Intelligent identification method and device for atomic-scale laser processing wafer substrate

Abstract

The invention relates to an intelligent identification method and device for an atomic-scale laser processing wafer substrate, belongs to the technical field of laser processing, and solves the problem that a wafer substrate with an ultra-smooth surface, a complete and defect-free surface lattice and no surface or subsurface damage is difficult to obtain at the same time in the existing laser finishing technology. The method comprises the following steps: collecting defects and morphology feature information generated in the interaction process of laser finishing and a wafer substrate surface atomic layer, and storing the defects and morphology feature information in a feature database; establishing an inversion deep learning model and carrying out training and testing; identifying propagable destructive defects and rough surface microtopography structures by using the optimal prediction model to obtain early warning, prediction or prevention and treatment information, and feeding back the information to the industrial control system; and adjusting atomic-scale laser finishing parameters to form closed loop control. According to the method, in the wafer substrate laser finishing process, defects are early warned, predicted and prevented in advance, and the defects generated in the wafer substrate atomic layer removing process are rapidly, accurately and effectively recognized.

Description

Technical Field

[0001] The present invention relates to the technical field of laser processing, and in particular, to an intelligent recognition method and device for atomic-level laser processing of a wafer substrate. Background Art

[0002] Wafers are widely used in fields such as aerospace, electronics, automotive, and marine as ideal substrate materials for various microelectronics and integrated circuit chips. Before using crystal materials such as Si and SiC as wafer substrates, multiple processes such as trimming and grinding, slicing, mechanical grinding, chemical mechanical polishing, cleaning, and detection are required to ensure their quality and reliability.

[0003] Among them, mechanical grinding and chemical mechanical polishing are key processes for removing cutting marks and damaged layers on the wafer substrate and achieving planarization and super-smoothness. However, due to the high hardness, brittleness, and strong chemical inertness of these crystal materials themselves, it is easy to cause defect damage and load failure on the wafer substrate, as well as the defect of low material removal efficiency, which will directly spread widely during the wafer epitaxy process.

[0004] Laser finishing, as one of the important technologies for surface finishing, has received high attention from domestic and foreign research institutions and related enterprises in recent years by regulating the surface finish of workpieces through selective melting, redistribution, and precise removal of materials by the movement of the molten pool. However, the existing laser finishing methods are difficult to obtain a wafer substrate with super-smooth surface, intact surface lattice without defects, and no surface or subsurface damage at the same time.

[0005] With the improvement of the development requirements of intelligent manufacturing, artificial intelligence technologies represented by deep learning have been widely penetrated into various fields of industry. In the application of object defect detection and recognition, the demand for using image processing technology is increasing. In order to early warn, predict, and prevent a series of defects that occur during the laser finishing of the wafer substrate, it is necessary to combine the ability of deep learning to quickly identify object features with laser finishing. Therefore, in this field, it is particularly crucial to establish an intelligent database and combine deep learning methods to quickly, accurately, and effectively identify the processing defects that occur during the laser finishing of the wafer substrate, especially the series of defects generated during the atomic layer removal process of the wafer substrate, and to accurately quantify them.

[0006] Therefore, it is urgent to develop a processing method that is simple and highly intelligent, can efficiently and accurately identify and predict various defects that occur during the atomic-level laser finishing of the wafer substrate, and perform accurate quantitative analysis on these defects, so as to obtain a complete set of methods for intelligent laser finishing of the wafer substrate with a super-smooth surface, a complete and defect-free surface lattice, and no surface or subsurface damage, so as to replace the two traditional processes of mechanical grinding and chemical mechanical polishing of the wafer substrate, and at the same time improve the manufacturing material utilization rate and processing efficiency of the wafer substrate. This has become an urgent problem for current scientific researchers to solve. Summary of the Invention

[0007] In view of the above analysis, embodiments of the present invention aim to provide an intelligent recognition method and device for atomic-level laser processing of wafer substrates, so as to solve the problem that it is difficult to obtain a wafer substrate with a super-smooth surface, a complete and defect-free surface lattice, and no surface or subsurface damage by existing laser finishing methods.

[0008] On the one hand, embodiments of the present invention provide an intelligent recognition method for atomic-level laser processing of wafer substrates, including: collecting defect and morphology feature information generated during the interaction between laser finishing and the atomic layer on the surface of the wafer substrate and storing it in a feature database; establishing an inversion deep learning model and using the defect and morphology feature information to perform distributed collaborative training and integrated regression testing on the inversion deep learning model to obtain an optimal prediction model; using the optimal prediction model to identify the propagable destructive defects and rough surface micro-morphology structures that occur during the laser finishing of the wafer substrate, so as to obtain early warning information, prediction information or prevention and control information about the defects and morphology during the laser finishing process and feedback it to the laser industrial control system; and the laser industrial control system adjusts the atomic-level laser finishing parameters according to the early warning information, prediction information or prevention and control information to form a closed-loop control, so as to precisely regulate and improve the quality of the laser processing technology of the wafer substrate.

[0009] The beneficial effects of the above technical solutions are as follows: The intelligent recognition method for atomic-level laser processing of wafer substrates according to embodiments of the present invention can monitor and predict processing defects in real time, avoid the large-scale spread of defects, and improve the manufacturing material utilization rate. Secondly, through accurate quantitative analysis of defects, the system can precisely process different defect types and degrees, avoid the problem of over-grinding or over-polishing that may occur in traditional methods, and improve the processing efficiency. Most importantly, the intelligent laser finishing method can obtain a wafer substrate with a super-smooth surface, a complete and defect-free surface lattice, and no surface or subsurface damage, meeting the requirements of high-precision microelectronics and integrated circuit chip manufacturing.

[0010] Based on further improvements to the above method, the defect and morphological feature information generated during the interaction between laser finishing and the atomic layer on the surface of the wafer substrate includes: collecting the morphological information, crystal structure information, defect and interface information, and composition analysis information of the wafer substrate through a transmission electron microscope; collecting the surface morphological information, surface mechanical properties, and surface charge distribution information through an atomic force microscope; and collecting the molecular structure information, lattice vibration information, and chemical composition analysis information of the wafer substrate through Raman spectroscopy.

[0011] Based on further improvements to the above method, the defect and morphological feature information generated during the interaction between laser finishing and the atomic layer on the surface of the wafer substrate includes: comprehensively collecting the defect and morphological feature information generated during the interaction between laser finishing and the atomic layer on the surface of the wafer substrate by combining the transmission electron microscope, the atomic force microscope, and the Raman spectroscopy with region-by-region and type-by-type sorting. Among them, the region-by-region means obtaining representative data on the surface of the wafer substrate by selecting different regions or sample points to reveal the regional distribution of surface defects; the type-by-type means distinguishing different types of defects and using different characterization methods to identify the morphology of the defects. Among them, different types of defects include point defects, line defects, and surface defects; and the sorting means sorting a large amount of acquired characterization data according to specific rules, where the specific rules include sorting according to defect type, defect size, or defect density, or sorting according to the surface region.

[0012] Based on further improvements to the above method, the feature database includes the following data: defect type, defect size, defect distribution, defect density, surface morphology, surface structure, surface chemical composition, electronic structure, and sample attribute data related to the laser finishing process.

[0013] Based on further improvements to the above method, the inversion deep learning model includes: an encoder-decoder module, an image processing module, a relationship processing module, an attention mechanism module, and an activation function module. Among them, the encoder-decoder module is used to segment and reconstruct the input image, and identify atomic-level features and defects; the image processing module is used to analyze the input image reconstructed by the encoder-decoder module, extract surface topography information, and perform various transformation and analysis processes to obtain denoised and enhanced image features and provide them to the relationship processing module; the relationship processing module is used to process the topological structure or spatial correlation of surface defects for the surface topography information, analyze the connections between surface atomic-level defects, so as to obtain a new feature representation integrating the relationships between images and provide it to the attention mechanism module; the attention mechanism module is used to enhance the perception ability of the inversion deep learning model for the wafer micro-nano surface structure, improve the accuracy of identification and localization of atomic-level defects, so as to obtain a feature subset of key attention and provide it to the activation function module; and the activation function module is used to enhance the network representation ability and adaptability, and at the same time enhance the identification and expression ability of the inversion deep learning model for atomic-level flaws and structural features.

[0014] Based on further improvements to the above method, the attention mechanism module includes a basic feature extraction unit, a structured attention unit, a channel attention unit, a redundant attention unit, and a hierarchical fusion unit. Among them, the basic feature extraction unit is used to extract multi-scale visual features from the wafer substrate image using a convolutional neural network and provide them to the structured attention unit and the channel attention unit. Among them, the multi-scale visual features include edges, textures, and shape features; the structured attention unit is used to adopt a pyramid pooling structure, connect feature maps of different scales, learn the weighted representation of features at all levels, focus on key structured detail features to generate attention weights for different regions and provide them to the redundant attention unit and the hierarchical fusion unit; the channel attention unit is used to adaptively re-weight the convolutional channels through the squeeze-excitation mechanism, enhance the ability to depict atomic-level structures to generate attention weights for different channels and provide them to the redundant attention unit and the hierarchical fusion unit; the redundant attention unit is used to identify irrelevant regions using attention scores, remove redundant interference terms in the intermediate layer features of the image to generate fused attention weights and provide them to the hierarchical fusion unit; and the hierarchical fusion unit fuses the attention features of different modules through an anchor box or a skip connection structure to form a feature representation that is more sensitive to atomic-level details.

[0015] Based on the further improvement of the above method, the warning information is that when the optimal prediction model identifies that a propagable destructive defect appears in the wafer substrate or the surface roughness exceeds the roughness threshold, the propagable destructive defect or the roughness that does not meet the requirements is used as warning information; the prediction information is that when the optimal prediction model identifies that no propagable destructive defect appears in the wafer substrate or the surface roughness does not exceed the roughness threshold, the defects that will appear in subsequent processing are predicted according to the current state of the substrate wafer and the type, size, and location distribution information of the defects are predicted as prediction information; and the prevention and control information is determined according to the prediction results of the optimal prediction model. The improvement measures recommended according to the surface morphology evolution law are used as prevention and control information.

[0016] Based on a further improvement of the above method, the laser industrial control system adjusts the atomic-level laser polishing parameters according to the early warning, prediction and prevention information, including: when the laser industrial control system receives the early warning information, the laser industrial control system reduces the laser power, increases the scanning line spacing, adjusts the spot shape, scanning rate, laser energy or focal position to weaken or enhance the intensity of the laser effect on the defective area; when the laser industrial control system receives the prediction information, the laser industrial control system specifically adjusts the laser parameters before laser polishing to avoid the critical area, or reduces the spot size and scanning overlap of the critical area, wherein the critical area is a critical area where faults may occur; when the laser industrial control system receives the prevention and control information, the laser industrial control system changes the scanning strategy, power distribution, or uses other wavelength lasers to rescan according to the suggestions provided by the optimal prediction model to prevent or eliminate the predicted defects.

[0017] On the other hand, an embodiment of the present invention provides an intelligent identification device for atomic-level laser processing of wafer substrates, including: an information acquisition module for collecting defect and morphological feature information generated during the interaction between laser polishing and the atomic layer on the surface of the wafer substrate; a feature database for storing the defect and morphological feature information; a model construction and training module for establishing an inversion deep learning model and using the defect and morphological feature information to perform distributed collaborative training and integrated regression testing on the inversion deep learning model to obtain an optimal prediction model; the optimal prediction model is used to identify propagable destructive defects and rough surface micromorphological structures that appear during the laser polishing of the wafer substrate, so as to obtain early warning information, prediction information or prevention information of the defects and morphology in the laser polishing process and feed it back to the laser industrial control system; and a laser industrial control system for adjusting the atomic-level laser polishing parameters according to the early warning information, prediction information or prevention information to form a closed-loop control, so as to accurately regulate and improve the quality of the laser processing process of the wafer substrate.

[0018] Based on further improvements to the above system, the information acquisition module includes: a transmission electron microscope, an atomic force microscope and a Raman spectrum, wherein the transmission electron microscope is used to collect morphology information, crystal structure information, defect and interface information and composition analysis information of the wafer substrate; the atomic force microscope is used to collect surface morphology information, surface mechanical properties and surface charge distribution information; and the Raman spectrum is used to collect molecular structure information, lattice vibration information and chemical composition analysis information of the wafer substrate.

[0019] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0020] (1) Compared with traditional mechanical grinding and chemical mechanical polishing, the present invention can monitor and predict processing defects in real time, avoiding the large-scale propagation of defects and improving the manufacturing yield rate. Secondly, by accurately quantitatively analyzing defects, the system can accurately process different defect types and degrees, avoiding the problems of over-grinding or polishing that may be caused by traditional methods, and improving processing efficiency. Most importantly, the intelligent laser finishing method can obtain wafer substrates with ultra-smooth surfaces, complete and defect-free surface lattices, and no surface or subsurface damage, meeting the needs of high-precision microelectronics and integrated circuit chip manufacturing.

[0021] (2) The present invention uses advanced atomic-level characterization technologies including transmission electron microscopy, atomic force microscopy and high-resolution Raman spectroscopy characterization technology to collect images and data of various defects and morphological characteristics generated during the laser finishing of wafer substrates, which can capture the microscopic details and morphological characteristics of the wafer surface and crystal structure. These atomic-level characterization technologies can observe and analyze the structure, morphology and chemical composition of materials at the nanoscale, and have the characteristics of high resolution and high sensitivity. The present invention uses transmission electron microscopy, atomic force microscopy and high-resolution Raman spectroscopy characterization technology combined with a self-programmed sorting by region and type to comprehensively collect defect and morphological characteristics generated during the interaction between laser finishing and the atomic layer on the surface of the wafer substrate. In contrast, existing technologies can often only obtain limited local information and cannot provide comprehensive atomic-level data. By combining data obtained by different atomic-level characterization technologies (such as image data, topological data and spectral data), multimodal data fusion can be achieved. This fusion can provide more comprehensive and diverse information, which helps to improve the robustness and generalization ability of deep learning models.

[0022] (3) This invention uses a self-developed inversion deep learning intelligent algorithm combined with an adaptive interactive model to select the optimal model through distributed collaborative training and integrated regression testing. This intelligent algorithm can more accurately predict design defects and morphological characteristics, and provide higher prediction accuracy and precision.

[0023] (4) The present invention combines an optimal model to identify the propagable destructive defects and rough surface microtopography structures that occur during the laser finishing of the wafer substrate, and reveals the evolution law of the combined products that appear during the laser finishing of the wafer substrate, as well as the non-thermal removal and planarization formation mechanism of the surface material. This enables the processing equipment to give early warnings, predictions, and prevention and control information about the defects and topography during the laser finishing process, helps avoid potential problems and losses, and improves production efficiency.

[0024] (5) The present invention feeds back the warning, prediction, and prevention and control information in step (3) to the laser industrial control system to achieve real-time closed-loop control. This means that the laser industrial control system can adjust the atomic-level laser finishing parameters in advance according to the prediction results to achieve the optimal atomic-level laser finishing method. This real-time feedback and adjustment can ensure the stability and consistency of the processing process, and improve the quality and reliability of the product.

[0025] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification, or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained through the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings are only for the purpose of showing specific embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference signs represent the same components.

[0027] Figure 1 is a flowchart of an intelligent recognition method for atomic-level laser processing of a wafer substrate according to an embodiment of the present invention;

[0028] Figure 2 is a specific flowchart of an intelligent recognition method for atomic-level laser processing of a wafer substrate according to an embodiment of the present invention;

[0029] Figure 3 is a structural diagram of an inversion deep learning model according to an embodiment of the present invention;

[0030] Figure 4 is a block diagram of an intelligent recognition device for atomic-level laser processing of a wafer substrate according to an embodiment of the present invention;

[0031] Figure 5 is a structural diagram of an encoder-decoder module according to an embodiment of the present invention;

[0032] Figure 6 is a structural diagram of an image processing module according to an embodiment of the present invention;

[0033] Figure 7 Structural diagram of a relationship processing module according to an embodiment of the present invention;

[0034] Figure 8 Structural diagram of an attention mechanism module according to an embodiment of the present invention;

[0035] Figure 9 Structural diagram of an activation function module according to an embodiment of the present invention. Detailed implementation manners

[0036] The preferred embodiments of the present invention will be specifically described below in conjunction with the accompanying drawings. The accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0037] The object of the present invention is to provide an intelligent recognition method for atomic-level laser processing of a wafer substrate, which replaces the traditional two processes of mechanical grinding and chemical mechanical polishing, can efficiently and accurately identify and predict various defects occurring in the process of atomic-level laser finishing of the wafer substrate, and perform accurate quantitative analysis on these defects, so as to obtain a wafer substrate with ultra-smooth surface, complete surface lattice without defects, and no surface or subsurface damage. This method uses deep learning technology to establish an intelligent database, quickly, accurately, and effectively identify the processing defects and morphological feature information in the process of laser finishing the wafer substrate, so as to realize the real-time monitoring and prediction of the atomic-level surface quality in the laser finishing process, and prevent and control by actively adjusting the equipment parameters, which can effectively improve the processing effect of atomic-level laser finishing.

[0038] As Figure 1 shown, a specific embodiment of the present invention discloses an intelligent recognition method for atomic-level laser processing of a wafer substrate, including: in step S101, collecting the defect and morphological feature information generated during the interaction between laser finishing and the atomic layer on the surface of the wafer substrate and storing it in the feature database; in step S102, establishing an inverse deep learning model and using the defect and morphological feature information to perform distributed collaborative training and integrated regression testing on the inverse deep learning model to obtain an optimal prediction model; in step S103, using the optimal prediction model to identify the propagable destructive defects and rough surface micro-morphological structures occurring in the process of laser finishing the wafer substrate, so as to obtain the early warning information, prediction information or prevention and control information of the defects and morphology in the laser finishing process and feedback it to the laser industrial control system; and in step S104, the laser industrial control system adjusts the atomic-level laser finishing parameters according to the early warning information, prediction information or prevention and control information to form a closed-loop control, so as to precisely control and improve the quality of the laser processing process of the wafer substrate.

[0039] Compared with traditional mechanical grinding and chemical mechanical polishing, the intelligent identification method for atomic-level laser processing of wafer substrates provided in this embodiment can monitor and predict processing defects in real time, avoid the large-scale propagation of defects, and improve the manufacturing yield rate. Secondly, by accurately quantitatively analyzing the defects, the system can accurately process different defect types and degrees, avoiding the problems of over-grinding or polishing that may be caused by traditional methods, and improving processing efficiency. Most importantly, the intelligent laser finishing method can obtain wafer substrates with ultra-smooth surfaces, complete and defect-free surface lattices, and no surface or subsurface damage, meeting the needs of high-precision microelectronics and integrated circuit chip manufacturing.

[0040] In the following, reference Figure 1 Each step of the intelligent identification method for atomic-level laser processing of wafer substrates according to an embodiment of the present invention is described in detail.

[0041] In step S101, information on defects and morphological features generated during the interaction between laser finishing and the atomic layer on the wafer substrate surface is collected and stored in a feature database. The feature database includes the following data: defect type, defect size, defect distribution, defect density, surface morphology, surface structure, surface chemical composition, electronic structure, and sample attribute data related to the laser finishing process.

[0042] For example, images directly collected by transmission electron microscopy, atomic force microscopy, and Raman spectroscopy serve as input images for the inversion deep learning model. Specifically, the acquisition of defect and morphological feature information generated during the interaction between laser finishing and the atomic layer on the wafer substrate surface includes: TEM acquisition of wafer substrate morphology information, crystal structure information, defect and interface information, and composition analysis information; AFM acquisition of surface morphology information, surface mechanical properties, and surface charge distribution information; and Raman spectroscopy acquisition of wafer substrate molecular structure information, lattice vibration information, and chemical composition analysis information.

[0043] Specifically, the collection of defect and morphology feature information generated during the interaction between laser finishing and the atomic layer on the wafer substrate surface includes: comprehensively collecting defect and morphology feature information generated during the interaction between laser finishing and the atomic layer on the wafer substrate surface by using transmission electron microscopy, atomic force microscopy, and Raman spectroscopy in combination with region-by-region and type-by-type sorting. Herein, region-by-region means obtaining representative data of the wafer substrate surface by selecting different regions or sample points to reveal the regional distribution of surface defects; type-by-type means distinguishing different types of defects and using different characterization methods to identify the morphology of the defects. Among them, different types of defects include point defects, line defects, and surface defects; and sorting means sorting a large amount of obtained characterization data according to specific rules. Herein, the specific rules include sorting according to defect type, defect size, or defect density, or sorting according to the surface region.

[0044] In step S102, an inverse deep learning model is established and the defect and morphology feature information is used to perform distributed collaborative training and integrated regression testing on the inverse deep learning model to obtain an optimal prediction model.

[0045] Reference Figure 3 , the inverse deep learning model includes: an encoder-decoder module, an image processing module, a relationship processing module, an attention mechanism module, and an activation function module. Among them, the encoder-decoder module is used to segment and reconstruct the input image, identify atomic-level features and defects. Specifically, the defect and morphology feature information generated during the interaction between laser finishing and the atomic layer on the wafer substrate surface is comprehensively collected by using transmission electron microscopy, atomic force microscopy, and Raman spectroscopy in combination with region-by-region and type-by-type sorting as the image information input to the inverse learning model; the image processing module is used to analyze the input image reconstructed by the encoder-decoder module and extract surface morphology information for various conversion and analysis processes to obtain denoised and enhanced image features and provide them to the relationship processing module. For example, the image processing module converts the reconstructed input image and the extracted surface morphology information into information such as the distribution and quantity of different types of defects; analyzes the defect type and severity, as well as the evolution law of the surface morphology, and determines whether the defect or roughness exceeds a predetermined threshold; and further obtains the following results: the digital features of the surface morphology and defect type / distribution of the wafer sample; the relationship processing module is used to process the surface morphology information, the topological structure or spatial correlation of surface defects, analyze the connection between surface atomic-level defects to obtain a new feature representation of the relationship between integrated images and provide it to the attention mechanism module; the attention mechanism module is used to enhance the perception ability of the inverse deep learning model for the wafer micro-nano surface structure, improve the accuracy of identifying and locating atomic-level defects to obtain a feature subset of key attention and provide it to the activation function module; and the activation function module is used to enhance the network representation ability and adaptability, and at the same time enhance the identification and expression ability of the inverse deep learning model for atomic-level defects and structural features.

[0046] Specifically, referring to Figure 8 , the attention mechanism module 800 includes a basic feature extraction unit 801, a structured attention unit 802, a channel attention unit 803, a redundant attention unit 804, and a hierarchical fusion unit 805. Among them, the basic feature extraction unit 801 is used to extract multi-scale visual features from the wafer substrate image using a convolutional neural network and provide them to the structured attention unit and the channel attention unit. The multi-scale visual features include edge, texture, and shape features. The structured attention unit 802 is used to adopt a pyramid pooling structure to connect feature maps of different scales, learn the weighted representation of features at all levels, focus on key structured detail features to generate attention weights for different regions, and provide them to the redundant attention unit and the hierarchical fusion unit. The channel attention unit 803 is used to adaptively re-weight the convolutional channels through the squeeze-excitation mechanism, enhance the ability to depict atomic-level structures to generate attention weights for different channels, and provide them to the redundant attention unit and the hierarchical fusion unit. The redundant attention unit 804 is used to identify irrelevant regions using attention scores, remove redundant interference terms in the intermediate layer features of the image to generate fused attention weights, and provide them to the hierarchical fusion unit. And the hierarchical fusion unit 805 fuses the attention features of different modules through an anchor box or a skip connection structure to form a feature representation that is more sensitive to atomic-level details.

[0047] In step S103, the optimal prediction model is used to identify the propagable destructive defects and rough surface micro-topography structures that occur during the laser finishing of the wafer substrate, so as to obtain early warning information, prediction information, or prevention information about the defects and topography during the laser finishing process and feedback them to the laser industrial control system.

[0048] Specifically, the early warning information is that when the optimal prediction model identifies that there are propagable destructive defects or the surface roughness exceeds the roughness threshold in the wafer substrate, the propagable destructive defects or the roughness that does not meet the requirements are used as early warning information. The prediction information is that when the optimal prediction model identifies that there are no propagable destructive defects or the surface roughness does not exceed the roughness threshold in the wafer substrate, the defects that will occur in the subsequent processing are predicted according to the current state of the substrate wafer, and the type, size, and position distribution information of the defects are predicted as prediction information. The prevention information is to determine the surface topography evolution law of the wafer substrate according to the prediction result of the optimal prediction model, and use the improvement measures recommended according to the surface topography evolution law as prevention information.

[0049] In step S104, the laser industrial control system adjusts the atomic-level laser finishing parameters according to the early warning information, prediction information, or prevention information to form a closed-loop control loop, so as to precisely regulate and improve the quality of the laser processing process of the wafer substrate.

[0050] Specifically, the laser industrial control system adjusts the atomic-level laser polishing parameters according to the early warning, prediction and prevention information, including: when the laser industrial control system receives early warning information, the laser industrial control system reduces the laser power, increases the scanning line spacing, adjusts the spot shape, scanning rate, laser energy or focal position to weaken or enhance the intensity of the laser effect on the defective area; when the laser industrial control system receives prediction information, the laser industrial control system specifically adjusts the laser parameters before laser polishing to avoid the critical area, or reduces the spot size and scanning overlap in the critical area, where the critical area is the critical area where faults may occur; when the laser industrial control system receives prevention and control information, the laser industrial control system changes the scanning strategy, power distribution, or uses other wavelength lasers to rescan according to the suggestions provided by the optimal prediction model to prevent or eliminate the predicted defects.

[0051] Another specific embodiment of the present invention, referring to Figure 4 , discloses an intelligent identification device for atomic-level laser processing of wafer substrates, including: an information acquisition module 401, used to collect defect and morphological feature information generated during the interaction between laser finishing and the atomic layer on the surface of the wafer substrate; a feature database 402, used to store defect and morphological feature information; a model construction and training module 403, used to establish an inversion deep learning model and use the defect and morphological feature information to perform distributed collaborative training and integrated regression testing on the inversion deep learning model to obtain an optimal prediction model; an optimal prediction model 404, used to identify propagable destructive defects and rough surface micromorphological structures that appear in the laser finishing process of the wafer substrate, so as to obtain early warning information, prediction information or prevention information of defects and morphology in the laser finishing process and feed it back to the laser industrial control system; and a laser industrial control system 405, used to adjust the atomic-level laser finishing parameters according to the early warning information, prediction information or prevention information to form a closed-loop control, so as to accurately regulate and improve the quality of the laser processing process of the wafer substrate.

[0052] The information acquisition module includes: transmission electron microscopy, atomic force microscopy and Raman spectroscopy. Among them, transmission electron microscopy is used to collect the morphology information, crystal structure information, defect and interface information and composition analysis information of the wafer substrate; atomic force microscopy is used to collect surface morphology information, surface mechanical properties and surface charge distribution information; and Raman spectroscopy is used to collect molecular structure information, lattice vibration information and chemical composition analysis information of the wafer substrate.

[0053] Hereinafter, the intelligent identification method for atomic-level laser processing of wafer substrates according to an embodiment of the present invention is described in detail by way of specific examples.

[0054] refer to [[ID=, an intelligent recognition method for atomic-level laser processing of a wafer substrate according to the present invention includes the following steps: Step 1, using transmission electron microscopy, atomic force microscopy, and high-resolution Raman spectroscopy characterization techniques in combination with a self-written sorting program for sub-region and sub-type sorting to establish a database of defect and morphological feature information generated during the interaction between laser finishing and the atomic layer on the surface of the wafer substrate; Step 2, using a self-written inverse deep learning intelligent algorithm in combination with an adaptive interaction model prediction design to perform distributed collaborative training and integrated regression testing on the defect and morphological feature information data in Step 1, and selecting the optimal model; Step 3, combining the optimal model in Step 2 to identify the propagable destructive defects and rough surface micro-morphological structures that occur during the laser finishing of the wafer substrate, revealing the evolution law of the joint products that occur during the laser finishing process of the wafer substrate, as well as the non-thermal removal and planarization formation mechanism of the surface material, and giving early warnings, predictions, and prevention and control information on defects and morphology during the laser finishing process in advance; Step 4, feeding back the warning, prediction, and prevention and control information in Step 3 to the laser industrial control system, adjusting the atomic-level laser finishing parameters in advance, forming a stable closed-loop control, and achieving the optimal atomic-level laser finishing method.

[0055] The self-written sorting program for sub-region and sub-type sorting described in Step 1 is used to organize, analyze, and mine the characterization data of transmission electron microscopy, atomic force microscopy, and high-resolution Raman spectroscopy.

[0056] The database of defect and morphological feature information described in Step 1 includes defect type, defect size, defect distribution, defect density, surface morphology, surface structure, surface chemical composition, electronic structure, and sample attribute data related to the laser finishing process.

[0057] In the present invention, by using transmission electron microscopy (TEM), atomic force microscopy (AFM), and high-resolution Raman spectroscopy characterization techniques, various information before and after laser finishing of the wafer substrate can be obtained.

[0058] 1. Transmission electron microscopy (TEM) can characterize:

[0059] Morphological information: TEM can provide high-resolution morphological images of the wafer substrate, showing the lattice structure, surface morphology, and particle distribution of the wafer substrate, etc.

[0060] Crystal structure information: Through the diffraction pattern of TEM, the crystal structure, lattice constant, and crystal orientation of the wafer substrate can be determined, etc.

[0061] Defect and interface information: TEM can detect defects, grain boundaries, and interfaces in the wafer substrate, providing information on defect type, density, and interface structure.

[0062] Component analysis: Through techniques such as energy-dispersive X-ray spectroscopy (EDS) or electron energy loss spectroscopy (EELS), the composition and elemental distribution of the wafer substrate can be analyzed.

[0063] 2. Atomic force microscopy (AFM) can characterize:

[0064] Surface topography information: AFM can obtain three-dimensional topography images of the wafer substrate surface, including atomic-level height and topological information.

[0065] Surface mechanical properties: AFM can measure the mechanical properties of the wafer substrate, such as hardness, elastic modulus, and adhesion force, etc.

[0066] Surface charge distribution: By introducing current or applying an external voltage to the AFM probe, the charge distribution on the wafer substrate surface can be measured.

[0067] 3. High-resolution Raman spectroscopy can characterize:

[0068] Molecular structure information: High-resolution Raman spectroscopy can provide vibration information of molecules in the wafer substrate, thereby determining the structure of molecules and the types of chemical bonds.

[0069] Lattice vibration information: By measuring the lattice vibration modes of the wafer substrate, the lattice dynamics properties and phonon spectra of the wafer substrate can be understood.

[0070] Chemical composition analysis: Raman spectroscopy can be used to analyze the chemical composition, phase transformation, and chemical reactions, etc. of the wafer substrate.

[0071] Furthermore, the sub-regions in the self-written program for sorting by region and type: When characterizing the wafer substrate surface, different regions or sample points are selected for characterization. This is to obtain representative data of the surface and reveal the regional distribution of surface defects; Sorting by type: refers to distinguishing different types of defects, such as point defects, line defects, surface defects, etc., and conducting classification statistics and analysis. By utilizing the advantages of different characterization techniques, the morphology of defects can be identified; Sorting: refers to sorting a large amount of obtained characterization data according to specific rules. This may include sorting according to defect type, size, density, etc. indicators, or sorting according to surface regions, etc.; Self-written program: that is, the algorithm or program code written by researchers themselves for data processing and analysis. Through this program, data sorting, analysis, and mining are realized.

[0072] Specifically, the defect and morphology feature information includes the following data information:

[0073] (1) Defect type: Classification of different morphologies such as point defects, line defects, and surface defects, or counting the number of defects, distinguishing more subtypes of defects such as single / clustered defects, and lattice defect structures such as vacancies and substitution atoms.

[0074] (2) Defect size: Geometric parameters of the defect such as length, width, depth, and height;

[0075] (3) Defect shape: circular, elliptical, long strip, etc., as well as statistical parameters such as maximum, minimum, and average values;

[0076] (4) Defect distribution: coordinate position distribution information of defects on the wafer surface, as well as distribution functions and parameters of Gaussian and Poisson distributions, and clustered distributions at different distance scales and spatially related scales;

[0077] (5) Defect density: the number of various defects per unit area; surface morphology: information such as surface roughness parameters, peak-to-valley height, morphology curve, etc., morphology such as isotropy / anisotropy of roughness parameters in different directions, and two-dimensional autocorrelation function reflecting surface correlation;

[0078] (6) Surface structure: surface atomic arrangement determined by transmission electron microscopy, local structure around defects, dislocation vector structure of linear defects such as dislocation and misfit dislocation;

[0079] (7) Surface chemical composition: High-resolution Raman spectroscopy determines the surface chemical bond structure, composition change information, component and chemical valence state XPS peaks;

[0080] (8) Electronic structure: surface electronic state density, band structure information, state density and localized state electronic structure caused by defects measured by atomic force microscopy;

[0081] (9) Other attributes: It may also include some sample attribute data related to the processing technology, including laser wavelength, laser frequency, laser pulse width, laser scanning speed, laser power, defocus, processing size, etc.

[0082] Establish a database: Based on the above processing, a database of different defect types containing rich characterization information is formed, laying the foundation for further analysis and modeling.

[0083] Furthermore, the self-compiled inversion deep learning intelligent algorithm described in step 2 includes a deep neural network with an encoding-decoding structure, an image processing network based on a convolutional neural network, an attenion mechanism external memory module, and an autoencoder regularization module, thereby forming a mathematical model of a high-order self-compiled inversion deep learning intelligent algorithm, and performing distributed collaborative training and integrated regression testing on laser finishing data without precise model information to ensure the stability and convergence of the model.

[0084] Furthermore, the distributed collaborative training and integrated regression testing described in Step 2 are mainly used to accelerate the model training process, improve the model generalization performance, conduct integrated regression prediction, evaluate the model quality, and reduce the overfitting risk, so as to provide a basis for result decision-making and enhance the performance and credibility of a single model.

[0085] Furthermore, the warning, prediction, and prevention information described in Step 2 refers to sending warning signals to the operator in a timely manner during the laser surface finishing process, predicting in real time and giving the distribution and severity level of surface defects and topography types for the next processing step, so as to prevent or mitigate the expected defects and problems.

[0086] The self-developed inversion deep learning intelligent algorithm in the present invention mainly includes the following components:

[0087] A deep neural network with an encoder-decoder structure, which is used to extract features and map the potential mapping relationship between the input and output, similar to the Seq2Seq or autoencoder structure.

[0088] An image processing network based on a convolutional neural network, which is used to analyze image input and extract spatial information such as surface topography, and can be UNet, Convolutional VAE, etc.

[0089] A recurrent or graph neural network, which is used to process the topological structure or spatial correlation of surface defects, as well as the relationship between multi-modal heterogeneous data.

[0090] The Attenion mechanism or external memory module, which is used to capture long-range dependencies and helps to model the interaction between multiple regions on the surface.

[0091] The activation function is learnable, such as PReLU, learnable ELU, etc., which enhances the network representation ability and the ability to adapt to specific problems.

[0092] The interpretability method of convolutional neural network, such as Grad-CAM and other techniques, is used to explain the surface regions concerned by the network and understand the data features.

[0093] The autoencoder regularization method, such as sparse, denoising, factorization, etc., enhances the data-driven feature learning ability.

[0094] The integration method integrates multiple networks to improve generalization and enhance the interpretability of results.

[0095] In the present invention, the adaptive interactive model prediction design improves and optimizes the model design by interacting with experimental or simulation data. This process generally includes the following:

[0096] Data collection and preparation: Collect and organize a dataset of defect and morphological feature information, which may include experimental observation data, computational simulation data, literature-reported data, process parameter data, equipment status data, wafer description data, process inspection data, and process log data; as well as historical process data, external environment data, raw material data, operation and maintenance and service data, model training and evaluation data, model optimization and iteration data, etc.

[0097] Deep learning model construction: Use a self-developed inversion deep learning intelligent algorithm to construct a model that can receive input data and learn the complex relationships between the data. This may involve deep learning architectures such as Convolutional Neural Networks (CNN) or Recurrent Neural Networks (RNN).

[0098] Reference ​ , the inversion deep learning model in the intelligent recognition method for atomic-level laser processing of wafer substrates includes 5 modules, namely: encoder-decoder module, image processing module, relationship processing module, attention mechanism module, activation function module. The connections and functions of these 5 modules are as follows: The encoder-decoder module is responsible for extracting semantic features of the wafer image and reconstructing context space information, which is the representation learning of the input image. Specifically, input the original image data into this encoder-decoder module, and the output of this encoder-decoder module through learning is the image feature representation. The image processing module performs multi-scale image transformation on the features extracted by the encoder, enhances the perception of key microstructures, and provides an image prior for feature expression. Specifically, input the image feature representation into this image processing module, and through the processing of this image processing module, generate the denoised and enhanced image features as the output of this image processing module. The relationship processing module analyzes the topological connections between entities in the denoised and enhanced image features, finds the association patterns of atomic defects, and realizes the structured modeling of features, so as to output a new feature representation integrating the relationships between images and provide it to the attention mechanism module. The attention mechanism module plays a guiding role in feature extraction and association analysis by suppressing indirectly related regions and weighting relevant details to generate a feature subset of key attention and provide it to the activation function module. The activation function module mainly acts in various neural networks, strengthens the non-linearity of feature expression, and enhances the ability of the model to fit complex mappings to generate the final feature expression obtained after transformation by the activation function.

[0099] These 5 modules have their own focuses and at the same time cooperate with each other to support more accurate and robust atomic-level defect characterization, and finally realize the automated inspection and quality control of wafer substrates.

[0100] (1) Reference ​, the encoder-decoder module 500 can be used to extract feature information and map the mapping relationship between input data and output data through the combination of its various constituent units. At the same time, it can perform fine image segmentation and reconstruction on the micro-nano surface, effectively identifying atomic-level features and defects. Its structure includes: Encoder unit 501: Usually using a recurrent neural network (RNN), convolutional neural network (CNN), or attention mechanism (Attention), it encodes the input sequence into a fixed-length vector representation to capture its semantic information; Decoder unit 503: Using an RNN or a network with an attention mechanism, it decodes the vector representation output by the encoder into the desired output sequence; It also includes a downsampling path unit, a skip connection unit, and an attention gating mechanism unit. The working process of the encoder-decoder is as follows: The encoder reads the input sequence and outputs the context vector c. The decoder generates the output of the first time step based on the context vector c, uses the output of the first time step as the input of the decoder at the second time step, predicts the output of the second time step, and loops in turn until a complete output sequence is generated; It is used to extract feature information and map the mapping relationship between input data and output data.

[0101] Downsampling path unit 502: Gradually downsamples the encoded features through a max pooling layer to expand the receptive field and obtain a more abstract global representation.

[0102] Skip connection unit 505: Adopts a skip connection of the UNet-like structure, directly transmits shallow features to the deep layer, bypasses the non-linear transformation, and retains more original signals.

[0103] Attention gating mechanism unit 504: Uses a gating unit to selectively strengthen the semantic information flow and specifically focuses on key atomic structures.

[0104] Specifically, the skip connection unit and the attention gating mechanism unit are used to cooperate with the encoder unit, decoder unit, and downsampling path unit to expand and enhance these three units. The skip connection unit accelerates the encoding-decoding process, and the attention gating mechanism aggregates the most important part of the information, enabling the entire encoder-decoder module to complete feature extraction and input-output mapping more efficiently and accurately. The relationship with other framework units is as follows.

[0105] (2) Refer to ​ , the image processing module 600 can be used to analyze the input image information and extract surface topography information for various conversion and analysis processes through the combination of its various constituent units. It enhances the perception of the atomic-scale surface shape, boundary, and defects, improves the representation ability of atomic-scale fine features, and is beneficial to subsequent recognition and analysis. Its constituent units mainly include: low-level feature extraction unit, multi-scale fusion unit, shape semantic unit, sparse representation unit, prior constraint unit, and image segmentation network unit.

[0106] Low-level feature extraction unit 601: Extracts directional features using a rod waveguide filter bank to enhance the description of edges and textures, so as to generate image features such as edges and textures. Specifically, the rod waveguide filter bank used in the low-level feature extraction unit directly processes the input original image to extract directional features, where the original image is the original data obtained from an image acquisition device (transmission electron microscope, atomic force microscope, Raman spectrometer, etc.), that is, the image before encoding into the encoder-decoder module. The directional features are output by the image processing module and become the input of subsequent modules such as the encoder-decoder. Specifically, for the encoder-decoder module, the directional features (low-level directional features) extracted by the low-level feature extraction unit can provide edge features in different directions to help the encoder better understand the line direction and shape contour information in the image. For the relationship processing module, texture features in different directions can help analyze the spatial relationship between pixels or regions in the image, such as determining that regions with the same texture direction belong to the same object. For the attention mechanism module, the low-level directional features can provide low-level line direction cues to guide the attention mechanism to focus on regions in a specific direction. For the activation function module, the response of the low-level directional features can help select activation functions that are sensitive to lines and edges, such as ReLu (a machine learning activation function). Therefore, the low-level directional features can provide low-level clues such as shape, edge, and texture for each module, helping to better analyze and understand the local structure of the image, thereby improving the performance of the entire model.

[0107] Multi-scale fusion unit 602: Adopts a scale space pyramid structure to connect feature maps of different scales, fuses multi-scale information to generate image pyramid representations of different scales, and provides them to the shape semantic unit.

[0108] Shape semantic unit 603: Constructs a shape bag of words, learns the word embedding representation of predefined primitives, and describes the regional shape features. The input of the shape semantic unit includes image features such as edges and textures output by the low-level feature extraction unit, and image pyramid representations of different scales output by the multi-scale fusion unit (these image features reflect the regional shape information); the output of the shape semantic unit is the shape bag of words embedding representation learned by the shape semantic unit, which describes the shape features of the local area of the image (these shape features are important information of the microscopic surface topography). The output features will be used as the result of the image processing module and provided to subsequent modules such as the encoder-decoder module, as well as other modules such as the attention mechanism module and the relationship processing module to assist in further image understanding and analysis processing.

[0109] Sparse Representation Unit 604: Adopting dictionary learning and sparse coding algorithms, it efficiently represents key atomic-level structures. The input of the sparse representation unit mainly comes from the outputs of multiple previous component units in the image processing module, including the image features output by units such as low-level feature extraction, multi-scale fusion, and shape semantics. These features reflect atomic-level structure and surface topography information.

[0110] The output of the sparse representation unit is an efficient representation of the key atomic-level structure obtained through dictionary learning and sparse coding, that is, the learned sparse dictionary that can represent atomic-level key information and the corresponding coding. The output of this sparse representation unit will be provided to subsequent encoder-decoder modules, as well as relationship processing modules, etc., for feature representation and relationship learning, which is beneficial to the identification of atomic-level defects and structures.

[0111] Prior Constraint Unit 605: Combining the physical model of crystal growth, it adds prior knowledge of atomic arrangement and kinematic constraints to regularize feature expressions. The input of this prior constraint unit is the surface topography image features output by the previous units in the image processing module, the prior knowledge of atomic arrangement provided by the physical model of crystal growth, and kinematic constraints.

[0112] The output of this prior constraint unit is a regularized feature expression obtained by combining prior knowledge, which reflects the constrained surface topography information and is an optimization of the original image features. This output will be provided to the encoder-decoder module, as well as subsequent relationship processing modules, etc. The enhanced surface topography features will help improve the accuracy of atomic-level defect and structure characterization.

[0113] (3) Reference ​ , through the combination of each component unit, the relationship processing module 700 can be used to process the topological structure or spatial correlation of surface defects, analyze the connections between surface atomic-level defects, so as to accurately identify the relationships between multi-modal heterogeneous data. Its structure mainly includes: entity recognition unit, relationship extraction unit, knowledge graph integration unit, joint reasoning unit, and interpretation and decision-making unit.

[0114] Entity Recognition Unit 701: Using a convolutional network to recognize image regions and locate various defect entities in the form of points, lines, and planes. The input of the entity recognition unit mainly comes from the output results of the image processing module, including image feature representations processed by low-level feature extraction, shape semantics learning, etc. These features contain information about surface topography, boundaries, and potential defects. Based on these input features, the entity recognition unit further analyzes and recognizes using a convolutional network to locate and label various defect entities in the form of points, lines, and planes. The input of the entity recognition unit is the low-level features of the image, and the output is information such as the positions and categories of the detected surface defect entities.

[0115] Relationship extraction module unit 702: Based on the instance segmentation results, extract spatial relationships such as topological connections and relative directional positions between entities. The input of the relationship extraction unit is the information of each defective entity output by the entity recognition unit, and the output is the spatial / topological relationship between different defective entities. For example, there is an inclusion relationship between A and B, and there is an adjacent relationship between C and D, etc.

[0116] Knowledge graph integration unit 703: Integrate the extracted entities and entity relationships into a predefined crystal structure knowledge graph. The input of the knowledge graph integration unit is the relationship between defective entities output by the relationship extraction unit, as well as the externally injected surface defect domain knowledge graph, and the output is an updated knowledge graph that incorporates new relationships. The predefined crystal structure knowledge graph mainly includes: a. Crystal type graph: covering the structural knowledge of various common crystal materials, such as diamond crystal, SiC crystal, Si crystal, gallium arsenide, gallium nitride, zinc blende structure, etc. b. Lattice arrangement graph: describing the spatial arrangement of atoms and molecules in different crystals, such as face-centered cubic lattice, body-centered cubic lattice, hexagonal close-packed lattice, etc. c. Crystal defect template: the structural template of typical crystal defects, such as point defects, line defects, plane defects, etc. d. Crystal surface reconstruction knowledge: reflecting the relevant constraint rules of surface structure reconstruction, embodying the expected surface shape, mainly including the principle of minimum surface energy, diffusion-limited model, bond optimization mechanism, dangling bond elimination principle, vacancy stability model, surface reset template, etc. e. Material property relationship graph: a relationship graph that associates material types with properties (optical properties, conductivity, etc.), mainly including: conductivity-doping relationship graph, energy gap-temperature relationship graph, refractive index-wavelength relationship graph, dielectric constant-frequency relationship graph, magnetic permeability-magnetic field relationship graph, loss-stress relationship graph, etc.

[0117] Joint reasoning module unit 704: Based on the knowledge graph, associate known atomic defect patterns and conduct relationship reasoning and analysis. The input of the joint reasoning unit is the updated knowledge graph, and the output is new relationships or conclusions obtained through joint reasoning using the knowledge graph. For example, A contains B and B is adjacent to C, so it can be inferred that A has a spatial relationship with C.

[0118] Explanation and decision-making unit 705: According to the results of the reasoning network, generate a text description using a rule template to explain the final atomic-level defect recognition decision-making process. The input of the explanation and decision-making unit is the relationship conclusion of the joint reasoning, and the output is an explanation of the final relationship reasoning decision-making process in natural language for easy understanding.

[0119] (4) Reference ​, the attention mechanism module 800, with a module design incorporating a rich attention mechanism, can enhance the model's perception ability of the wafer micro-nano surface structure, improve the accuracy of atomic-level defect recognition and localization, and is used to capture long-range dependencies to help model the interactions between multiple regions of the surface. Specifically, the attention mechanism module can enhance the model's perception ability of the wafer micro-nano surface structure: the attention mechanism can help the model learn to focus on important regions in the image, conduct targeted analysis of the surface micro-nano structure, and improve the ability to identify key details. Improve the accuracy of atomic-level defect recognition and localization: the attention mechanism can guide the model to focus on areas where defects may exist, reduce interference from irrelevant content, and thus more accurately identify the defect locations. Capture long-range dependencies: through attention calculations for different positions, the mutual influence between two distant regions can be modeled, capturing more complex dependencies. Model the interactions between multiple regions of the surface: different attention heads can simultaneously focus on different regions of interest, model the associations between regions, and achieve global modeling of surface information. Improve the interpretability and explainability of the model: the attention weight distribution can provide an explanation for the model, indicating the regions the model focuses on and increasing the explainability of the results. The attention mechanism module learns the attention distribution features of the image from different perspectives through different attention units and gradually fuses them to obtain a comprehensive multi-granularity attention representation for final image analysis. Its structure mainly includes: a basic feature extraction unit, a structured attention unit, a channel attention unit, a redundant attention unit, and a hierarchical fusion unit.

[0120] The basic feature extraction unit 801: uses a convolutional neural network to extract multi-scale visual features from the wafer substrate image, including information such as edges, textures, and shapes. The input of the basic feature extraction unit is the original image, and the output is the basic visual features extracted through operations such as convolution.

[0121] The structured attention unit 802: adopts a pyramid pooling structure to connect feature maps of different scales, learns the weighted representations of features at all levels, and pays more attention to key structured detail features. The input of the structured attention unit is the basic features, and the output is the attention weights for different regions obtained through the structured attention mechanism.

[0122] The channel attention unit 803: through the squeeze-excitation mechanism, adaptively re-weights the convolutional channels to enhance the ability to depict atomic-level structures. The input of the channel attention unit is also the basic features, and the output is the attention weights for different channels obtained through the channel attention mechanism.

[0123] Redundant Attention Unit 804: Identify irrelevant regions using attention scores and remove redundant interference terms in the intermediate-layer features of the image. The input of the redundant attention unit is the above two attention features, and the output is the fused attention weights obtained by calculating the redundancy between the attention features.

[0124] Hierarchical Fusion Unit 805: Fuse the attention features of different modules together through anchor boxes or skip connection structures to form a feature representation that is more sensitive to atomic-level details. The input of the hierarchical fusion unit is the above attention features, and the output is the final multi-level attention features obtained through hierarchical fusion.

[0125] (5) Reference ​ , Activation Function Module 900, which is used to enhance the network representation ability and adaptability. At the same time, the combination of each component unit can enhance the model's recognition and expression ability for atomic-level defects and structural features. Its structure mainly includes: Squeeze unit, Excitation unit, Sigmoid gating unit, Swish activation function unit, Mish activation function unit, Hard Swish activation function unit.

[0126] Squeeze Unit 901: Compress convolutional features into vectors using global average pooling to emphasize global information. The input of the Squeeze unit is the atomic-level defect multi-scale feature map output from the image processing module, and the output is the compressed atomic-level defect feature vector.

[0127] Excitation Unit 902: Implement an adaptive channel adjustment mechanism using a fully connected network to model the channel relationship of the input features and enhance important feature channels. The input of the Excitation unit is the atomic-level defect feature vector output by the Squeeze unit, and the output is the learned importance weights for different atomic-level defect features.

[0128] Sigmoid Gating Unit 903: Achieve flexible regulation of the output through the Sigmoid function to prevent gradient disappearance. The input of the Sigmoid gating unit is the original atomic-level defect multi-scale feature map and the feature weights learned by the Excitation unit, and the output is the new feature map after feature gating adjustment.

[0129] Swish Activation Function Unit 904: Swish = x * sigmoid(x). The Swish function can smoothly adjust the activation degree of neurons and enhance non-linearity. Swish Activation Unit: The input is the feature map adjusted by the Sigmoid gating unit, and the output is the enhanced feature representation obtained through transformation by the Swish activation function.

[0130] Mish Activation Function Unit 905: Mish = x * tanh(softplus(x)). The Mish function combines the advantages of Swish and ReLU at the same time.

[0131] Hard Swish Activation Function Unit 906: Hard Swish is a piecewise approximation of Swish, which is more computationally efficient.

[0132] These function units are executed in series to gradually enhance the representation of atomic-level defect features and the model's adaptability to them.

[0133] Distributed Collaborative Training: Distribute data and models to multiple computing nodes or devices, and accelerate the training process and improve the model's performance through parallel computing and communication collaboration. This distributed training can be implemented using clusters, cloud computing, or distributed computing frameworks. The data refers to the defect and morphology feature information generated during the interaction between laser finishing and the atomic layer of the wafer substrate surface stored in the feature database, and the model refers to the established inversion deep learning model.

[0134] Model Optimization and Improvement: Iteratively optimize and improve the model through interaction with experimental or simulation data. This may include techniques such as model parameter adjustment, feature engineering, data augmentation, etc., to improve the model's accuracy and generalization ability.

[0135] Integrated Regression Testing: Select the optimal model and conduct regression testing to evaluate the model's prediction performance on unseen data. This can help verify the reliability and applicability of the model and provide guidance for further design and prediction. Select the established inversion deep learning model for regression testing to evaluate the prediction ability of the prediction model for real-time data, and then form the optimal prediction model; this regression testing refers to using test data to evaluate the established deep learning prediction model to see how well it predicts actual data. By verifying the model's prediction performance in an independent test set in this way, the prediction ability of the deep learning model can be objectively evaluated, and the best model can be selected. Specifically, it may include the following steps: Collect actual wafer substrate sample data as the test set, input the test set sample data into the already trained deep learning prediction model, obtain the model's prediction output for these samples, compare the model's prediction output with the actual labels in the test set, and calculate some evaluation metrics such as MSE, R 2 etc. These metrics can evaluate the accuracy of the model's prediction. If the evaluation metrics perform well, it means that the deep learning model has strong prediction ability for actual data and can be used as the final prediction model. If the metrics are poor, it is necessary to return to the above steps to further optimize the model.

[0136] In the present invention, distributed collaborative training and integrated regression testing can have the following functions:

[0137] Accelerate the model training process. Through distributed computing, parallel training on multiple devices can greatly shorten the training time on a single device and improve the efficiency of research and optimization.

[0138] Improve the model generalization performance. Distributed training can learn features from more diverse examples, and different device parameter settings can form a model ensemble to enhance the generalization performance.

[0139] Conduct integrated regression prediction. Single model prediction may have problems such as bias or excessive variance. Model integration can integrate multiple models to smooth and reduce the variance of regression prediction.

[0140] Evaluate the quality of models. By comparing the regression prediction effects of different models on the test set, the best single model in the integrated model or the final optimal model of the entire integration can be selected.

[0141] Reduce the risk of overfitting. Distributed collaborative training introduces a certain degree of randomness to avoid overfitting of a single model to a specific training sample set.

[0142] Provide a basis for result decision-making. Regression testing can evaluate the prediction quality of different models and provide support for decision-makers to select a reliable final model.

[0143] Generally, utilize distributed and integrated technologies to enhance the performance and credibility of a single model.

[0144] In the present invention, the meanings of early warning, prediction, and prevention and control information in this scenario are as follows:

[0145] Early warning information: During the laser surface finishing process, when the model detects certain types of high-risk defects (such as cracks) or surface roughness exceeding the threshold, it can promptly send a warning signal to the operator.

[0146] Prediction information: The optimal model can real-time predict and give the distribution and severity level of the surface defect types for the next processing, as well as quantitative indicators of surface topography parameters such as roughness, Hole density, etc.

[0147] Prevention and control information: Based on a deep understanding of the evolution law of the surface topography, the model can recommend corresponding improvement measures, such as adjusting the laser wavelength, laser frequency, laser pulse width, laser scanning speed, laser power, defocus amount, processing size, etc., or switching the beam mode, etc., to prevent or mitigate the expected defects and problems.

[0148] In short, the intelligent model should make comprehensive use of computing power, physical knowledge and data-driven to achieve real-time monitoring and prediction of defects and surface morphology during the processing, risk assessment and device optimization decision-making, and play the guiding and auxiliary role of the automation system to improve the stability of the processing process and product quality.

[0149] The laser industrial control system adjusts atomic-level laser finishing parameters according to early warning, prediction and prevention information to form closed-loop control.

[0150] Early warning, forecast and prevention information specifically refers to:

[0151] Early warning information: When the optimal prediction model detects the presence of propagable destructive defects in the wafer substrate or the surface roughness exceeds the threshold, it sends a warning signal to the laser industrial control system, allowing the industrial control system to be aware of possible problems in advance;

[0152] Prediction information: This refers to the optimal prediction model predicting the type, size, and location distribution of defects that may occur in subsequent processing based on the current wafer status. This allows the industrial control system to predict possible failure modes and provide the optimal laser finishing parameters in a timely manner;

[0153] Prevention and control information: When receiving the above-mentioned warning or prediction information, the optimal prediction model will propose corresponding solutions or laser finishing parameter adjustment suggestions based on the experience knowledge base, intelligent algorithms, etc., and send them to the laser industrial control system. This is to prevent or alleviate the predicted problems.

[0154] The method for adjusting the atomic-level laser light-smoothing parameters can be as follows:

[0155] Warning information: When the monitoring system detects defects or abnormal surface roughness, it will issue a warning signal. At this time, the laser industrial control system can automatically or manually reduce the laser power, increase the scanning line spacing, adjust the spot shape, scanning rate, laser energy or focus position to weaken or enhance the laser effect on the area;

[0156] Prediction information: When the prediction model provides information such as the possible failure mode and location, the industrial control system can adjust the laser parameters before processing, such as avoiding critical areas where failure may occur or reducing the spot size and scanning overlap in that area.

[0157] Prevention and control information: The industrial control system will prevent or eliminate predicted defects and problems based on the suggestions provided by the prediction model, such as changing the scanning strategy, power distribution, and rescanning with other wavelength lasers.

[0158] The entire process of cooperation feedback and control can be continuously optimized and improved through machine learning algorithms, that is, online process control, or re-collecting defect and morphology feature information to repeat the entire model optimization process, thereby achieving precise regulation and quality improvement of laser processing.

[0159] Those skilled in the art can understand that all or part of the processes for implementing the methods of the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a disk, an optical disc, a read-only memory, or a random access memory, etc.

Claims

1. An intelligent recognition method for atomic-level laser processing of a wafer substrate, characterized in that, Including: Collecting defect and morphology feature information generated during the interaction between laser finishing and the atomic layer of the wafer substrate surface and storing it in the feature database; Establishing an inverse deep learning model and using the defect and morphology feature information to perform distributed collaborative training and integrated regression testing on the inverse deep learning model to obtain an optimal prediction model; Using the optimal prediction model to identify the transmissible destructive defects and rough surface micro-morphology structures that occur during the laser finishing of the wafer substrate to obtain early warning information, prediction information or prevention and control information about defects and morphology during the laser finishing process and feedback it to the laser industrial control system; And The laser industrial control system adjusts the atomic-level laser finishing parameters according to the early warning information, prediction information or prevention and control information to form a closed-loop control, so as to precisely regulate and improve the quality of the laser processing technology of the wafer substrate.

2. The intelligent identification method for atomic-level laser processing wafer substrate according to claim 1, characterized in that: Collecting defect and morphology feature information generated during the interaction between laser finishing and the atomic layer of the wafer substrate surface includes: Collecting the morphology information, crystal structure information, defect and interface information, and composition analysis information of the wafer substrate through a transmission electron microscope; Collecting surface morphology information, surface mechanical properties and surface charge distribution information through an atomic force microscope; and Collecting the molecular structure information, lattice vibration information and chemical composition analysis information of the wafer substrate through Raman spectroscopy.

3. The intelligent recognition method for atomic-level laser processing of a wafer substrate according to claim 2, characterized in that, Collecting defect and morphology feature information generated during the interaction between laser finishing and the atomic layer of the wafer substrate surface includes: comprehensively collecting defect and morphology feature information generated during the interaction between laser finishing and the atomic layer of the wafer substrate surface by combining the transmission electron microscope, the atomic force microscope and the Raman spectroscopy with regional and type-based sorting. Among them, The regional sorting means obtaining representative data on the wafer substrate surface by selecting different regions or sample points to reveal the regional distribution of surface defects; The type-based sorting means distinguishing different types of defects and using different characterization methods to identify the morphology of the defects. Among them, different types of defects include point defects, line defects and surface defects; and The sorting means sorting a large amount of obtained characterization data according to specific rules. Among them, the specific rules include sorting according to defect type, defect size or defect density, or sorting according to the surface area.

4. The intelligent identification method for atomic-level laser processing wafer substrate according to claim 1, characterized in that: The feature database includes the following data: defect type, defect size, defect distribution, defect density, surface morphology, surface structure, surface chemical composition, electronic structure and sample attribute data related to the laser finishing process.

5. The intelligent identification method for atomic-level laser processing wafer substrate according to claim 1, characterized in that: The inverse deep learning model includes: an encoder-decoder module, an image processing module, a relationship processing module, an attention mechanism module and an activation function module. Among them, The encoder-decoder module is used to segment and reconstruct the input image and identify atomic-level features and defects; The image processing module is used to analyze the input image reconstructed by the encoder-decoder module and extract surface morphology information for various conversion and analysis processes to obtain denoised and enhanced image features and provide them to the relationship processing module; The relationship processing module is used to process the topological structure or spatial correlation of surface defects from the surface topography information, analyze the connections between surface atomic-level defects, so as to obtain a new feature representation of the relationship between integrated images and provide it to the attention mechanism module; The attention mechanism module is used to enhance the perception ability of the inverse deep learning model for the wafer micro-nano surface structure, improve the accuracy of identification and localization of atomic-level defects, so as to obtain a feature subset of key attention and provide it to the activation function module; and The activation function module is used to enhance the network representation ability and adaptability, and at the same time enhance the identification and expression ability of the inverse deep learning model for atomic-level defects and structural features.

6. The intelligent recognition method for atomic-level laser processing of a wafer substrate according to claim 5, characterized in that, The attention mechanism module includes a basic feature extraction unit, a structured attention unit, a channel attention unit, a redundant attention unit and a hierarchical fusion unit. Among them, The basic feature extraction unit is used to extract multi-scale visual features from the wafer substrate image using a convolutional neural network and provide them to the structured attention unit and the channel attention unit. Among them, the multi-scale visual features include edges, textures, and shape features; The structured attention unit is used to adopt a pyramid pooling structure to connect feature maps of different scales, learn the weighted representation of features at all levels, focus on key structured detail features to generate attention weights for different regions and provide them to the redundant attention unit and the hierarchical fusion unit; The channel attention unit is used to adaptively re-weight the convolutional channels through the squeeze-excitation mechanism, enhance the ability to depict atomic-level structures to generate attention weights for different channels and provide them to the redundant attention unit and the hierarchical fusion unit; The redundant attention unit is used to identify irrelevant regions using attention scores, remove redundant interference terms in the intermediate layer features of the image to generate fused attention weights and provide them to the hierarchical fusion unit; and The hierarchical fusion unit fuses the attention features of different modules through an anchor box or a skip connection structure to form a feature representation that is more sensitive to atomic-level details.

7. The intelligent identification method for atomic-level laser processing of a wafer substrate according to claim 1, wherein The warning information is that when the optimal prediction model identifies that a propagable destructive defect or surface roughness exceeding the roughness threshold appears in the wafer substrate, the propagable destructive defect or non-compliant roughness is used as the warning information; The prediction information is that when the optimal prediction model identifies that no propagable destructive defect appears in the wafer substrate or the surface roughness does not exceed the roughness threshold, the defects that will appear in the subsequent processing are predicted according to the current state of the substrate wafer, and the type, size, and position distribution information of the defects are predicted as the prediction information; And The prevention and control information is to determine the surface topography evolution law of the wafer substrate according to the prediction result of the optimal prediction model, and use the improvement measures recommended according to the surface topography evolution law as the prevention and control information.

8. The intelligent recognition method for atomic-level laser processing of a wafer substrate according to claim 7, characterized in that The laser industrial control system adjusts the atomic-level laser finishing parameters according to the early warning, prediction and prevention information, including: When the laser industrial control system receives the warning information, the laser industrial control system reduces the laser power, increases the scanning line spacing, and adjusts the spot shape, scanning rate, laser energy or focus position to weaken or enhance the intensity of the laser effect on the defect area; When the laser industrial control system receives the prediction information, the laser industrial control system specifically adjusts the laser parameters before laser finishing to avoid the critical area, or reduces the spot size and scanning overlap of the critical area, wherein the critical area is the area where failure occurs in the subsequent processing process; When the laser industrial control system receives the prevention and control information, the laser industrial control system changes the scanning strategy, power distribution, scanning speed, spot size, spot position, or uses other wavelength lasers to rescan based on the suggestions provided by the optimal prediction model to prevent or eliminate the predicted defects.

9. An intelligent recognition device for atomic-level laser processing of a wafer substrate, characterized in that, include: An information acquisition module is used to collect information on defects and morphological features generated during the interaction between laser finishing and the atomic layer on the wafer substrate surface; A feature database, used to store the defect and morphological feature information; A model building and training module is used to establish an inversion deep learning model and use the defect and morphological feature information to perform distributed collaborative training and integrated regression testing on the inversion deep learning model to obtain an optimal prediction model; The optimal prediction model is used to identify propagable destructive defects and rough surface microstructures that appear during the laser finishing process of wafer substrates. This allows for early warning, prediction, or prevention information on defects and morphology during the laser finishing process and provides feedback to the laser industrial control system. as well as The laser industrial control system is used to adjust the atomic-level laser finishing parameters according to early warning information, prediction information or prevention information to form a closed-loop control, so as to accurately control and improve the quality of the laser processing technology of the wafer substrate.

10. The intelligent recognition device for atomic-level laser processing of a wafer substrate according to claim 9, characterized in that, The information acquisition module includes: transmission electron microscope, atomic force microscope and Raman spectroscopy, wherein, The transmission electron microscope is used to collect morphology information, crystal structure information, defect and interface information, and composition analysis information of the wafer substrate; The atomic force microscope is used to collect surface morphology information, surface mechanical properties and surface charge distribution information; and The Raman spectrum is used to collect molecular structure information, lattice vibration information and chemical composition analysis information of the wafer substrate.

Images