An intelligent identification method and device for atomic-level laser processing of wafer substrates

Abstract

The present application relates to an intelligent identification method and device for atomic-level laser processing of wafer substrates, belonging to the technical field of laser processing, and solving the problem that it is difficult to simultaneously obtain a wafer substrate with a super-smooth surface, a complete surface lattice without defects, and no surface or subsurface damage in the existing laser finishing technology. The method comprises collecting defect and topographic feature information generated in the interaction process between laser finishing and the atomic layer of the wafer substrate surface and storing it in a feature database; establishing an inversion deep learning model and training and testing it; using the optimal prediction model to identify propagating destructive defects and rough surface microtopography structures to obtain early warning, prediction or prevention information and feed it back to the industrial control system; and adjusting the atomic-level laser finishing parameters to form a closed-loop control. Through early warning, prediction and prevention of defects occurring during laser finishing of the wafer substrate, defects generated during the atomic layer removal process of the wafer substrate are quickly, accurately and effectively identified.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser processing, and in particular to an intelligent identification method and device for atomic-level laser processing of wafer substrates. BACKGROUND

[0002] Wafers are widely used in the fields of aerospace, electronics and electrical appliances, automobiles and ships, etc., as ideal substrate materials for various microelectronic 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 need to be carried out to ensure their quality and reliability.

[0003] Among them, mechanical grinding and chemical mechanical polishing are key processes for removing cutting marks and metamorphic layers of wafer substrates, achieving planarization and super-smoothness. However, due to the characteristics of high hardness, high brittleness and strong chemical inertness of these crystal materials, defects and damage of wafer substrates are easily caused, and the material removal efficiency is low, which will directly spread in a large range in the wafer epitaxy process.

[0004] Laser finishing, as one of the important technologies for surface finishing, controls the surface finish of workpieces by selective melting, redistribution and precise removal of material through molten pool movement. In recent years, it has been highly valued by domestic and foreign research institutes and related enterprises. However, the existing laser finishing method is difficult to obtain a wafer substrate with super-smooth surface, complete and defect-free surface lattice, and no surface or subsurface damage.

[0005] With the increasing demand for intelligent manufacturing development, artificial intelligence technologies represented by deep learning have been widely applied in various fields of industry. In the application of object defect detection and identification, the demand for image processing technology is increasing. In order to early warn, predict and prevent a series of defects in the process of laser finishing of wafer substrates, 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 processing defects in the process of laser finishing of wafer substrates, especially a series of defects produced in the process of atomic layer removal of wafer substrates, and to accurately quantify them.

[0006] Therefore, it is urgent to develop a simple and highly intelligent processing method that can efficiently and accurately identify and predict various defects in the process of atomic-level laser finishing wafer substrates, and accurately quantify these defects to obtain a set of intelligent laser finishing wafer substrates with super-smooth surfaces, complete and defect-free surface lattices, and no surface or subsurface damage, thereby replacing the traditional two-step process of mechanical grinding and chemical mechanical polishing of wafer substrates, and improving the manufacturing yield and processing efficiency of wafer substrates. SUMMARY

[0007] In view of the above analysis, the embodiments of the present application aim to provide an intelligent identification method and device for atomic-level laser processing wafer substrates to solve the problem that the existing laser finishing method cannot simultaneously obtain wafer substrates with super-smooth surfaces, complete and defect-free surface lattices, and no surface or subsurface damage.

[0008] In one aspect, the embodiments of the present application provide an intelligent identification method for atomic-level laser processing wafer substrates, comprising: collecting defect and topography feature information generated in the interaction process between laser finishing and the atomic layer of the wafer substrate surface and storing it in a feature database; establishing an inverse deep learning model and using the defect and topography 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 propagating destructive defects and rough surface micro-topography structure that occur in the process of laser finishing wafer substrates to obtain early warning information, prediction information or prevention information of defects and topography in the laser finishing process and feed back 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 information to form a closed-loop control to accurately regulate and improve the quality of the laser processing technology of the wafer substrate.

[0009] The beneficial effects of the above technical solution are as follows: the intelligent identification method for atomic-level laser processing wafer substrates according to the embodiments of the present application can monitor and predict processing defects in real time, avoiding the large-scale propagation of defects and improving the manufacturing yield. Secondly, by accurately quantifying the defects, the system can accurately process different defect types and degrees, avoiding the problem of excessive grinding or polishing in traditional methods and improving the processing efficiency. Most importantly, the intelligent laser finishing method can obtain wafer substrates with super-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.

[0010] Based on the further improvement of the above method, the defect and topography characteristic information generated in the process of laser finishing and atomic layer interaction of wafer substrate surface includes: collecting the topography information, crystal structure information, defect and interface information and composition analysis information of wafer substrate by transmission electron microscope; collecting the surface topography information, surface mechanical properties and surface charge distribution information by atomic force microscope; and collecting the molecular structure information, lattice vibration information and chemical composition analysis information of wafer substrate by Raman spectrum.

[0011] Based on the further improvement of the above method, the defect and topography characteristic information generated in the process of laser finishing and atomic layer interaction of wafer substrate surface includes: utilizing the transmission electron microscope, the atomic force microscope and the Raman spectrum combined with sorting by region and type, comprehensively collecting the defect and topography characteristic information generated in the process of laser finishing and atomic layer interaction of wafer substrate surface, wherein the sorting by region means that representative data of wafer substrate surface is obtained by selecting different regions or sample points to reveal the regional distribution of surface defects; the sorting by type means that different types of defects are distinguished and the morphology of defects is identified by different characterization methods, wherein different types of defects include point defects, line defects and surface defects; and the sorting means that a large amount of characterization data obtained is sorted and processed according to a specific rule, wherein the specific rule includes sorting according to defect type, defect size or defect density, or sorting according to surface region.

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

[0013] Based on the further improvement of the above method, the inversion deep learning model comprises an encoder-decoder module, an image processing module, a relationship processing module, an attention mechanism module and an activation function module, wherein the encoder-decoder module is used to segment and reconstruct the input image, 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 topography information for various conversion and analysis processing 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 relationship between atomic-level defects on the surface to obtain new feature representation of integrated image relationship 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 to the micro-nano surface structure of the wafer, improve the accuracy of atomic-level defect identification and positioning, 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 to atomic-level defects and structural features.

[0014] Based on the further improvement of the above method, the attention mechanism module comprises a basic feature extraction unit, a structured attention unit, a channel attention unit, a redundant attention unit and a hierarchical fusion unit, wherein 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, wherein the multi-scale visual features include edge, texture 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 each level, focus on key structured detail features to generate attention weights of different regions and provide them to the redundant attention unit and the hierarchical fusion unit; the channel attention unit is used to adaptively reweight the convolutional channels through a squeeze-excitation mechanism, enhance the ability to depict atomic-level structures to generate attention weights of 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 of intermediate layer features in 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 expression more sensitive to atomic-level details.

[0015] Based on the further improvement of the above method, the early warning information is that when the optimal prediction model identifies that a transmissible destructive defect appears in the wafer substrate or the surface roughness exceeds the roughness threshold, the transmissible destructive defect or the non-compliant roughness is taken as the early warning information; the prediction information is that when the optimal prediction model identifies that no transmissible destructive defect appears in the wafer substrate or the surface roughness does not exceed the roughness threshold, the defects appearing in the subsequent processing are predicted according to the current state of the substrate wafer, and the type, size, position distribution information of the defects are taken as the prediction information; and the prevention and control information is that the surface topography evolution law of the wafer substrate is determined according to the prediction result of the optimal prediction model, and the improvement measures recommended according to the surface topography evolution law are taken as the prevention and control information.

[0016] Based on the further improvement of the above method, the laser industrial control system adjusts the atomic-level laser finishing parameters according to the early warning, prediction and prevention and control 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 point position, so as to weaken or enhance the action intensity of the laser on the defect area; when the laser industrial control system receives the prediction information, the laser industrial control system adjusts the laser parameters before laser finishing in a targeted manner to avoid key areas, or to reduce the spot size and scanning overlap of the key areas, wherein the key area is a key area that may cause failure; when the laser industrial control system receives the prevention and control information, the laser industrial control system changes the scanning strategy, power distribution according to the suggestion provided by the optimal prediction model, or uses other wavelength lasers to re-scan to prevent or eliminate the predicted defects.

[0017] In another aspect, the embodiment of the present application provides an intelligent identification device for atomic-level laser processing of wafer substrates, comprising: an information acquisition module for acquiring defect and topographic feature information generated in the interaction process between laser finishing and the atomic layer of the wafer substrate surface; a feature database for storing the defect and topographic feature information; a model construction and training module for establishing an inverse deep learning model and performing distributed collaborative training and integrated regression testing on the inverse deep learning model by using the defect and topographic feature information, to obtain an optimal prediction model; an optimal prediction model for identifying transmissible destructive defects and rough surface micro-topography structures appearing in the process of laser finishing of wafer substrates, to obtain early warning information, prediction information or prevention and control information of defects and topography in the laser finishing process and feed back to the laser industrial control system; and a laser industrial control system for adjusting 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 accurately control and improve the quality of the laser processing technology of the wafer substrate.

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

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

[0020] (1) Compared with traditional mechanical grinding and chemical mechanical polishing, the present application can monitor and predict processing defects in real time, avoiding the wide spread of defects and improving the manufacturing yield. Secondly, through accurate quantitative analysis of defects, the system can accurately process different defect types and degrees, avoiding the problem of excessive grinding or polishing in traditional methods and improving 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 needs of high-precision microelectronics and integrated circuit chip manufacturing.

[0021] (2) The present application adopts 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 topography information generated during the laser finishing of wafer substrates. The microscopic details and topography features of the wafer surface and crystal structure can be captured. These atomic-level characterization technologies can observe and analyze the structure, topography and chemical composition of materials at the nanoscale, with high resolution and high sensitivity. The present application uses transmission electron microscopy, atomic force microscopy and high-resolution Raman spectroscopy characterization technology combined with region-by-region and type-by-type sorting self-programming to comprehensively collect defect and topography feature information generated during the interaction between laser finishing and the atomic layer of the wafer substrate surface. In contrast, existing technologies can only obtain limited local information and cannot provide comprehensive atomic-level data. By combining data obtained from different atomic-level characterization technologies (such as image data, topological data and spectral data), multi-modal data fusion can be achieved, which can provide more comprehensive and diverse information and help improve the robustness and generalization ability of deep learning models.

[0022] (3) The present application adopts a self-programmed inverse deep learning intelligent algorithm combined with a self-adaptive interactive model, selects the optimal model through distributed collaborative training and integrated regression testing. Such an intelligent algorithm can more accurately predict design defects and topography feature information and provide higher prediction accuracy and precision.

[0023] (4) The present application combines the optimal model, can identify the propagating destructive defects and rough surface micro-topography structure appeared in the process of laser finishing wafer substrate, and reveals the evolution law of joint product appeared in the process of laser finishing wafer substrate, and the forming mechanism of non-thermal removal and planarization of surface layer material. This makes the processing equipment can give the early warning, prediction and prevention information of defects and morphology in the process of laser finishing, help to avoid potential problems and loss, and improve the production efficiency.

[0024] (5) The present application feeds back the early warning, prediction and prevention information in step (3) to the laser industrial control system, realizes 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, so as 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, improve the quality and reliability of the product.

[0025] In the present application, the above-mentioned technical solutions can be combined with each other to realize more preferred combination schemes. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification, or will be understood by implementing the present application. The purpose and other advantages of the present application can be realized and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0026] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0027] Figure 1 Flow chart of the intelligent identification method for atomic level laser processing wafer substrate according to the embodiment of the present application;

[0028] Figure 2 Specific flow chart of the intelligent identification method for atomic level laser processing wafer substrate according to the embodiment of the present application;

[0029] Figure 3 Structure diagram of the inversion deep learning model according to the embodiment of the present application;

[0030] Figure 4 Block diagram of the intelligent identification device for atomic level laser processing wafer substrate according to the embodiment of the present application;

[0031] Figure 5 Structure diagram of the encoder-decoder module according to the embodiment of the present application;

[0032] Figure 6 Structure diagram of the image processing module according to the embodiment of the present application;

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

[0034] Figure 8 This is a structural diagram of the attention mechanism module according to an embodiment of the present invention;

[0035] Figure 9 This is a structural diagram of the activation function module according to an embodiment of the present invention. Detailed Implementation

[0036] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0037] The purpose of this invention is to provide an intelligent identification method for atomic-level laser-processed wafer substrates, replacing the traditional two-step process of mechanical grinding and chemical mechanical polishing. This method can efficiently and accurately identify and predict various defects occurring during atomic-level laser-processed wafer substrates, and accurately quantify these defects, thereby obtaining wafer substrates with ultra-smooth surfaces, intact and defect-free crystal lattices, and no surface or subsurface damage. This method utilizes deep learning technology to establish an intelligent database, enabling rapid, accurate, and effective identification of processing defects and morphological features during laser-processed wafer substrates. This allows for real-time monitoring and prediction of atomic-level surface quality during laser processing, and preventative measures are taken by actively adjusting equipment parameters, effectively improving the processing effect of atomic-level laser processing.

[0038] like Figure 1 As shown, a specific embodiment of the present invention discloses an intelligent identification method for atomic-level laser-processed wafer substrates, comprising: in step S101, 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; in step S102, 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; in step S103, using the optimal prediction model to identify propagable destructive defects and rough surface microstructures that occur during the laser finishing of the wafer substrate, to obtain early warning information, prediction information, or prevention information on defects and morphology during the laser finishing process and feeding it back 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 information to form a closed-loop control, so as to precisely control and improve the quality of the laser processing technology 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 the embodiment can monitor and predict processing defects in real time, avoid large-scale propagation of defects, and improve the manufacturing yield. Secondly, through accurate quantitative analysis of defects, the system can accurately process different defect types and degrees, avoiding the problem of excessive grinding or polishing that may be caused by traditional methods, and improving the processing efficiency. Most importantly, the intelligent laser finishing method can obtain wafer substrates with super-smooth surface, complete and defect-free surface lattice, and no surface or subsurface damage, meeting the needs of high-precision microelectronics and integrated circuit chip manufacturing.

[0040] Hereinafter, reference will be made to Figure 1 The various steps of the intelligent identification method for atomic-level laser processing of wafer substrates according to the embodiment of the application will be described in detail.

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

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

[0043] Specifically, the defect and topography feature information generated in the interaction process of laser finishing and atomic layer of wafer substrate surface includes: using a transmission electron microscope, an atomic force microscope, and a Raman spectrum combined with sorting by region and type to comprehensively collect the defect and topography feature information generated in the interaction process of laser finishing and atomic layer of wafer substrate surface, wherein the sorting by region means that representative data of the wafer substrate surface is obtained by selecting different regions or sample points to reveal the regional distribution of surface defects; the sorting by type means that different types of defects are distinguished, and the morphology of the defects is identified by using different characterization methods, wherein the different types of defects include point defects, line defects, and surface defects; and the sorting means that a large amount of characterization data is sorted according to a specific rule, wherein the specific rule includes sorting according to the defect type, the defect size, or the defect density, or sorting according to the surface region.

[0044] In step S102, an inversion deep learning model is established, and distributed collaborative training and integrated regression testing of the inversion deep learning model are performed by using the defect and topography feature information to obtain an optimal prediction model.

[0045] Reference Figure 3 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, wherein the encoder-decoder module is used to segment and reconstruct the input image, identify atomic-level features and defects, and specifically, the defect and topography feature information generated in the interaction process of laser finishing and atomic layer of wafer substrate surface is comprehensively collected by using a transmission electron microscope, an atomic force microscope, and a Raman spectrum combined with sorting by region and type as image information input into the inversion learning model; the image processing module is used to analyze the input image reconstructed by the encoder-decoder module and extract surface topography information for various conversion and analysis processing 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 topography information into information such as the distribution and quantity of different types of defects; analyzes the defect type and severity, and the surface topography evolution law to determine whether the defect or roughness exceeds a predetermined threshold; and further obtains the following results: digital features of the surface topography and defect type / distribution of the wafer sample; the relationship processing module is used to process the topological structure or spatial correlation of surface defects on the surface topography information, analyze the relationship 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 inversion deep learning model to the wafer micro-nano surface structure, improve the accuracy of the identification and positioning of atomic-level defects to obtain a feature subset that needs to be focused on 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 to atomic-level flaws 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, wherein the basic feature extraction unit 801 is configured to extract multi-scale visual features from the wafer substrate image using a convolutional neural network and provide the multi-scale visual features to the structured attention unit and the channel attention unit, wherein the multi-scale visual features include edge, texture, and shape features; the structured attention unit 802 is configured to adopt a pyramid pooling structure, connect feature maps of different scales, learn a weighted representation of features at each level, focus on key structured detail features to generate attention weights for different regions, and provide the attention weights to the redundant attention unit and the hierarchical fusion unit; the channel attention unit 803 is configured to use a squeeze-excitation mechanism to adaptively reweight the convolutional channels, enhance the ability to depict atomic-level structures, generate attention weights for different channels, and provide the attention weights to the redundant attention unit and the hierarchical fusion unit; the redundant attention unit 804 is configured to use attention scores to identify irrelevant regions, remove redundant interference terms of intermediate layer features in the image, generate fusion attention weights, and provide the fusion attention weights to the hierarchical fusion unit; and the hierarchical fusion unit 805 fuses the attention features of different modules together 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 propagating destructive defects and rough surface micro-topography structures that occur in the laser finishing wafer substrate process, so as to obtain the early warning information, prediction information, or prevention and control information of the defects and topography in the laser finishing process and feed back to the laser industrial control system.

[0048] Specifically, when the optimal prediction model identifies that the wafer substrate has a propagating destructive defect or the surface roughness exceeds the roughness threshold, the propagating destructive defect or the roughness that does not meet the requirements is taken as the early warning information. When the optimal prediction model identifies that the wafer substrate does not have a propagating destructive defect or the surface roughness does not exceed the roughness threshold, the type, size, and position distribution information of the defect that occurs in the subsequent processing are predicted according to the current state of the substrate wafer, and the prediction information is taken as the prediction information. The prevention and control information is determined according to the prediction result of the optimal prediction model to determine the surface topography evolution law of the wafer substrate, and the improvement measures recommended according to the surface topography evolution law are taken as the prevention and control information.

[0049] In step S104, the laser industrial control system adjusts the atomic-level laser finishing parameters according to the early warning information, the prediction information, or the prevention and control 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.

[0050] Specifically, 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 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 point position, to weaken or enhance the action intensity of the laser on the defect area; when the laser industrial control system receives the prediction information, the laser industrial control system adjusts the laser parameters before laser finishing to avoid key areas, or to reduce the spot size and scanning overlap of the key areas, wherein the key area is a key area that may cause failure; when the laser industrial control system receives the prevention information, the laser industrial control system changes the scanning strategy, power distribution according to the suggestion provided by the optimal prediction model, or uses other wavelength lasers to rescan to prevent or eliminate the predicted defects.

[0051] Another specific embodiment of the present application, with reference to Figure 4 , discloses an intelligent identification device for atomic-level laser processing of wafer substrates, comprising: an information acquisition module 401 for acquiring defect and topographic feature information generated in the interaction process between laser finishing and the atomic layer of the wafer substrate surface; a feature database 402 for storing the defect and topographic feature information; a model construction and training module 403 for establishing an inverse deep learning model and using the defect and topographic feature information to perform distributed collaborative training and integrated regression testing on the inverse deep learning model to obtain an optimal prediction model; an optimal prediction model 404 for identifying transmissible destructive defects and rough surface micro-topographic structures that occur during laser finishing of the wafer substrate to obtain early warning information, prediction information or prevention information of defects and topography during the laser finishing process and feed back to the laser industrial control system; and a laser industrial control system 405 for adjusting atomic-level laser finishing parameters according to the early warning information, prediction information or prevention information to form a closed loop control to accurately regulate and improve the quality of the laser processing of the wafer substrate.

[0052] 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 acquire topographic information, crystal structure information, defect and interface information and composition analysis information of the wafer substrate; the atomic force microscope is used to acquire surface topographic information, surface mechanical properties and surface charge distribution information; and the Raman spectrum is used to acquire 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 the embodiments of the present application is described in detail in the form of specific examples.

[0054] Reference Figure 2The application discloses an intelligent identification method for atomic-level laser processing of wafer substrates, and comprises the following steps: step one, a defect and morphology characteristic information database generated in the interaction process of laser finishing and the surface atomic layer of a wafer substrate is established by using a transmission electron microscope (TEM), an atomic force microscope (AFM) and high-resolution Raman spectrum characterization technology combined with a region-type sorting self-programming; step two, a self-adaptive interactive model prediction design is used to perform distributed collaborative training and integrated regression testing on the defect and morphology characteristic information data in step one, and an optimal model is selected; step three, the optimal model in step two is combined to identify the propagating destructive defects and rough surface micro-morphology structure generated in the laser finishing process of the wafer substrate, the evolution law of the joint product generated in the laser finishing process of the wafer substrate is revealed, and the non-thermal removal and planarization formation mechanism of the surface material is revealed, and early warning, prediction and prevention information of the defects and morphology in the laser finishing process is provided; and step four, the early warning, prediction and prevention information in step three is fed back to a laser industrial control system, atomic-level laser finishing parameters are adjusted in advance, a stable closed-loop control is formed, and an optimal atomic-level laser finishing method is achieved.

[0055] The region-type sorting self-programming in step one is used to realize collation, analysis and mining of the TEM, AFM and high-resolution Raman spectrum characterization data.

[0056] The defect and morphology characteristic information database in step one comprises defect types, defect sizes, defect distributions, defect densities, surface morphologies, surface structures, surface chemical components, electronic structures and sample attribute data related to the laser finishing process.

[0057] In the application, the TEM, AFM and high-resolution Raman spectrum characterization technology can obtain various information before and after laser finishing of the wafer substrate.

[0058] 1. The TEM can be used to characterize:

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

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

[0061] Defect and interface information: the TEM can detect defects, grain boundaries and interfaces in the wafer substrate, and provide information about defect types, densities and interface structures.

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

[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.

[0066] Surface charge distribution: By introducing current or external voltage on 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 type of chemical bonds.

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

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

[0071] Further, the sub-regional sorting in the self-programming program refers to selecting different regions or sample points for characterization when characterizing the wafer substrate surface, in order to obtain representative data of the surface and reveal the regional distribution of surface defects. The sub-type refers to distinguishing different types of defects, such as point defects, line defects, and surface defects, for classification and analysis, and using the advantages of different characterization techniques to identify the morphology of defects. The sorting refers to sorting the large amount of characterization data obtained according to specific rules, which may include sorting according to defect type, size, density, etc., or sorting according to surface region, etc. The self-programming program refers to the algorithm or program code written by researchers for data processing and analysis, through which the data is sorted, analyzed, and mined.

[0072] Specifically, the defect and morphology characteristic 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 sub-type defects of single / group defects, and lattice defect structures such as vacancies and substitution atoms;

[0074] (2) Defect size: geometric size parameters of the defect, such as length, width, depth, height, etc.

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

[0076] (4) Defect distribution: coordinate position distribution information of defects on the wafer surface, distribution functions and parameters of Gaussian and Poisson distribution, and cluster distribution on different distance scales and spatial correlation scales;

[0077] (5) Defect density: the number of defects per unit area; surface morphology: surface roughness parameters, peak-to-valley height, morphology curve information, and 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 mismatch dislocation;

[0079] (7) Surface chemical composition: high-resolution Raman spectroscopy to determine surface chemical bond structure and composition change information, component and chemical valence XPS peak;

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

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

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

[0083] Further, the self-compiled inverse deep learning intelligent algorithm of step two includes a deep neural network with an encoding-decoding structure, an image processing network based on a convolutional neural network, an Attention mechanism external module, and a self-encoder regularization module, thereby forming a mathematical model of a high-order self-compiled inverse deep learning intelligent algorithm. Distributed collaborative training and integrated regression testing are performed on laser finishing data without accurate model information to ensure the stability and convergence of the model.

[0084] Further, the distributed collaborative training and integrated regression testing in step two are mainly used to accelerate the model training process, improve the model generalization performance, conduct integrated regression prediction, evaluate the model and reduce the risk of overfitting, provide the basis for result decision-making, and enhance the performance and credibility of single model.

[0085] Further, the early warning, prediction and prevention information in step two refers to timely sending early warning signals to the operator during the laser finishing process, real-time prediction and giving the surface defect and topography type distribution and severity level of the next step processing, so as to prevent or reduce the expected defects and problems.

[0086] In the present application, the self-compiled inversion deep learning intelligent algorithm comprises the following main components:

[0087] The deep neural network of encoder-decoder structure is used to extract features and map the potential mapping relationship between input and output, similar to Seq2Seq or auto-encoder structure.

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

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

[0090] The attention mechanism or external storage module is used to capture long-distance dependence relationship, which helps to model the interaction between multiple regions of the surface.

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

[0092] The convolutional neural network interpretability method, such as Grad-CAM technology, explains the surface area focused by the network and understands the data features.

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

[0094] The integrated method integrates multiple networks to improve the generalization and enhance the result interpretability.

[0095] In the present application, the adaptive interaction model prediction design improves and optimizes the design of the model through interaction with experimental or simulation data. This process usually includes the following:

[0096] Data collection and preparation: Collect and organize datasets of defect and topography information, which can include experimental observation data, computational simulation data, literature reported data, process parameter data, equipment state data, wafer description data, process detection data, and process log data; as well as historical process data, external environment data, raw material data, operation and service data, model training and evaluation data, model optimization iteration data, etc.

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

[0098] Reference Figure 3 The inversion deep learning model in the intelligent identification method of atomic-level laser processing wafer substrate includes five modules: encoder-decoder module, image processing module, relationship processing module, attention mechanism module, and activation function module. The connection and role of the five modules are as follows: the encoder-decoder module is responsible for extracting semantic features and reconstructing context spatial information of wafer images, which is the representation learning of input images. Specifically, the original image data is input into the encoder-decoder module, and the output of the encoder-decoder module is learned as the image feature representation. The image processing module performs multi-scale image conversion on the features extracted by the encoder, enhances the perception of key microstructures, and provides image prior for feature expression. Specifically, the image feature representation is input into the image processing module, and the denoised and enhanced image features generated by the image processing module are output as the output of the image processing module. The relationship processing module analyzes the topological relationship between entities of the denoised and enhanced image features, finds the correlation pattern of atomic defects, realizes the structured modeling of features, and outputs new feature representation integrated with the relationship between images to the attention mechanism module. The attention mechanism module plays a guiding role in feature extraction and correlation analysis by suppressing indirectly related areas and weighting relevant details to generate a feature subset that needs to be focused on and provide it to the activation function module. The activation function module mainly acts in various neural networks, strengthens the nonlinearity of feature expression, and enhances the ability of the model to fit complex mappings to generate the final feature representation based on the activation function conversion.

[0099] These five modules each have their own focus while supporting each other to achieve more accurate and robust atomic-level defect characterization, ultimately realizing automated detection and quality control of wafer substrates.

[0100] (1) Reference Figure 5, the encoder-decoder module 500 can be used to extract feature information and map the mapping relationship between input data and output data by the combination of various constituent units, while performing fine image segmentation and reconstruction of micro-nano surfaces, effectively identifying atomic-level features and defects; the structure comprises: an encoder unit 501: usually adopting a recurrent neural network (RNN), a convolutional neural network (CNN) or an attention mechanism (Attention), encoding the input sequence into a fixed-length vector representation, capturing its semantic information; a decoder unit 503: adopting an RNN or a network with an attention mechanism, decoding the vector representation output by the encoder into the desired output sequence; further comprising a down-sampling 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 a context vector c, the decoder generates the output of the first time step according to the context vector c, takes the output of the first time step as the input of the second time step decoder, predicts the output of the second time step, and sequentially cycles until the complete output sequence is generated; for extracting feature information and mapping the mapping relationship between input data and output data.

[0101] The down-sampling path unit 502: gradually down-samples the encoded features through the max-pooling layer, expands the receptive field, and obtains more abstract global representation.

[0102] The skip connection unit 505: adopts the skip connection of the UNet type structure, directly transmits the shallow features to the deep layer, bypasses the nonlinear conversion, and preserves more original signals.

[0103] The attention gating mechanism unit 504: uses a gating unit to selectively strengthen the semantic information flow and focuses on the key atomic structure.

[0104] Specifically, the skip connection unit and the attention gating mechanism unit are used to cooperate with the encoder unit, the decoder unit and the down-sampling path unit to expand and enhance the three units, the skip connection unit accelerates the encoding-decoding process, and the attention gating mechanism gathers the most important part of the information, so that the entire encoder-decoder module can more efficiently and accurately complete feature extraction and input-output mapping. The relationship between the other framework units.

[0105] (2) Reference Figure 6 The image processing module 600 can be used to analyze input image information and extract surface topography information for various conversion and analysis processing, enhance the perception of atomic-micron surface shape, boundary and defects, improve the representation ability of atomic-level fine features, and is beneficial to subsequent identification and analysis, and its constituent units mainly include: a low-level feature extraction unit, a multi-scale fusion unit, a shape semantic unit, a sparse representation unit, a prior constraint unit, and an image segmentation network unit.

[0106] The low-level feature extraction unit 601 extracts directional features using a rod waveguide filter bank, enhances the description of edges and textures, and generates image features such as edges, textures, etc. Specifically, the rod waveguide filter bank used in the low-level feature extraction unit directly processes the input raw image, which is the original data obtained from the image acquisition device (transmission electron microscope, atomic force microscope, Raman spectrometer, etc.), before being encoded by the encoder-decoder module. The directional features are output by the image processing module and serve as input for subsequent modules such as the encoder-decoder. For the encoder-decoder module, the directional features extracted by the low-level feature extraction unit (low-level directional features) 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 relationships 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, low-level directional features can provide bottom-line line direction cues to guide the attention mechanism to focus on regions in a specific direction. For the activation function module, the response of low-level directional features can help select activation functions sensitive to lines and edges, such as ReLu (a machine learning activation function). Therefore, low-level directional features can provide bottom-line shape, edge, texture, etc. cues for each module, helping to better analyze and understand the local structure of the image, thereby improving the performance of the entire model.

[0107] The multi-scale fusion unit 602 uses a scale space pyramid structure to connect feature maps of different scales and fuse multi-scale information to generate image pyramid representations of different scales and provide them to the shape semantic unit.

[0108] The shape semantic unit 603 constructs a shape bag of words, learns word embedding representations of predefined graphemes, and describes regional shape features. The input to the shape semantic unit includes edge, texture, etc. image features 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 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 region of the image (these shape features are important information for microscopic surface topography). This output feature is provided as the result of the image processing module to subsequent encoder-decoder modules and other modules such as the attention mechanism module and the relationship processing module, helping to further image understanding and analysis processing.

[0109] Sparse representation unit 604: using dictionary learning and sparse coding algorithm, efficiently represent the key atomic structure. The input of the sparse representation unit mainly comes from the output of the previous multiple component units in the image processing module, including the image features output by the low-level feature extraction, multi-scale fusion, shape semantics, etc. These features reflect the atomic structure and surface topography information.

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

[0111] Prior constraint unit 605: combined with the physical model of crystal growth, add atomic arrangement priori and kinematic constraints, regularize feature expression. The input of the prior constraint unit is the surface topography image features output from the previous units in the image processing module, and the atomic arrangement priori knowledge and kinematic constraints provided by the physical model of crystal growth.

[0112] The output of the prior constraint unit is the 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, and the subsequent relationship processing module, etc. Enhanced surface topography features will help improve the accuracy of atomic level defect and structure representation.

[0113] (3) Reference Figure 7 The relationship processing module 700 can be used to process the topological structure or spatial correlation of surface defects, analyze the relationship between surface atomic level defects, and accurately identify the relationship between multi-modal heterogeneous data by combining various component units. Its structure mainly includes: entity recognition unit, relationship extraction unit, knowledge graph integration unit, joint reasoning unit, and explanation decision unit.

[0114] Entity recognition unit 701: using convolutional network to identify image regions, positioning various types of point, line and surface defects entities. The input of the entity recognition unit mainly comes from the output of the image processing module, including the image feature representation processed by low-level feature extraction, shape semantic learning, etc. These features contain surface topography, boundary, and potential defect information. Based on these input features, the entity recognition unit uses convolutional network for further analysis and identification, positioning and labeling various types of point, line and surface defects. The input of the entity recognition unit is the low-level features of the image, and the output is the position, category, etc. Information of each surface defect entity detected.

[0115] The relationship extraction module unit 702: based on the instance segmentation result, the topological connection, the direction relative position and other spatial relationships between entities are extracted. The input of the relationship extraction unit is the defect entity information output by the entity recognition unit, and the output is the spatial / topological relationship between different defect entities, for example, A and B have a containing relationship, C and D have a neighboring relationship, etc.

[0116] The knowledge graph integration unit 703: the extracted entities and entity relationships are fused into the pre-defined crystal structure knowledge graph. The input of the knowledge graph integration unit is the relationship between the defect entities output by the relationship extraction unit, and the externally injected surface defect field knowledge graph, and the output is the updated knowledge graph fused with new relationships. The pre-defined crystal structure knowledge graph mainly includes: a, crystal type graph: covering the structure 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 structure template of typical crystal defects, such as point defects, linear defects, and planar defects. 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 restriction model, bond optimization mechanism, dangling bond elimination principle, vacancy stability model, and surface reset template. e, material attribute relationship graph: the relationship graph that associates material types with attributes (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, and loss-stress relationship graph.

[0117] The joint reasoning module unit 704: based on the knowledge graph, the known atomic defect patterns are associated, and relationship reasoning analysis is performed. The input of the joint reasoning unit is the updated knowledge graph, and the output is the new relationship or conclusion obtained by joint reasoning of the knowledge graph, for example, A contains B, B is adjacent to C, so it can be inferred that A has a spatial relationship with C.

[0118] The explanation decision unit 705: according to the reasoning network result, a rule template is used to generate a textual description to explain the final atomic level defect recognition decision process. The input of the explanation decision unit is the relationship conclusion of the joint reasoning, and the output is an explanation in natural language form for the final relationship reasoning decision process for easy understanding.

[0119] (4) Reference Figure 8, the attention mechanism module 800, which contains a module design of rich attention mechanism, can enhance the model's perception of wafer micro-nano surface structure, improve the accuracy of atomic-level defect recognition and positioning, and be used to capture long-distance dependencies to help model the interaction between multiple regions of the surface. Specifically, the attention mechanism module can enhance the model's perception of wafer micro-nano surface structure: the attention mechanism can help the model learn to focus on important areas in the image, analyze the surface micro-nano structure in a targeted manner, and improve the ability to identify key details. Improve the accuracy of atomic-level defect recognition and positioning: the attention mechanism can guide the model to focus on areas where defects may exist, reducing the interference of irrelevant content, so as to more accurately identify the defect location. Capture long-distance dependencies: through attention calculation for different positions, the mutual influence between two distant regions can be modeled, and more complex dependencies can be captured. Model the interaction between multiple regions of the surface: different attention heads can focus on different regions of interest at the same time, model the correlation 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, pointing out the areas 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 angles through different attention units, and gradually fuses to obtain a comprehensive multi-granularity attention representation for the final image analysis. Its structure mainly includes: a basic feature extraction unit, a structured attention unit, a channel attention unit, a redundancy 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 edge, texture, shape, and other information. The input of the basic feature extraction unit is the original image, and the output is the basic visual features extracted through convolution and other operations.

[0121] The structured attention unit 802 uses a pyramid pooling structure to connect feature maps of different scales and learn the weighted representation of each level of features, paying 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 uses a squeeze-excitation mechanism to adaptively reweight the convolution channels and 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] Redundancy attention unit 804: identify irrelevant regions using attention scores, remove redundant interference items of image middle layer features. The input of the redundancy attention unit is the two attention features described above, and the output is the fusion attention weight obtained by calculating the redundancy between the attention features.

[0124] Hierarchical fusion unit 805: fuse the attention features of different modules through anchor boxes or skip connections to form a feature expression more sensitive to atomic-level details. The input of the hierarchical fusion unit is each of the above attention features, and the output is the final multi-level attention feature obtained by hierarchical fusion.

[0125] (5) Reference Figure 9 , activation function module 900, for enhancing the representation and adaptability of the network, and at the same time, through the combination of various component units, the model's ability to identify and express atomic-level defects and structural features can be enhanced, and 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: use global average pooling to compress convolutional features into vectors, emphasizing 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: uses a fully connected network to implement an adaptive channel adjustment mechanism, models the channel relationship of the input features, and enhances 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 weight of different atomic-level defect features.

[0128] Sigmoid gating unit 903: uses the Sigmoid function to achieve flexible regulation of the output, preventing gradient vanishing. The input of the Sigmoid gating unit is the original atomic-level defect multi-scale feature map and the feature weight 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 nonlinearity. Swish activation unit: the input is the feature map adjusted by the Sigmoid gating unit, and the output is the enhanced feature expression obtained by converting through 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.

[0131] Hard Swish activation function unit 906: Hard Swish is a piecewise approximation of Swish, more computationally efficient.

[0132] These function units are executed in series, gradually enhancing 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 collaborate through parallel computing and communication to speed up the training process and improve the performance of the model. This distributed training can be achieved using clusters, cloud computing or distributed computing frameworks. The data refers to the defect and topography feature information generated during the interaction between laser finishing and the atomic layer of wafer substrate surface stored in the feature database, and the model refers to the established inversion deep learning model.

[0134] Model optimization and improvement: through interaction with experimental or simulation data, iteratively optimize and improve the model. This may include model parameter adjustment, feature engineering, data augmentation, etc. to improve the accuracy and generalization ability of the model.

[0135] Integrated regression testing: select the optimal model and perform regression testing to evaluate the predictive performance of the model on unseen data. This can help verify the reliability and applicability of the model and provide guidance for further design and prediction. The established inversion deep learning model is selected for regression testing to evaluate the predictive ability of the predictive model on real-time data, and then the optimal predictive model is formed; the regression testing refers to using test data to evaluate the established deep learning prediction model to see how well it predicts actual data. Through this way of verifying the predictive performance of the model on an independent test set, the predictive 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 a test set, input the test set sample data into the already trained deep learning prediction model, obtain the model's predicted output for these samples, compare the model's predicted output with the actual labels in the test set, calculate some evaluation indicators such as MSE, R 2 If the evaluation indicators perform well, it means that the deep learning model has strong predictive ability for actual data and can be used as the final prediction model. If the indicators are poor, it needs to return to the above steps to further optimize the model.

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

[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 generalization performance of the model, distributed training can learn features from more samples, and different device parameter settings can form a model ensemble to enhance the generalization performance.

[0139] Perform integrated regression prediction, single model prediction may have bias or excessive variance problems, model integration can integrate multiple models to smooth and reduce the variance of regression prediction.

[0140] Evaluate the pros and cons of the model, 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 randomness to avoid overfitting of a single model to a specific training sample set.

[0142] Provide basis for result decision, regression test can evaluate the prediction quality of different models to support decision makers in selecting a reliable final model.

[0143] Overall, the performance and credibility of a single model are enhanced by using distributed and integrated technologies.

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

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

[0146] Prediction information: The optimal model can predict and give the distribution and severity level of the next step of processing surface defects, as well as quantitative indicators of surface morphology parameters such as roughness and Hole density.

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

[0148] In short, the intelligent model is to comprehensively utilize computing power, physical knowledge and data-driven to realize real-time monitoring and prediction of defects and surface topography in the processing, risk assessment and device optimization decision, play the guiding and auxiliary role of the automation system, and improve the stability of the processing and the product quality.

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

[0150] The early warning, prediction and prevention information specifically refers to:

[0151] Early warning information: refers to the warning signal sent to the laser industrial control system when the optimal prediction model detects that the wafer substrate appears a transmissible destructive defect or the surface roughness exceeds the threshold value, which can make the industrial control system pay attention to the possible problems in advance;

[0152] Prediction information: refers to the information about the defect type, size, position distribution and the like that may occur in subsequent processing predicted by the optimal prediction model according to the current wafer state, which can make the industrial control system predict the possible failure mode and give the optimal laser finishing parameters in time;

[0153] Prevention information: refers to the corresponding solutions or laser finishing parameter adjustment suggestions sent to the laser industrial control system by the optimal prediction model according to the experience knowledge base, intelligent algorithm and the like after receiving the above early warning or prediction information, which is to prevent or alleviate the predicted problems.

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

[0155] Early warning information: when the monitoring system detects defects or surface roughness abnormalities, it will send an early warning signal, at which 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 focal point position to weaken or enhance the action intensity of the laser on the region;

[0156] Prediction information: after the prediction model gives the possible failure mode, position and the like, the industrial control system can adjust the laser parameters in advance, for example, avoiding the key area where the failure may occur, or reducing the spot size and scanning overlap of the area;

[0157] Prevention information: the industrial control system will change the scanning strategy, power distribution, use other wavelength lasers to rescan and the like according to the suggestions provided by the prediction model to prevent or eliminate the predicted defects and problems.

[0158] The whole matching feedback and control process can be continuously optimized and improved by machine learning algorithm, that is, online process control, or re-collecting defect and topographic feature information to repeat the whole model optimization process, thereby realizing accurate regulation and quality improvement of laser processing.

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

Claims

1. An intelligent identification method for atomic level laser processing of wafer substrates, characterized by, The method comprises the following steps: Collecting defect and topography feature information generated in the interaction process of laser finishing and atomic layer of wafer substrate surface and storing the information in a feature database; Establishing an inversion deep learning model and using the defect and topography 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 propagating destructive defects and rough surface micro-topography structure in the laser finishing process of wafer substrate to obtain early warning information, prediction information or prevention and treatment information of defects and topography in the laser finishing process and feed back to a laser industrial control system; And The laser industrial control system adjusts atomic-level laser finishing parameters according to the early warning information, prediction information or prevention and treatment information to form a closed-loop control to accurately control and improve the quality of the laser processing technology of the wafer substrate, Wherein, the inversion deep learning model comprises an encoder-decoder module, an image processing module, a relationship processing module, an attention mechanism module and an activation function module, The encoder-decoder module is used to extract semantic features of wafer images and reconstruct context spatial information, and output image feature representation; The image processing module is used to convert and analyze the image feature representation to obtain denoised and enhanced image features and provide them to the relationship processing module; The relationship processing module is used to analyze the topological relationship between entities of the denoised and enhanced image features and the relationship between surface atomic-level defects to obtain new feature representation of integrated image relationship and provide it to the attention mechanism module; The attention mechanism module is used to obtain a feature subset of focus by suppressing indirectly related areas and weighting features of focus and provide them to the activation function module; and The activation function module is used to enhance the representation of atomic-level defect features to output feature representation for identifying and classifying the propagating destructive defects and rough surface micro-topography structure.

2. The intelligent identification method of atomically-precise laser processing wafer substrate according to claim 1, wherein, Collecting defect and topography feature information generated in the interaction process of laser finishing and atomic layer of wafer substrate surface comprises: Collecting topography information, crystal structure information, defect and interface information and composition analysis information of wafer substrate by transmission electron microscopy; Collecting surface topography information, surface mechanical properties and surface charge distribution information by atomic force microscopy; and Collecting molecular structure information, lattice vibration information and chemical composition analysis information of wafer substrate by Raman spectroscopy.

3. The intelligent identification method of atomically-precise laser processing wafer substrate according to claim 2, characterized in that, Collecting defect and topography feature information generated in the interaction process of laser finishing and atomic layer of wafer substrate surface comprises: using the transmission electron microscopy, the atomic force microscopy and the Raman spectroscopy in combination with region-based and type-based sorting to comprehensively collect defect and topography feature information generated in the interaction process of laser finishing and atomic layer of wafer substrate surface, wherein, The region-based representation obtains representative data of wafer substrate surface by selecting different regions or sample points to reveal the regional distribution of surface defects; The type distinguishing representation distinguishes different types of defects and identifies the morphology of the defects using different representation methods, wherein the different types of defects include point defects, line defects, and surface defects; and The sorting represents that a large amount of obtained representation data is sorted according to a specific rule, wherein the specific rule includes sorting according to defect types, defect sizes, or defect densities, or sorting according to surface regions.

4. The intelligent identification method of atomically-precise laser processing wafer substrate according to claim 1, wherein, The feature database includes the following data: defect types, defect sizes, defect distributions, defect densities, surface topographies, surface structures, surface chemical compositions, electronic structures, and sample attribute data related to the laser finishing process.

5. The intelligent identification method of atomically-precise laser processing wafer substrate according to claim 1, wherein, The attention mechanism module includes a basic feature extraction unit, a structured attention unit, a channel attention unit, a redundancy attention unit, and a hierarchical fusion unit, wherein The basic feature extraction unit is configured to extract multi-scale visual features from the wafer substrate image using a convolutional neural network and provide the multi-scale visual features to the structured attention unit and the channel attention unit, wherein the multi-scale visual features include edge, texture, and shape features; The structured attention unit is configured to adopt a pyramid pooling structure, connect feature maps of different scales, learn weighted representations of features at different levels, focus on key structured detail features to generate attention weights of different regions, and provide the attention weights to the redundancy attention unit and the hierarchical fusion unit; The channel attention unit is configured to use a squeeze-excitation mechanism to adaptively reweight convolution channels, enhance the ability to depict atomic-level structures, generate attention weights of different channels, and provide the attention weights to the redundancy attention unit and the hierarchical fusion unit; The redundancy attention unit is configured to use attention scores to identify irrelevant regions, remove redundant interference terms of intermediate layer features in the image, generate fusion attention weights, and provide the fusion attention weights to the hierarchical fusion unit; and The hierarchical fusion unit fuses attention features of different modules together through an anchor box or a skip connection structure to form a feature expression that is more sensitive to atomic-level details.

6. The intelligent identification method for atomic-level laser processed wafer substrates according to claim 1, wherein The early warning information is the transmissible destructive defects or the roughness that does not meet the requirements when the optimal prediction model identifies that the wafer substrate has transmissible destructive defects or the roughness exceeds the roughness threshold; The prediction information is the type, size, and position distribution information of the defects predicted according to the current state of the wafer substrate when the optimal prediction model identifies that the wafer substrate does not have transmissible destructive defects or the roughness does not exceed the roughness threshold; The prevention and control information is the surface topography evolution law of the wafer substrate determined according to the prediction result of the optimal prediction model, and the improvement measures recommended according to the surface topography evolution law are taken as the prevention and control information. The laser industrial control system adjusts the atomic-level laser finishing parameters according to the early warning, prediction, and prevention and control information, including:

7. The intelligent identification method of atomically-precise laser processing wafer substrate according to claim 6, wherein, ​ When the laser industrial control system receives the early warning information, the laser industrial control system reduces laser power, increases scanning line spacing, adjusts spot shape, scanning rate, laser energy or focal point position to weaken or enhance the effect of laser on the defect area; When the laser industrial control system receives the prediction information, the laser industrial control system adjusts the laser parameters before laser finishing to avoid the key area or reduce the spot size and scanning overlap of the key area, wherein the key area is the area that causes failure in the subsequent processing process; When the laser industrial control system receives the prevention and treatment information, the laser industrial control system changes the scanning strategy, power distribution, scanning speed, spot size, spot position according to the suggestion provided by the optimal prediction model, or uses other wavelength laser to rescan to prevent or eliminate the predicted defects.

8. An intelligent identification device for atomic level laser processing of a wafer substrate, characterized by, It comprises: An information acquisition module for acquiring defect and topographic feature information generated in the interaction process between laser finishing and atomic layer of wafer substrate surface; A feature database for storing the defect and topographic feature information; A model construction and training module for establishing an inverse deep learning model and performing distributed collaborative training and integrated regression testing on the inverse deep learning model using the defect and topographic feature information to obtain an optimal prediction model; An optimal prediction model for identifying transmissible destructive defects and rough surface micro-topography structure in the process of laser finishing wafer substrate to obtain early warning information, prediction information or prevention and treatment information of defects and topography in the laser finishing process and feed back to the laser industrial control system; And A laser industrial control system for adjusting atomic-level laser finishing parameters according to early warning information, prediction information or prevention and treatment information to form a closed-loop control to accurately control and improve the quality of the laser processing technology of the wafer substrate, Wherein, the inverse deep learning model comprises an encoder-decoder module, an image processing module, a relationship processing module, an attention mechanism module and an activation function module, The encoder-decoder module is used to extract semantic features and reconstruct context spatial information of wafer images, and output image feature representation; The image processing module is used to convert and analyze the image feature representation to obtain denoised and enhanced image features and provide them to the relationship processing module; The relationship processing module is used to analyze the topological relationship between entities of the denoised and enhanced image features and the relationship between surface atomic-level defects to obtain new feature representation of integrated image relationship and provide it to the attention mechanism module; The attention mechanism module is used to obtain a feature subset of focus by suppressing indirectly related areas and weighting features of focus and provide it to the activation function module; and The activation function module is used to enhance the representation of atomic-level defect features to output feature representation for identifying and classifying transmissible destructive defects and rough surface micro-topography structure.

9. The apparatus of claim 8, wherein the apparatus is configured to identify the wafer substrate as a wafer substrate that is suitable for atomic level laser processing when the measured value of the parameter is within a predetermined range of values. The information acquisition module comprises a transmission electron microscope, an atomic force microscope and a Raman spectrum, wherein The transmission electron microscope is used for collecting topographic information, crystal structure information, defect and interface information and composition analysis information of the wafer substrate; The atomic force microscope is used for collecting surface topographic information, surface mechanical properties and surface charge distribution information; and The Raman spectrum is used for collecting molecular structure information, lattice vibration information and chemical composition analysis information of the wafer substrate.

Images