A Method and System for Optimizing Wafer Dicing Parameters

Abstract

The present invention discloses a method and system for optimizing wafer cutting parameters. The method includes data collection and processing, as well as function fitting and machine learning model training based on the collected data. The range of the modification area and the initial cutting process parameters are determined according to the wafer characteristic data required by the project. The K-means clustering algorithm is used to determine the initial cutting process parameters. Finally, the initial cutting process parameters are optimized and adjusted based on the machine learning model and the fitting function. Compared with the prior art, the technical solution can significantly improve the stability of the cutting process and the product quality, while reducing the production cost. Through the application of the machine learning model, the degree of automation is greatly improved, the dependence on manual experience is reduced, the parameter optimization of the semiconductor stealth cutting process is made more efficient and accurate, and the increasingly strict industrial production requirements can be met.

Description

Technical Field

[0001] The present invention belongs to the technical field of parameter control in semiconductor stealth dicing process, and particularly relates to a method and system for optimizing wafer dicing parameters. Background Art

[0002] The stealth dicing process in semiconductor manufacturing is a key technology, which is widely used in the cutting of ultra-thin wafers, the manufacturing of optoelectronic devices, and micro sensors. With the increasing requirements for wafer size and precision in the integrated circuit industry, the stealth dicing process has become the mainstream choice in modern semiconductor processing because it can avoid the surface damage of wafers caused by traditional cutting techniques. This process forms a modified layer by focusing a laser beam at a specific depth inside the wafer, and microcracks or stress concentration areas are generated inside the material. When an external force acts on the wafer, these cracks will expand and eventually cause the wafer to separate, thus achieving the segmentation of the wafer. However, the realization of the stealth dicing process depends on the optimization of multiple process parameters. Due to the complex interaction between these process parameters and the possible influence of factors such as wafer material, thickness, and surface quality in the actual production process, how to effectively control and adjust these parameters to ensure cutting quality and yield has become an urgent problem in semiconductor processing.

[0003] The traditional debugging of the stealth dicing process usually relies on experienced engineers for repeated experiments and adjustments. In actual operation, engineers determine the appropriate process parameter window through trial and error. However, since the relationship between process parameters and cutting effects is often non-linear and difficult to predict, this experience-based adjustment method is not only inefficient but also may lead to long downtime and increased production costs.

[0004] In addition, the optimization process of the stealth dicing process usually requires a large number of processing equipment, measurement tools, and experimental time. Each parameter adjustment may involve a large number of tests and verifications, further exacerbating the time cost and resource consumption in the production process. Therefore, how to quickly and efficiently adjust and optimize process parameters while ensuring cutting quality has become a key issue in improving production efficiency, reducing costs, and enhancing product stability.

[0005] To achieve this goal, the traditional manual debugging method urgently needs to develop in a more intelligent and automated direction. In recent years, with the development of machine learning and artificial intelligence technologies, these emerging technologies have made breakthrough progress in many fields and have begun to be applied in semiconductor manufacturing. Currently, in the semiconductor manufacturing process, although there are already some process optimization methods based on machine learning, most of them still rely on a large amount of training data and show certain limitations under complex and changeable process conditions. Summary of the Invention

[0006] To solve the problems of parameter optimization in the invisible cutting process and the uncontrollable trial cutting time cost, the present invention provides a wafer cutting parameter optimization method and system, which can effectively cope with complex and variable wafer incoming materials and design requirements, and improve the production efficiency of proofing and the consistency of product quality.

[0007] The technical solution provided by the present invention is specifically as follows:

[0008] A wafer cutting parameter optimization method, comprising the steps of:

[0009] Obtain actual wafer cutting data, including wafer characteristic data, cutting process parameters and corresponding cutting result data; the wafer characteristic data includes the target thickness, safety distance and wafer thickness, the cutting process parameters include the laser mode, laser defocus and laser power, and the cutting result data includes the length of the modified region, the distance of the modified region from BCD and the cutting effect BHC;

[0010] Preprocess the collected data, and quantitatively score the cutting quality under the corresponding wafer characteristic data and cutting process parameters respectively according to the cutting result data;

[0011] Perform function fitting on the relationship between the laser defocus, laser power and the generated length of the modified region under different laser modes to obtain a fitting function for the length of the modified region; perform function fitting on the relationship between the distance of the modified region from BCD and the laser defocus to obtain a fitting function for the laser defocus;

[0012] Train multiple machine learning models respectively, the input of the machine learning model is the cutting process parameter, and the output of the machine learning model is the cutting quality score;

[0013] Determine the range of the modified region interval and the initial cutting process parameters according to the wafer characteristic data required by the project, and then optimize and adjust the initial cutting process parameters based on the machine learning model and the fitting function.

[0014] Further, the optimization and adjustment includes: inputting the initial cutting process parameters into multiple machine learning models to obtain the average cutting quality score under this parameter, if the cutting quality score meets the requirements, then use this parameter as the final cutting process parameter; otherwise, calculate the length of the modified region according to the fitting function of the length of the modified region under the current parameter, and based on this, adjust the defocus value in the parameter according to the fitting function of the laser defocus under the condition of meeting the range of the modified region interval, and input the adjusted parameter into the machine learning model to obtain the average cutting quality score until it meets the requirements.

[0015] Further, the cutting effect BHC reflects the overall quality of the wafer after cutting through the electron measurement photos of the wafer circuit surface and the electron measurement photos of the laser scattering distance. The scoring criteria for the quantitative scoring include the state of the wafer cutting marks.

[0016] Further, the modified region length fitting function is a polynomial function obtained by fitting based on the least squares method. The laser defocus fitting function is:

[0017] BCD down = T - ( n 2 + k) L ， BCD up = BCD down + h,

[0018] where BCD down represents the lowest height value of the modified region in the vertical direction, BCD up represents the maximum height value of the modified region in the vertical direction, h represents the length of the modified region, T represents the wafer thickness, n 2 represents the refractive index of the laser incident on the wafer, k is the laser refraction correction coefficient, L represents the laser defocus.

[0019] Further, the calculation formula for the range of the modified region interval is:

[0020] BCD up >De + 0.5D, BCD down <De - 0.5D,

[0021] where De represents the target thickness in the wafer characteristic data, and D represents the safety distance.

[0022] Further, the initial parameters of the cutting process are determined based on the K-means clustering algorithm, including the steps:

[0023] First, cluster the obtained cutting result data to obtain the cluster centers of each category:

[0024] S1. Select m groups of data in the cutting result data as the initial cluster centers. The dimensions of each sample point include the cutting quality score, the length h of the modified region, and the distance BCD of the modified region;

[0025] S2. Traverse each sample in the cutting result data, calculate its distances to the m initial cluster centers respectively, and assign it to the category corresponding to the cluster center with the minimum distance;

[0026] S3. Calculate the centroid of all samples included in each category respectively and use it as the clustering center of this category;

[0027] S4. Repeat S2 and S3 until the iteration termination condition is reached;

[0028] Then, calculate the Euclidean distance between the wafer characteristic data corresponding to the clustering center of each category and the wafer characteristic data required by the project, and use the cutting process parameters corresponding to the clustering center with the smallest Euclidean distance as the initial process parameters.

[0029] Further, the iteration termination condition is that the clustering centers of each category no longer change or the change error is less than the set threshold or the number of iterations reaches the set value; if the clustering center is not the existing sample data, select the sample data with the smallest distance from this clustering center as the substitution value.

[0030] Preferably, the machine learning model uses a support vector machine.

[0031] A wafer cutting parameter optimization system based on the above method, comprising:

[0032] A data acquisition and storage module, configured to acquire and store actual wafer cutting data, including wafer characteristic data, cutting process parameters, and corresponding cutting result data;

[0033] A data processing module, configured to preprocess the acquired data, and respectively quantify and score the cutting quality under the corresponding wafer characteristic data and cutting process parameters according to the cutting result data;

[0034] A function fitting module, configured to perform function fitting on the relationships between the laser off-focus, laser power, and the length of the modified region generated under different laser modes, and perform function fitting on the relationship between the distance of the modified region from the BCD and the laser off-focus;

[0035] A model training module, configured to train multiple machine learning models respectively, where the input of the machine learning model is the cutting process parameter, and the output of the machine learning model is the cutting quality score;

[0036] An initial parameter determination module, configured to determine the initial cutting process parameters according to the wafer characteristic data required by the project;

[0037] A parameter optimization module, configured to optimize and adjust the initial cutting process parameters according to the machine learning model and the fitting function under the condition of satisfying the modified region interval range.

[0038] Further, it further includes a model update module, configured to update and optimize the machine learning model by using the continuously acquired actual wafer cutting data.

[0039] Compared with the prior art, the present invention significantly improves the efficiency and accuracy of optimizing the parameters of the ultra-thin wafer stealth dicing process by introducing a machine learning model. The traditional optimization of stealth dicing process parameters mainly relies on the experience of engineers and a large number of trial-and-error processes, which has a long debugging cycle and high labor costs. However, the present invention can establish an effective model with relatively few data samples through small-sample learning, quickly capture the complex relationships between different process parameters, thus greatly reducing the number of experiments required and the optimization time.

[0040] Traditional process optimization methods usually cannot accurately identify the non-linearity and interactions between process parameters, resulting in unstable results during the optimization process and even in some cases being unable to find the best process window. This is because traditional optimization methods cannot effectively handle the interdependence between different parameters, and their debugging process is usually blind and exploratory, making it difficult to meet the complex process adjustment requirements. The present invention, on the other hand, can make effective predictions based on the existing small-sample data through a machine learning model. The model can learn the impact of each process parameter on the dicing effect from historical data and can dynamically optimize the parameter combination, avoiding the blindness of manual debugging.

[0041] In addition, the present invention can find a suitable process parameter window for newly introduced wafers or new materials in a relatively short time. In actual production, changes in wafer materials, sizes, and thicknesses may lead to a decrease in the effectiveness of process parameters under traditional methods, requiring a large number of re-adjustments and tests. Through machine learning, the system can automatically adjust parameters based on a small amount of test data and quickly adapt to the production requirements of new products, greatly improving the flexibility and production efficiency of the production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention.

[0043] Figure 1 is a schematic diagram of the parameters related to the modified area provided by an embodiment of the present invention;

[0044] Figure 2 is a schematic flow diagram of function fitting and model training provided by an embodiment of the present invention;

[0045] Figure 3 is a schematic flow diagram of process parameter optimization provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0047] Embodiment 1

[0048] This embodiment provides a method for optimizing wafer cutting parameters, including the steps of:

[0049] 1. Data collection

[0050] Conduct a DOE (Design of Experiments) test to collect the actual wafer cutting data set, including wafer characteristic data, cutting process parameters, and corresponding cutting result data.

[0051] The wafer characteristic data includes the target thickness De, the safety distance D, and the wafer thickness T. The target thickness De refers to the ideal thickness value of the wafer after cutting. The safety distance D refers to the thickness of the material that needs to be removed during the grinding stage of the wafer. The wafer thickness T refers to the thickness of the wafer before cutting. The cutting process parameters include the laser mode SC, the laser off-focus L, and the laser power P. The selection of the laser mode SC affects the cutting path and efficiency, such as the continuous mode or the pulse mode, etc. The off-focus L determines the relative distance between the laser beam and the material surface. The magnitude of the laser power determines the energy input during the cutting process, thereby affecting the cutting speed and quality.

[0052] The cutting result data includes the cutting effect BHC, the modified region distance BCD, and the modified region length h (the relevant parameters of the modified region are as Figure 1 shown), and these data are used to evaluate the cutting quality. Figure 1 In the figure, De1~De3 are the respective target thicknesses during multi-target thickness requirement cutting, D1~D3 are the safety distances corresponding to the respective target thicknesses, P is the laser power used during cutting, and d is the modified layer interval. The cutting effect BHC can be evaluated through the electron measurement photos of the wafer circuit surface and the electron measurement photos of the laser scattering distance, and is used to reflect the overall quality of the wafer after cutting, including defects such as cracks and chipping. By obtaining the photos of the wafer circuit surface after cutting through an electron microscope or other equipment, it can be used to analyze the surface morphology and defects of the circuit surface. The laser scattering distance reflects the scattering situation of the laser beam on the material surface and can indirectly evaluate the cutting depth and quality.

[0053] 2. Data processing

[0054] Preprocess the collected data, including operations such as data cleaning and normalization, to remove outliers and noise data, and convert data with different dimensions into a unified scale for subsequent analysis and calculation.

[0055] Use statistical analysis methods (such as analysis of variance, regression analysis, etc.) to identify key process parameters that are highly correlated with product quality. For example, the laser power, defocus position, and laser mode have a significant impact on the cutting quality. At the same time, quantify the features in the electron measurement photos of the wafer circuit surface and the electron measurement photos of the laser scattering distance, that is, quantify the BHC from 1 to 10 according to the cutting quality on the front side of the wafer. The scoring criteria include whether the cutting marks are separated, whether there are serpentine bifurcations, etc. Generally speaking, the cutting score gradually increases from the cutting marks that are never separated, have serpentine bifurcations, have a small number of serpentine bifurcations to a straight line. Evaluate the cutting quality under each process parameter through data scoring and use it for subsequent machine learning model training, as Figure 2 shown.

[0056] 3. Function fitting

[0057] Perform function fitting on the relationship between the cutting result data of the wafer and the cutting process parameter data.

[0058] Different laser heads and their working modes generate different shapes of the modified region. Therefore, the collected data can be used to perform function fitting on this corresponding relationship, that is:

[0059] Fitting function 1:

[0060] Among them, the defocus L and the laser power P are both input process parameters of the laser cutting equipment, h is the length of the modified region. SC is the laser mode, which is a beam mode with fixed optical path characteristics and cannot be quantified. f is the function expression obtained by fitting, and its specific form depends on the specific application scenario. For example:

[0061]

[0062] , , …, are fitting coefficients. The accuracy of the fitting polynomial function can meet the engineering requirements, and algorithms such as the least squares method can be used for the fitting method.

[0063] During laser cutting, the light beam is incident from the gaseous medium onto the back grinding surface of the wafer, and its projection angle can be calculated according to the law of refraction of light: , where n 1 and n2 are the refractive indices of the incident and refracting media respectively, θ 1 and θ 2 are the angle of incidence and the angle of refraction respectively.

[0064] In stealth dicing, the laser processes the wafer to generate single or multiple modified regions. Cracks in these modified regions will expand along the direction perpendicular to the laser focus direction, eventually causing the wafer to separate. To estimate the distance BCD of the modified region, fitting can be performed using measured values:

[0065] Fitting function 2: BCD down = T - ( n 2 + k) L ， BCD up = BCD down + h

[0066] where k is the laser refraction correction coefficient, which is usually obtained through experimental calibration and parameter fitting. Due to the absorption, reflection, scattering, etc. losses of the wafer to the laser, and the refractive index difference will change the actual focal position, which makes the geometric shape (such as depth, width) of the modified region deviate from the theoretical calculation. The role of k is to correct the errors caused by these optical effects through experience. BCD down represents the lowest height value of the modified region in the vertical direction, and BCD up represents the highest height value of the modified region in the vertical direction.

[0067] 4. Model training

[0068] Since the amount of data that can be generated by each DOE experiment is limited, appropriate machine learning algorithms must be used. Support Vector Machine (SVM) is a supervised learning method, which is particularly suitable for small sample, non-linear and high-dimensional pattern recognition problems. Using a small amount of historical data to train multiple groups of SVM models respectively, taking the dicing process parameters as the model input and the dicing result data (BHC score) as the output, so that the trained SVM models can predict the wafer dicing state under a given combination of process parameters.

[0069] 5. Process parameter optimization

[0070] Optimize the dicing process parameters through the trained multiple groups of machine learning models in combination with the fitting function.

[0071] 5.1. Determine the modified region interval

[0072] To ensure that the ground wafer after dicing will not grind to the modified region, the modified region needs to avoid the target thickness position and leave a certain safety distance. That is: BCDup >De + 0.5D, BCD down <De - 0.5D.

[0073] 5.2. Determine the initial values of process parameters

[0074] According to the wafer characteristic data (target thickness De and safety distance D) required by the project, use the K-means algorithm to find the matching parameters in the existing cutting process parameters (obtained through DOE experiments) as the initial process parameters, and exact matching is not required.

[0075] Specifically, first cluster the cutting result data obtained from the DOE experiment to obtain the cluster centers of each category;

[0076] (1) Select m groups of data in the cutting result data as the initial cluster centers, and the dimensions of each sample point include BHC score, distance BCD from the modified area, and length h of the modified area;

[0077] (2) Traverse each sample in the cutting result data, calculate its distances to the m initial cluster centers respectively, and assign it to the category corresponding to the cluster center with the minimum distance;

[0078] (3) Calculate the centroid of all samples included in each category and use it as the cluster center of this category;

[0079] (4) Repeat the above two steps (2) and (3) until the termination condition is reached, such as the cluster center no longer changes, or the change error is less than the set threshold, or the number of iterations reaches a certain value.

[0080] Then, calculate the Euclidean distances between the wafer characteristic data corresponding to the cluster centers of each category and the wafer characteristic data required by the project respectively, and use the cutting process parameters corresponding to the cluster center with the minimum Euclidean distance as the initial process parameters. If the cluster center is not the sample data existing in the DOE experiment, select the sample with the minimum distance from this cluster center as the substitute value.

[0081] 5.3. Optimize process parameters

[0082] In order to meet the requirements of multiple chip thicknesses, use machine learning models and fitting functions to optimize the initial process parameters. That is, input the initial values of different process parameters and output the optimized and adjusted process parameters based on the initial values.

[0083] Specifically, first input the initial process parameters into multiple trained machine learning models to obtain the average BHC score under this process parameter. If the average BHC score meets the requirements, use this process parameter as the final process parameter. If the average BHC score does not meet the requirements, then within the range of the modified area interval (BCD up>De + 0.5D, BCD down When <De - 0.5D), adjust the defocus in the process parameters according to fitting function 1 and fitting function 2 L value, and input the adjusted process parameters into the machine learning model to obtain the BHC score until the requirements are met. The optimization process of the entire process parameters is as Figure 3 shown.

[0084] 6. Model Update

[0085] By continuously collecting the latest data in the production process, the machine learning model and the actual wafer cutting data set are continuously updated to adapt to the changes in process conditions and improve the robustness and adaptability of the model.

[0086] Example Two

[0087] Based on the above method, this example provides a wafer cutting parameter optimization system, which mainly includes the following modules:

[0088] Data acquisition and storage module, used to obtain and store actual wafer cutting data, including wafer characteristic data, cutting process parameters, and corresponding cutting result data;

[0089] Data processing module, used to preprocess the collected data and quantitatively score the cutting quality under the corresponding wafer characteristic data and cutting process parameters according to the cutting result data;

[0090] Function fitting module, used to fit the relationship between laser defocus, laser power, and the length of the modified area generated under different laser modes, and fit the relationship between the distance of the modified area from BCD and the laser defocus;

[0091] Model training module, used to train multiple machine learning models respectively. The input of the machine learning model is the cutting process parameter, and the output of the machine learning model is the cutting quality score; for large-scale and high-dimensional data, a deep neural network model is preferably used, while for medium and small-scale data sets, traditional machine learning algorithms such as random forest or support vector machine can be selected.

[0092] Initial parameter determination module, used to determine the initial cutting process parameters according to the wafer characteristic data required by the project;

[0093] Parameter optimization module, used to optimize and adjust the initial cutting process parameters according to the machine learning model and fitting function under the condition of meeting the modified area interval range.

[0094] In some embodiments, the system further includes a model update module, used to update and optimize the machine learning model by using the continuously collected actual wafer cutting data.

[0095] The above system can execute the wafer cutting parameter optimization method described in Embodiment 1, and has the corresponding functional modules and beneficial effects of the method. For the technical details not described in detail in this embodiment, reference can be made to the wafer cutting parameter optimization method provided in Embodiment 1 of the present invention.

[0096] In summary, compared with the traditional experience-based process parameter optimization scheme, the present invention can use machine learning algorithms to analyze the relationships between complex process parameters, shorten the optimization time, and improve the process optimization efficiency; by optimizing the process parameter combination, it can maintain a high level of product quality within a wide range of process parameter windows and reduce the defective product rate. Adaptive optimization is achieved through the update and iteration of new models, reducing human intervention and lowering the input of manpower and time.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; under the concept of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and the steps can be implemented in any order, and there are many other variations in different aspects of the present invention as described above. For the sake of brevity, they are not provided in detail; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or equivalently replace some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present application.

Claims

1. A method for optimizing wafer cutting parameters, characterized in that: Includes steps: Obtain actual wafer cutting data, including wafer characteristic data, cutting process parameters and corresponding cutting result data; wafer characteristic data includes target thickness, safety distance and wafer thickness, cutting process parameters include laser mode, laser defocus point and laser power, and cutting result data includes modified area length, modified area distance BCD and cutting effect BHC; Pre-process the collected data, and quantify the scores of the corresponding wafer characteristic data and the cutting quality under the cutting process parameters according to the cutting result data; The relationship between the laser defocus point, laser power and the length of the modified area generated under different laser modes is fitted with a function to obtain a modified area length fitting function; the relationship between the modified area distance BCD and the laser defocus point is fitted with a function to obtain a laser defocus point fitting function; The modified region length fitting function is: Among them, the out-of-focus L and laser power P These are the input process parameters of the laser cutting equipment. h is the length of the modified region, f is a polynomial function obtained based on the least squares fitting method, and its specific form depends on different laser modes; The laser defocus point fitting function is: BCD down = T - ( n 2 + k) L ， BCD up = BCD down + h, Among them, BCD down Indicates the lowest vertical height of the modified area, BCD up represents the maximum height of the modified area in the vertical direction, h represents the length of the modified area, T Indicates the wafer thickness, n 2 represents the refractive index of the laser incident on the wafer, and k is the laser refraction correction coefficient; Training a plurality of machine learning models respectively, wherein the input of the machine learning model is a cutting process parameter, and the output of the machine learning model is a cutting quality score; Determine the modified area range and initial cutting process parameters based on the wafer characteristic data required for the project, and then optimize and adjust the initial cutting process parameters based on the machine learning model and fitting function; The optimization adjustment includes: inputting the initial parameters of the cutting process into multiple machine learning models to obtain the average cutting quality score under the parameters. If the cutting quality score meets the requirements, the parameters are used as the final cutting process parameters; otherwise, the length of the modified area is calculated according to the modified area length fitting function under the current parameters, and based on this, the defocus value in the parameter is adjusted according to the laser defocus fitting function while satisfying the modified area interval range, and the adjusted parameters are input into the machine learning model to obtain the average cutting quality score until it meets the requirements.

2. The wafer cutting parameter optimization method according to claim 1, characterized in that: The cutting effect BHC reflects the overall quality of the wafer after cutting through electronic measurement photos of the wafer circuit surface and electronic measurement photos of the laser scattering distance. The scoring criteria of the quantitative scoring include the state of the wafer cutting marks.

3. The wafer cutting parameter optimization method according to claim 1, characterized in that: The calculation formula of the modified area interval range is: BCD up > De + 0.5D, BCD down < De - 0.5D, where De represents the target thickness in the wafer characteristic data and D represents the safety distance.

4. The wafer cutting parameter optimization method according to claim 1, characterized in that: The initial parameters of the cutting process are determined based on the K-means clustering algorithm, including the steps of: First, cluster the obtained cutting result data to obtain the cluster centers of each category: S1. Select m groups of data from the cutting result data as the initial cluster centers, and the dimensions of each sample point include the cutting quality score, the length of the modified area h, and the distance BCD of the modified area; S2. Traverse each sample in the cutting result data, calculate its distance to the m initial cluster centers respectively, and divide it into the category corresponding to the cluster center with the smallest distance; S3. Calculate the centroid of all samples contained in each category and use it as the cluster center of the category; S4. Repeat S2 and S3 until the iteration termination condition is reached; Then, the Euclidean distances between the wafer characteristic data corresponding to the cluster centers of each category and the wafer characteristic data required for the project are calculated respectively, and the cutting process parameters corresponding to the cluster center with the smallest Euclidean distance are used as the initial process parameters.

5. The wafer cutting parameter optimization method according to claim 4, characterized in that: The iteration termination condition is that the cluster center of each category no longer changes or the change error is less than a set threshold or the number of iterations reaches a set value; if the cluster center is not an existing sample data, the sample data with the smallest distance to the cluster center is selected as a replacement value.

6. The wafer cutting parameter optimization method according to claim 1, characterized in that: The machine learning model adopts support vector machine.

7. A wafer cutting parameter optimization system based on the method according to any one of claims 1 to 6, characterized in that: include: A data acquisition and storage module is used to acquire and store actual wafer cutting data, including wafer characteristic data, cutting process parameters and corresponding cutting result data; The data processing module is used to pre-process the collected data and quantify the scores of the corresponding wafer characteristic data and the cutting quality under the cutting process parameters according to the cutting result data; Function fitting module, used for performing function fitting on the relationship between laser defocus point, laser power and length of the generated modified area under different laser modes, and performing function fitting on the relationship between modified area distance BCD and laser defocus point; A model training module, used to train multiple machine learning models respectively, wherein the input of the machine learning model is a cutting process parameter, and the output of the machine learning model is a cutting quality score; An initial parameter determination module is used to determine the initial parameters of the cutting process according to the wafer characteristic data required by the project; The parameter optimization module is used to optimize and adjust the initial parameters of the cutting process according to the machine learning model and fitting function while meeting the range of the modified area.

8. The wafer cutting parameter optimization system according to claim 7, characterized in that: It also includes a model update module, which is used to update and optimize the machine learning model using the actual wafer cutting data that is continuously collected.

Images