent large blocks with uniform color, spatial continuity, and uniform thickness, with clear color boundaries between different blocks.

[0045] Advantages: 1. Layered processing for precise boundaries: SLIC is used to refine pixel boundaries, followed by MeanShift for global color clustering, merging fine edge details with large homogeneous areas without losing the thin-thickness boundary line; 2. Unsupervised adaptive segmentation: No need to pre-label thickness or preset the number of blocks, automatically adapting to 2D materials of any shape and number of layers; 3. Solves mixed thickness prediction errors: Achieves "one thickness per region," avoiding distortion of prediction values ​​caused by mixing multiple features of the entire material, significantly improving the accuracy of local thickness measurement; 4. Matches the physical properties of thin film interference: Fully relies on the physical law that "thickness determines interference color" of 2D materials, using color as the core basis for partitioning, unifying the segmentation logic and the underlying principle of thickness measurement.

[0046] In this embodiment, removing the background of the target image specifically includes: The training process for the EfficientNet-B0 neural network may include the following steps: I. Pre-training: Dataset construction and image preprocessing, specifically including: 1. Sample acquisition and ground truth labeling: Input CVD 2D material optical micrographs; background removal: Use U-Net (ResNet34) segmentation network to remove the silicon substrate and dust background, retaining only the 2D material's colored foreground region; ground truth labeling: Use AFM / Raman spectroscopy / profilometer to measure the true nanometer thickness of each material region as a supervision label; data cleaning: Remove invalid samples without thickness labels, blurry images, or those with severe stains. 2. Image standardization: Crop and scale the segmented material regions to the network's standard input size of 224×224 RGB images, using them as image input for EfficientNet-B0. 3. Simultaneous extraction of 14-dimensional artificial features: For the same material region, simultaneously calculate geometric and color artificial features: area, centroid XY, LAB three channels, saturation, compactness, logarithmic area, various color ratios, etc., to obtain a 14-dimensional feature vector, which is stored separately for later use. When capturing optical micrographs of CVD two-dimensional materials, a selection process is required. The selection criteria are as follows: choose materials where the background and target are clearly distinguishable and where the background color is relatively uniform. Figure 3 A microscopic image (10x magnification) of a two-dimensional material on a silicon wafer substrate is shown. It can be observed that the silicon wafer appears as a relatively uniform purple area in the image, indicating good background uniformity. However, the thickness of the oxide layer on the silicon wafer surface may be uneven due to the fabrication process or other factors, resulting in variations in the purple distribution in the microscopic image. Furthermore, impurities such as dust may adhere to the silicon wafer surface, appearing as noticeable black spots under the microscope, which also disrupts the background uniformity. Images with these problems are generally not used.

[0047] II. Global Data Optimization, specifically including: 1. Feature Standardization: Robust normalization of all 14-dimensional artificial features: Scaling based on median and quartiles to suppress feature mutations caused by outliers such as scratches and light spots, avoiding training oscillations. Only artificial features are subjected to RobustScaler, and the depth features output by EfficientNet are normalized internally. 2. Logarithmic Transformation of Thickness Labels: The original thickness samples exhibit a long-tailed distribution (many ultra-thin single-layer samples and very few thick multi-layer samples). A logarithmic transformation is performed on the ground truth thickness t: During training, the network fits a logarithmic thickness; during inference, an inverse transformation is performed. To accurately represent nanometer-thickness, training weights are balanced for samples of varying thicknesses. 3. Dataset partitioning: The dataset is split proportionally into three sets: 70% training set, 20% validation set, and 10% test set, with strict isolation and no overlap among the three sets.

[0048] III. Overall Model Network Construction: The model is divided into two main branches, which are then concatenated at the end and connected to the regression prediction head: 1. Image Branch: The EfficientNet-B0 backbone is modified by removing the last fully connected classification layer of the original ImageNet classification; all MBConv lightweight convolutions, SE attention modules, and global average pooling layers are retained; after pooling, a one-dimensional depth visual feature vector is output; 2. Artificial Feature Branch: The 14-dimensional feature vector, standardized by RobustScaler, is input and connected to two fully connected layers to increase the dimensionality and match the depth feature dimension; 3. Feature Fusion and Regression Head: The deep features output by EfficientNet and the 14-dimensional artificial features are concatenated and fused, and then fed into multiple fully connected layers to finally output a one-dimensional prediction value (thickness in logarithmic space).

[0049] IV. Training Hyperparameters and Loss Function Configuration, Pre-training Weight Initialization: EfficientNet-B0 loads publicly available ImageNet pre-trained weights for transfer learning: The underlying convolutional layers reuse general texture / color extraction capabilities, while only the upper layers are adapted for thin film thickness regression tasks. The loss function (for regression tasks) uses MSE (mean squared error) loss, fitting the logarithmically transformed thickness labels. It can also be combined with MAE joint constraints to reduce prediction bias in extreme thicknesses. Optimizer and learning rate: Optimizer: AdamW; Learning rate strategy: cosine annealing decay, initial learning rate... Batch Size: 8 / 16 (matching the memory usage of 224×224 images); Regularization to prevent overfitting; Dropout layer: Add Dropout after feature fusion to randomly deactivate some features; Data augmentation: Randomly and slightly adjust the color, brightness, and slightly translate and rotate the training set to simulate the small fluctuations in microscope illumination; Early Stop: Monitor the loss on the validation set, and stop training if it does not decrease for 20 consecutive rounds to prevent overfitting.

[0050] V. Complete Iterative Training Steps 1. First 5 rounds: Freeze the main intervention training weights. Freeze all convolutional layers of EfficientNet-B0 and train only the artificial feature branches and terminal regression heads. Purpose: To allow the network to learn the basic relationship between artificial color / geometric features and thickness, avoiding random weights from disrupting the pre-trained visual extraction capabilities.

[0051] 2. Unfreeze the entire network for joint training (main training phase): Unfreeze the EfficientNet-B0 weights, and the entire dual-branch network participates in parameter updates, with simultaneous optimization: MBConv convolutional kernel: automatically learns thin-film interference colors, surface textures, and subtle edge thickness features; SE attention module: automatically weights thickness-sensitive color channels, weakening dust and noise; artificial feature fully connected layer: matches the depth feature dimension, enhancing the ability to distinguish subtle color differences in thin layers. After each training epoch, inference is performed on the validation set, calculating metrics: logarithmic domain MSE, MAE after thickness restoration, and RMSE; the model weights with the optimal validation set error (optimal checkpoint) are saved. Early stopping criteria: If the validation set loss does not decrease for several consecutive epochs and the test set accuracy declines, training is terminated, and overfitted weights are discarded.

[0052] VI. Model Inference and Weight Consolidation (Post-Training Processing): Load the optimal EfficientNet-B0 weights from training; Inference Flow: Input a 224×224 material RGB image → EfficientNet extracts depth features; Simultaneously input 14-dimensional standardized artificial features of the same block; After feature fusion, the network outputs logarithmic thickness; Inverse exponential transformation yields nanometer thickness values; Model Deployment and Consolidation: Export the EfficientNet backbone, feature fusion layer, and regression head as a whole into an inference model, deploy it to the detection device terminal, and realize real-time single-image thickness prediction.

[0053] The core design objectives of the training are: Transfer learning of EfficientNet-B0: Capturing subtle interference color differences in nanofilms that are indistinguishable to the naked eye through lightweight multi-scale convolutions; Dual feature fusion: Automatic deep features from the network + manually designed 14-dimensional film physical features to compensate for the shortcomings of single CNNs in distinguishing ultrathin materials; Logarithmic labels + RobustScaler: Solving the problems of long-tailed samples and abnormal light spot noise, significantly improving the stability of thickness prediction; Transfer learning + step-by-step frozen training: Reducing the risk of overfitting caused by insufficient two-dimensional material annotation datasets, and achieving convergence even in small sample scenarios.

[0054] Embodiment 2 of the present invention provides a device for measuring the thickness of a CVD-synthesized two-dimensional material, comprising the following modules: an image acquisition module for acquiring a target image of the CVD-synthesized two-dimensional material to be processed; an image preprocessing module for removing the background of the target image and dividing the target image into several sub-images to be processed, each sub-image having the same color; and a thickness measurement module for performing the following processing on each sub-image to be processed: inputting the sub-image to be processed into a trained EfficientNet-B0 neural network to obtain the corresponding thickness of the sub-image to be processed.

[0055] In this embodiment, the image preprocessing module is further configured to: process the target image based on the trained U-Net image segmentation framework to obtain a black and white mask, and remove the background of the target image based on the black and white mask. The U-Net image segmentation framework uses a ResNet34 residual network as the encoder.

[0056] In this embodiment, the image preprocessing module is further configured to: divide the target image into several sub-regions using the SLIC superpixel segmentation algorithm, and perform clustering processing on the several sub-regions using the Mean Shift clustering algorithm to obtain several sub-images to be processed, each sub-image to be processed having the same color.

[0057] Embodiment 3 of the present invention provides an electronic device, including: a memory for storing executable instructions; and a processor for implementing the measurement method in Embodiment 1 when executing the executable instructions stored in the memory.

[0058] Embodiment 4 of the present invention provides a storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the measurement method in Embodiment 1.

[0059] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0060] This invention can be a system, method, and / or computer program product. A computer program product may include a readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0061] A readable storage medium can be a tangible device that holds and stores instructions for use by an instruction execution device. Readable storage media can include, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0062] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for measuring the thickness of a CVD-synthesized two-dimensional material, characterized in that, Includes the following steps: Obtain the target image of the CVD-synthesized two-dimensional material to be processed; Remove the background from the target image and divide the target image into several sub-images to be processed, each sub-image having the same color. For each sub-image to be processed, the following processing is performed: the sub-image to be processed is input into the trained EfficientNet-B0 neural network to obtain the corresponding thickness of the sub-image to be processed.

2. The measurement method according to claim 1, characterized in that, The removal of the background from the target image specifically includes: The target image is processed based on the trained U-Net image segmentation framework to obtain a black and white mask. The background of the target image is removed based on the black and white mask. The U-Net image segmentation framework uses a ResNet34 residual network as the encoder.

3. The measurement method according to claim 1, characterized in that, The step of dividing the target image into several sub-images to be processed, wherein each sub-image to be processed has the same color, specifically includes: The target image is divided into several sub-regions using the SLIC superpixel segmentation algorithm. The mean shift clustering algorithm is then used to cluster these sub-regions, resulting in several sub-images to be processed, each with the same color.

4. A device for measuring the thickness of a CVD-synthesized two-dimensional material, characterized in that, Includes the following modules: The image acquisition module is used to acquire target images of the two-dimensional material to be synthesized by CVD. The image preprocessing module is used to remove the background of the target image and divide the target image into several sub-images to be processed, each sub-image having the same color. The thickness measurement module is used to perform the following processing on each sub-image to be processed: inputting the sub-image to be processed into a trained EfficientNet-B0 neural network to obtain the corresponding thickness of the sub-image to be processed.

5. The measuring device according to claim 4, characterized in that, The image preprocessing module is also used for: The target image is processed based on the trained U-Net image segmentation framework to obtain a black and white mask. The background of the target image is removed based on the black and white mask. The U-Net image segmentation framework uses a ResNet34 residual network as the encoder.

6. The measuring device according to claim 4, characterized in that, The image preprocessing module is also used for: The target image is divided into several sub-regions using the SLIC superpixel segmentation algorithm. The mean shift clustering algorithm is then used to cluster these sub-regions, resulting in several sub-images to be processed, each with the same color.

7. An electronic device, characterized in that, include: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the measurement method according to any one of claims 1 to 3.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the measurement method according to any one of claims 1 to 3.

Images