Surface treatment metal thin layer quality analysis method and system

Abstract

The invention relates to the field of material science and surface engineering, and provides a surface treatment metal thin layer quality analysis method and system, and the method comprises the steps: collecting the data of a metal thin layer through a multi-mode detection system; processing data acquired by the multi-modal detection system through a convolutional neural network, extracting and fusing feature vectors, and generating classification results of surface defects, internal defects and chemical component distribution; collecting data in a thin layer forming process through an ultrafast laser spectrometer and a micro sensor which are embedded in the surface treatment equipment, and adjusting deposition parameters of the surface treatment equipment through a reinforcement learning algorithm; and updating the weight of the convolutional neural network through an incremental learning algorithm, and outputting an analysis report including defect types. By integrating computer vision, Raman spectrum and terahertz wave imaging, comprehensive analysis of surface defects, internal defects and chemical components of the metal thin layer is realized, the detection comprehensiveness is remarkably improved, and physical and chemical characteristics can be captured at the same time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material science and surface engineering, and particularly relates to a surface treatment metal thin layer quality analysis method and system. BACKGROUND

[0002] In the prior art, the quality analysis method of the surface treatment metal thin layer mainly includes optical imaging, profile measurement and spectral analysis. These methods detect surface defects, thickness or composition distribution by collecting surface image, geometric shape or chemical property data of the metal thin layer, and are widely used in the fields of semiconductors, aerospace and photovoltaics.

[0003] However, the prior art usually relies on a single detection means, which is difficult to simultaneously obtain the physical defects and chemical properties of the metal thin layer, resulting in insufficient detection comprehensiveness. Single modal analysis often cannot meet the needs of semiconductor devices or aerospace coatings and other applications in complex multi-layer structures or high-precision scenarios, limiting the precision and scope of quality analysis. SUMMARY

[0004] The present application aims to at least solve one of the problems existing in the prior art, and provides a surface treatment metal thin layer quality analysis method and system to solve the problem that the prior art usually relies on a single detection means, which is difficult to simultaneously obtain the physical defects and chemical properties of the metal thin layer, resulting in insufficient detection comprehensiveness.

[0005] In one aspect of the present application, a surface treatment metal thin layer quality analysis method is provided, comprising the following steps: Step S1, collecting data of the metal thin layer using a multi-modal detection system, wherein the multi-modal detection system comprises a CMOS camera, a Raman spectrometer and a terahertz wave imager; Step S2, processing the data collected by the multi-modal detection system through a convolutional neural network, extracting and fusing feature vectors, and generating classification results of surface defects, internal defects and chemical composition distribution; Step S3, collecting data in the thin layer formation process through an ultrafast laser spectrometer and a micro sensor embedded in a surface treatment device, and adjusting deposition parameters of the surface treatment device through a reinforcement learning algorithm; Step S4, updating the weights of the convolutional neural network through an incremental learning algorithm, and outputting an analysis report including defect types.

[0006] Optionally, the resolution of the CMOS camera is 10 million to 20 million pixels, the CMOS camera is used to collect surface image data of the metal thin layer, and the surface defects include cracks and cavities.

[0007] Optionally, the Raman spectrometer uses a laser with a wavelength of 500 nm to 600 nm to acquire molecular vibrational spectral data of the metal thin layer, and the chemical composition distribution includes oxide content.

[0008] Optionally, the terahertz wave imager operates at a frequency of 0.1 terahertz to 4 terahertz, and is used to acquire internal structural data of thin metal layers, including delamination and bubbles.

[0009] Optionally, step S2 includes the following sub-steps: Sub-step S201: Extract the feature vectors of the metal thin layer data acquired by the CMOS camera, the Raman spectrometer, and the terahertz wave imager, respectively denoted as... , and ; Sub-step S202: Merge feature vectors using a weighted concatenation method. The calculation formula is as follows: ; in, For image texture feature vectors, For spectral peak eigenvectors, This is a terahertz depth feature vector. The merged feature vector, , , They are respectively , , The corresponding weighting coefficients; Sub-step S203: Input the merged feature vector into the classification layer to obtain the classification results of surface defects, internal defects and chemical composition distribution output by the classification layer.

[0010] Optionally, the ultrafast laser spectrometer has a pulse width of 50-150 femtoseconds, a wavelength of 700-900 nanometers, and a sampling frequency of once every 5-15 milliseconds. The ultrafast laser spectrometer is used to collect thickness and crystal structure data of metal thin layers.

[0011] Optionally, the microsensor has a size of 0.5 cubic centimeters to 1.5 cubic centimeters and is integrated into a physical vapor deposition or chemical vapor deposition device to collect temperature and pressure data during the metal thin-film formation process.

[0012] Optionally, the reinforcement learning algorithm in step S3 is based on a deep Q-network, and step S3 specifically includes the following sub-steps: Sub-step S301: Collect the thickness and preset threshold data of the metal thin layer; Sub-step S302: Calculate the reward function, the formula is: ; in, This is the reward value for the current round. For defect density, For thickness uniformity, For thickness deviation, To preset the target thickness, , , All are weighting coefficients; Sub-step S303: Adjust the deposition rate and temperature parameters of the surface treatment equipment according to the reward function.

[0013] Optionally, the incremental learning algorithm in step S4 uses the lifelong continuous learning-plastic weight consolidation method, and step S4 specifically includes the following sub-steps: Sub-step S401: Calculate the loss function for the current task; Sub-step S402: Use the Fisher information matrix to protect historical weights, and update the formula as follows: ; in, This represents the total loss after the incremental update. The classification loss for the current task, For the first convolutional neural network The current values ​​of the parameters, For the first in the historical model The old values ​​of the parameters, For the first Fisher information values ​​for each parameter. The regularization coefficient is used. Sub-step S403: Adjust weights to handle new metal thin-film materials and defect types.

[0014] In another aspect, the present invention provides a surface-treated metal thin film quality analysis system for implementing the surface-treated metal thin film quality analysis method described above, the surface-treated metal thin film quality analysis system comprising: The multimodal detection device includes a CMOS camera with a resolution of 10-20 megapixels, a Raman spectrometer with a wavelength of 500-600 nanometers, and a terahertz wave imager with a working frequency of 0.1-4 terahertz. The in-situ analysis device, including an ultrafast laser spectrometer with a pulse width of 50-150 femtoseconds and a miniature sensor with a size of 0.5-1.5 cubic centimeters, is integrated into the surface treatment equipment. The data processing device is used to run a convolutional neural network and an incremental learning algorithm to execute steps S2 and S4 in the surface-treated metal thin film quality analysis method described above. The control device is used to run the reinforcement learning algorithm, execute step S3 in the surface treatment metal thin film quality analysis method described above, and communicate with the surface treatment equipment.

[0015] Compared with the prior art, the present invention has the following advantages: 1. This invention integrates computer vision, Raman spectroscopy and terahertz wave imaging to achieve comprehensive analysis of surface defects, internal defects and chemical composition of thin metal layers, significantly improving the comprehensiveness of detection and enabling the simultaneous capture of physical and chemical properties.

[0016] 2. This invention transforms the detection results into quantifiable quality indicators and automatically generates process parameter adjustment instructions such as deposition rate and substrate temperature, thereby realizing closed-loop adaptive control of the production process. Compared with existing methods that rely on manual experience or fixed parameters, it can significantly reduce the probability of defects and improve film consistency.

[0017] 3. This invention adopts an incremental learning algorithm to dynamically update the weights of the convolutional neural network, adapting to different thin metal materials and new defect types. It eliminates the need for frequent retraining, reduces model maintenance costs, enhances cross-material and cross-scenario versatility, and is suitable for changing production environments. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0019] Figure 1 A flowchart of a method for analyzing the quality of surface-treated metal thin films according to an embodiment of the present invention; Figure 2 A multimodal data processing flowchart of a surface-treated metal thin-film quality analysis method provided in another embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0021] One embodiment of the present invention relates to a method for quality analysis of surface-treated metal thin films, the process of which is as follows: Figure 1 As shown, it includes the following steps: Step S1: Use a multimodal detection system to acquire data of the metal thin layer. The multimodal detection system includes a CMOS camera, a Raman spectrometer, and a terahertz wave imager. Step S2: Process the data collected by the multimodal detection system through a convolutional neural network, extract and fuse feature vectors, and generate classification results for surface defects, internal defects, and chemical composition distribution; Step S3: Data during the thin-layer formation process is collected using an ultrafast laser spectrometer and micro-sensors embedded in the surface treatment equipment, and the deposition parameters of the surface treatment equipment are adjusted using a reinforcement learning algorithm; Step S4: Update the weights of the convolutional neural network using an incremental learning algorithm, and output an analysis report including the defect type.

[0022] Specifically, in step S1, the multimodal detection system acquires two-dimensional image data of the metal thin layer surface using a CMOS camera as surface image data, obtains molecular vibrational spectrum data of the metal thin layer using a Raman spectrometer to detect chemical composition such as oxide and impurity distribution, and scans the internal structure of the metal thin layer using a terahertz wave imager to obtain three-dimensional defect distribution data. In step S2, the convolutional neural network first preprocesses the surface image data of the metal thin layer, including grayscale conversion and edge detection, and then extracts feature vectors from the spectral and terahertz data. These feature vectors are then weighted and concatenated to generate comprehensive features for classifying surface cracks, voids, internal delamination, and chemical inhomogeneities. Step S3 is performed in a physical vapor deposition (PVD) setup. An ultrafast laser spectrometer and micro-sensors acquire the thickness, crystal structure, and environmental parameters of the metal thin layer in real time. A reinforcement learning algorithm compares the acquired data with preset thresholds and adjusts the deposition rate and temperature accordingly. In step S4, an incremental learning algorithm updates the weights of the convolutional neural network based on newly acquired data, generating an analysis report containing defect types, locations, and chemical distributions. The analysis report is output in heatmap and tabular form for operator reference.

[0023] Step S1 aims to provide comprehensive raw data for subsequent defect analysis. The multimodal detection system includes a CMOS camera with a resolution of 10-20 megapixels and a wavelength of... The Raman spectrometer and the terahertz wave imager with a working frequency of 0.1THz-4THz were used to simultaneously acquire information about different dimensions of the metal thin layer. A CMOS camera acquires a one-dimensional image of the surface morphology of a thin metal layer as surface image data, which is then used to detect cracks and voids. Raman spectrometers acquire molecular vibrational spectral data of thin metal layers, and use this data to identify oxide content and chemical impurity distribution. Terahertz wave imagers acquire three-dimensional internal structure images of thin metal layers as internal structure data, which are then used to detect hidden defects such as delamination and bubbles.

[0024] Three types of raw data It will then be sent to the next step for feature extraction. This is surface image data of a thin metal layer. Molecular vibrational spectral data for thin metal layers. This is the internal structure data of a thin metal layer.

[0025] CMOS cameras have a resolution of 10-20 megapixels. They are used to acquire surface image data of thin metal layers, including surface defects such as cracks and voids.

[0026] Specifically, a CMOS camera can be deployed above the inspection platform to acquire high-resolution image data of the metal thin layer surface through continuous shooting as surface image data of the metal thin layer. Image processing, including noise filtering and edge enhancement, is performed on the surface image data to detect surface defects with crack widths of 1-10 micrometers and void diameters of 5-50 micrometers.

[0027] Raman spectrometers use lasers with wavelengths of 500-600 nanometers to collect molecular vibrational spectral data of thin metal layers, including the distribution of chemical composition, including oxide content.

[0028] Specifically, the Raman spectrometer collects molecular vibrational spectral data by irradiating the surface of a thin metal layer with an optical fiber probe. The molecular vibrational spectral data is then processed, including baseline correction and peak fitting, to identify the oxide content and impurity distribution, and generate a two-dimensional distribution map of the chemical composition.

[0029] Terahertz wave imagers operate at frequencies of 0.1 terahertz to 4 terahertz. They are used to acquire internal structural data of thin metal layers, including internal defects such as delamination and bubbles.

[0030] Specifically, the terahertz wave imager generates a three-dimensional image of the internal structure of the metal thin layer by scanning the metal thin layer in transmission mode. This image serves as the internal structure data of the metal thin layer. The internal structure data is then processed, including time-domain signal conversion and spatial reconstruction, to identify internal defects with a layer thickness of 1 micrometer to 10 micrometers and a bubble diameter of 5 micrometers to 20 micrometers.

[0031] Combined Figure 2 Step S2 includes the following sub-steps: Sub-step S201: Extract the feature vectors of the metal thin layer data acquired by the CMOS camera, Raman spectrometer, and terahertz wave imager, respectively denoted as... , and ; Sub-step S202: Merge feature vectors using a weighted concatenation method. The calculation formula is as follows: ; Sub-step S203: Input the merged feature vector into the classification layer and output the classification results of surface defects, internal defects and chemical composition distribution.

[0032] Specifically, step S2 aims to transform the multimodal data collected in step S1 into structured quality indicators, which serve as the basis for process control.

[0033] Convolutional Neural Networks (CNNs) extract texture feature vectors from CMOS images. Peak characteristics of Raman spectra and the depth characteristics of terahertz waveforms The fused feature vector, i.e., the merged feature vector, is obtained through weighted concatenation. : .

[0034] in, It is a 128-dimensional image texture feature vector. It is a 64-dimensional spectral peak eigenvector. It is a terahertz depth feature vector with a dimension of 64; The weighting coefficients corresponding to the three modal features are... , , The corresponding weight coefficients, with an initial value range of 0.2-0.4, are automatically optimized during training using the gradient descent algorithm.

[0035] Fusion feature vectors After inputting into the Softmax classification layer, the Softmax classification layer outputs a defect probability map, identifying types such as cracks, voids, delamination, bubbles, and chemical inhomogeneities. Based on the defect classification results, it calculates three quality indicators: defect density. Thickness uniformity and thickness deviation .

[0036] The ultrafast laser spectrometer has a pulse width of 50-150 femtoseconds, a wavelength of 700-900 nanometers, and a sampling frequency of once every 5-15 milliseconds. The ultrafast laser spectrometer is used to collect thickness and crystal structure data of thin metal layers.

[0037] Specifically, the ultrafast laser spectrometer is integrated into the cavity of the surface treatment equipment. Pulsed laser is transmitted through optical fiber. Data acquisition includes the spectral reflection signal of the thickness of the metal thin layer and the Raman scattering signal of the crystal structure. After processing the acquired spectral reflection signal of the thickness of the metal thin layer and the Raman scattering signal of the crystal structure, thickness curve and crystal diffraction peak shift map are generated respectively.

[0038] The microsensors, ranging in size from 0.5 to 1.5 cubic centimeters, are integrated into physical vapor deposition or chemical vapor deposition equipment to collect temperature and pressure data during the formation of thin metal layers.

[0039] Specifically, the micro-sensors transmit data via a wireless module to record temperature and pressure changes in real time during the deposition process. The recorded data is used to analyze the uniformity of the metal thin film and the causes of defect formation.

[0040] The reinforcement learning algorithm in step S3 is based on a deep Q-network, and step S3 specifically includes the following sub-steps: Sub-step S301: Collect the thickness and preset threshold data of the metal thin layer; Sub-step S302: Calculate the reward function, the formula is: ; Sub-step S303: Adjust the deposition rate and temperature parameters of the surface treatment equipment according to the reward function.

[0041] Specifically, step S3 aims to dynamically adjust the deposition parameters based on the defect results identified in step S2, thereby achieving closed-loop quality control. as a state vector , with deposition rate and substrate temperature As an action A deep Q-network (DQN) is built, and the policy update is guided by the following reward function: .

[0042] in, This is the reward value for the current round (dimensionless). Preset target thickness; , , These are the weighting coefficients for the three indicators, with recommended values ​​ranging from 0.4 to 0.6, 0.2 to 0.4, and 0.1 to 0.3, respectively, to balance the effects of defect density, uniformity, and thickness deviation. The Q-value iterative update formula is: .

[0043] in, For state Take action Expected cumulative reward The learning rate (can be set to 0.01). This is the discount factor (which can be set to 0.9). For the next state The set of optional actions below, For the next state The maximum expected cumulative reward.

[0044] When an increase in defect density or thickness deviation is detected, the control device automatically reduces the deposition rate (0.1 nm / s–1 nm / s) and increases the temperature (100°C–500°C) to correct the deposition process in real time. The new process status data will be re-sent to step S1 to form a feedback loop.

[0045] The incremental learning algorithm in step S4 uses a lifelong continuous learning-plastic weight consolidation method. Step S4 specifically includes the following sub-steps: Sub-step S401: Calculate the loss function for the current task; Sub-step S402: Use the Fisher information matrix to protect historical weights, and update the formula as follows: ; Sub-step S403: Adjust weights to handle new metal thin-film materials and defect types.

[0046] Specifically, step S4 is used to ensure that the model maintains high accuracy in identifying defects when process conditions change or new defect types appear.

[0047] After the adjusted process in step S3 generates new defect samples, these new samples are added to the training set, and the weights of the CNN are updated using the Elastic Weight Consolidation (EWC) algorithm. The calculation formula is as follows: .

[0048] in, This represents the total loss after the incremental update. The classification loss for the current task, For the first convolutional neural network The current values ​​of the parameters, For the first in the historical model The old values ​​of the parameters, For the first Fisher information values ​​for each parameter. The regularization coefficient is used. An EWC update is triggered every 100 new samples collected, ensuring the model maintains its original defect recognition performance while learning new defect features (such as peeling). The updated model immediately replaces the classification network in step S2, ensuring that subsequent detection results always match the current process state, thus achieving system self-evolution.

[0049] Another embodiment of the present invention relates to a surface-treated metal thin film quality analysis system, comprising: The multimodal detection device includes a CMOS camera with a resolution of 10-20 megapixels, a Raman spectrometer with a wavelength of 500-600 nanometers, and a terahertz wave imager with a working frequency of 0.1-4 terahertz. The in-situ analysis device, including an ultrafast laser spectrometer with a pulse width of 50-150 femtoseconds and a miniature sensor with a size of 0.5-1.5 cubic centimeters, is integrated into the surface treatment equipment. A data processing device is used to run a convolutional neural network and an incremental learning algorithm to execute steps S2 and S4 in the surface-treated metal thin film quality analysis method. The control device is used to run a reinforcement learning algorithm, execute step S3 in the surface treatment metal thin film quality analysis method, and communicate with the surface treatment equipment.

[0050] The surface-treated metal thin film quality analysis system provided in this embodiment of the invention is used to implement the surface-treated metal thin film quality analysis method provided in this embodiment of the invention.

[0051] Specifically, the multimodal detection device is used to perform full-domain detection of the metal thin layer in the post-deposition stage. It acquires surface morphology images through a CMOS camera, obtains chemical composition distribution information through a Raman spectrometer, and acquires internal structure images through a terahertz wave imager, thereby obtaining multi-source data such as the macroscopic appearance, molecular structure, and three-dimensional defects of the metal thin layer.

[0052] The in-situ analysis device is integrated into the surface treatment equipment to monitor key state parameters in real time during the deposition process. The in-situ analysis device includes an ultrafast laser spectrometer and a micro-sensor array. The ultrafast laser spectrometer is used to excite and capture transient spectral signals during the thin-layer formation process, while the micro-sensor array is used to collect process parameters such as temperature, stress, and thickness growth rate, enabling in-situ dynamic sensing of the deposition process.

[0053] The data processing device receives the aforementioned multimodal detection data and in-situ analysis data, fuses and processes them, and outputs quality indicators such as defect identification results, defect density, thickness uniformity, and thickness deviation. It also performs self-optimization and updates on historical detection data to maintain the system's adaptability to different materials and defect types.

[0054] The control device establishes a communication connection with the surface treatment equipment. Based on the quality indicators and defect results output by the data processing device, it automatically generates adjustment commands such as deposition rate and substrate temperature and sends them to the equipment. This enables adaptive closed-loop control of the deposition process, reduces defect generation, and improves film consistency.

[0055] The surface-treated metal thin-film quality analysis method of the present invention will be described below with reference to specific embodiments.

[0056] This embodiment provides a method for analyzing the quality of surface-treated metal thin films, and verifies its effectiveness and feasibility.

[0057] I. Experimental conditions and sample preparation.

[0058] A 50 nm thick titanium thin film was selected and prepared using magnetron sputtering physical vapor deposition at a deposition temperature of 300°C and a cavity substrate vacuum of 5 × 10⁻⁶. -3 Pa, deposition rate of 0.5 nm / s, sample size of 20 mm × 20 mm, and cooled to room temperature after preparation for later use.

[0059] 2. Step S1: Collect data using a multimodal detection system.

[0060] The multimodal detection system includes a CMOS camera, a Raman spectrometer, and a terahertz wave imager, which simultaneously acquire information about different dimensions of the metal thin layer. A CMOS camera acquires two-dimensional images of the surface topography of a thin metal layer as surface image data, which is then used to detect cracks and voids. Raman spectrometers acquire molecular vibrational spectral data of thin metal layers, and use this data to identify oxide content and chemical impurity distribution. Terahertz wave imagers acquire three-dimensional internal structure images of thin metal layers as internal structure data, which are then used to detect hidden defects such as delamination and bubbles.

[0061] Three types of raw data It is then sent to the next step for feature extraction.

[0062] 3. Step S2: Analyze and fuse data using a convolutional neural network.

[0063] Convolutional neural networks extract texture feature vectors from CMOS images. Peak characteristics of Raman spectra and the depth characteristics of terahertz waveforms The merged feature vector is obtained by weighted concatenation. : .

[0064] in, For image texture feature vectors, For spectral eigenvectors, For terahertz eigenvectors, These are the weight coefficients corresponding to the three modal features, which are automatically optimized during training.

[0065] Fusion feature vectors After inputting into the Softmax classification layer, the Softmax classification layer outputs a defect probability map, identifying types such as cracks, voids, delamination, bubbles, and chemical inhomogeneities, and calculates the defect density based on the defect classification results. Thickness uniformity and thickness deviation In this embodiment, the convolutional neural network includes 3 convolutional layers, 2 pooling layers, and 2 fully connected layers. The training set contains 1200 sets of image, spectral, and terahertz data. After 100 rounds of training, the accuracy of the convolutional neural network reaches [percentage missing]. above.

[0066] IV. Step S3: Perform reinforcement learning-driven adaptive control of the deposition process.

[0067] by as a state vector , with deposition rate and substrate temperature As an action Establish depth The network guides policy updates through the following reward function: .

[0068] in, This is the reward value for the current round. To preset the target thickness, , , These are the weighting coefficients for the three indicators.

[0069] The formula for iteratively updating the Q value is: .

[0070] in, For state Take action Expected cumulative reward For learning rate, As a discount factor, For the next state The set of optional actions below, For the next state The maximum expected cumulative reward.

[0071] When an increase in defect density or thickness deviation is detected, the control device automatically reduces the deposition rate and increases the temperature. The new process status data will be re-sent to step S1, forming a feedback loop.

[0072] 5. Step S4: Update the model online using an incremental learning algorithm.

[0073] After the adjusted process in step S3 generates new defect samples, these new samples are added to the training set, and the weights of the convolutional neural network are updated using the EWC algorithm. The calculation formula is as follows: .

[0074] in, This represents the total loss after the incremental update. The classification loss for the current task, For the first convolutional neural network The current values ​​of the parameters, For the first in the historical model The old values ​​of the parameters, For the first Fisherf information values ​​for each parameter. This is the regularization coefficient.

[0075] An update is triggered every time a certain number of new samples are collected, so that the model can maintain its original defect recognition performance when learning new defect features. The updated model replaces the classification network in step S2, so that the subsequent detection results always match the current process status.

[0076] VI. Experimental Results and Effect Verification.

[0077] In the testing of multiple titanium thin-film samples, the surface-treated metal thin-film quality analysis method provided in this embodiment of the invention took an average of approximately 6.5 seconds per sample. Results showed that all cracks and voids were successfully detected; internal delamination and bubble defects were identified; Raman spectroscopy showed increased TiO2 peak intensity in the edge region, indicating localized oxidation; compared with single optical imaging methods, the defect false negative rate was significantly reduced, and the average detection time was significantly shortened.

[0078] The surface treatment metal thin film quality analysis system provided in this embodiment of the invention can monitor the thin film quality in real time during the deposition process after being deployed on the production line. When the reinforcement learning algorithm determines that the defect density is increasing, it automatically adjusts the deposition parameters, significantly reducing the scrap rate and realizing online closed-loop quality control.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing the quality of surface-treated metal thin films, characterized in that, Includes the following steps: Step S1: Collect data of the metal thin layer using a multimodal detection system, which includes a CMOS camera, a Raman spectrometer, and a terahertz wave imager; Step S2: Process the data collected by the multimodal detection system through a convolutional neural network, extract and fuse feature vectors, and generate classification results for surface defects, internal defects, and chemical composition distribution; Step S3: Data during the thin-layer formation process is collected using an ultrafast laser spectrometer and micro-sensors embedded in the surface treatment equipment, and the deposition parameters of the surface treatment equipment are adjusted using a reinforcement learning algorithm; Step S4: Update the weights of the convolutional neural network using an incremental learning algorithm, and output an analysis report including the defect type.

2. The method for analyzing the quality of surface-treated metal thin films according to claim 1, characterized in that, The CMOS camera has a resolution of 10-20 megapixels and is used to acquire surface image data of a thin metal layer, including cracks and voids.

3. The method for analyzing the quality of surface-treated metal thin films according to claim 1, characterized in that, The Raman spectrometer uses a laser with a wavelength of 500 nm to 600 nm to collect molecular vibrational spectral data of the metal thin layer, and the chemical composition distribution includes oxide content.

4. The method for analyzing the quality of surface-treated metal thin films according to claim 1, characterized in that, The terahertz wave imager operates at a frequency of 0.1 terahertz to 4 terahertz and is used to acquire internal structural data of thin metal layers, including internal defects such as delamination and bubbles.

5. The method for analyzing the quality of surface-treated metal thin films according to claim 1, characterized in that, Step S2 includes the following sub-steps: Sub-step S201: Extract the feature vectors of the metal thin layer data acquired by the CMOS camera, the Raman spectrometer, and the terahertz wave imager, respectively denoted as... , and ; Sub-step S202: Merge feature vectors using a weighted concatenation method. The calculation formula is as follows: ； in, For image texture feature vectors, For spectral peak eigenvectors, This is a terahertz depth feature vector. The merged feature vector, , , They are respectively , , The corresponding weighting coefficients; Sub-step S203: Input the merged feature vector into the classification layer to obtain the classification results of surface defects, internal defects and chemical composition distribution output by the classification layer.

6. The method for analyzing the quality of surface-treated metal thin films according to claim 1, characterized in that, The ultrafast laser spectrometer has a pulse width of 50-150 femtoseconds, a wavelength of 700-900 nanometers, and a sampling frequency of once every 5-15 milliseconds. The ultrafast laser spectrometer is used to collect thickness and crystal structure data of thin metal layers.

7. The method for analyzing the quality of surface-treated metal thin films according to claim 1, characterized in that, The microsensor has a size of 0.5 cubic centimeters to 1.5 cubic centimeters and is integrated into a physical vapor deposition or chemical vapor deposition device to collect temperature and pressure data during the formation of a thin metal layer.

8. The method for analyzing the quality of surface-treated metal thin films according to claim 1, characterized in that, The reinforcement learning algorithm in step S3 is based on a deep Q-network, and step S3 specifically includes the following sub-steps: Sub-step S301: Collect the thickness and preset threshold data of the metal thin layer; Sub-step S302: Calculate the reward function, the formula is: ； in, This is the reward value for the current round. For defect density, For thickness uniformity, For thickness deviation, To preset the target thickness, , , All are weighting coefficients; Sub-step S303: Adjust the deposition rate and temperature parameters of the surface treatment equipment according to the reward function.

9. The method for analyzing the quality of surface-treated metal thin films according to claim 1, characterized in that, The incremental learning algorithm in step S4 uses a lifelong continuous learning-plastic weight consolidation method. Step S4 specifically includes the following sub-steps: Sub-step S401: Calculate the loss function for the current task; Sub-step S402: Use the Fisher information matrix to protect historical weights, and update the formula as follows: ； in, This represents the total loss after the incremental update. The classification loss for the current task. For the first convolutional neural network The current values ​​of the parameters, For the first in the historical model The old values ​​of the parameters, For the first Fisher information values ​​for each parameter. The regularization coefficient is used. Sub-step S403: Adjust weights to handle new metal thin-film materials and defect types.

10. A surface-treated metal thin-film quality analysis system, characterized in that, For implementing the surface-treated metal thin film quality analysis method according to any one of claims 1 to 9, the surface-treated metal thin film quality analysis system comprises: The multimodal detection device includes a CMOS camera with a resolution of 10-20 megapixels, a Raman spectrometer with a wavelength of 500-600 nanometers, and a terahertz wave imager with a working frequency of 0.1-4 terahertz. The in-situ analysis device, including an ultrafast laser spectrometer with a pulse width of 50-150 femtoseconds and a miniature sensor with a size of 0.5-1.5 cubic centimeters, is integrated into the surface treatment equipment. A data processing device is used to run a convolutional neural network and an incremental learning algorithm to execute steps S2 and S4 in the surface-treated metal thin film quality analysis method. The control device is used to run a reinforcement learning algorithm, execute step S3 in the surface treatment metal thin film quality analysis method, and communicate with the surface treatment equipment.

Images