Laser cutting protection liquid performance prediction and optimization method based on machine learning

Abstract

The invention provides a laser cutting protection liquid performance prediction and optimization method based on machine learning, and the method comprises the steps: constructing a multi-dimensional feature data set of a laser cutting protection liquid, and the data set comprises molecular descriptors of formula components and cross-scale fusion features of spin-coating process parameters; performing feature selection on the data set to compress the number of features, and screening key features related to viscosity; based on the screened data set, adopting a machine learning model to train a viscosity prediction model; based on the viscosity predicted value of the model and the spin-coating rotating speed, the film forming thickness is calculated through a Meyerhofer model, and the film forming thickness is in direct proportion to the cubic root of the viscosity and in inverse proportion to the square root of the spin-coating rotating speed; and optimizing the formula of the protection liquid according to the film forming thickness.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of semiconductor processing and machine learning, specifically to a method for predicting and optimizing the performance of laser cutting protective fluid by combining experimental data, theoretical models, and machine learning algorithms. The optimized protective fluid formulation is then applied to actual laser cutting processes. This invention is applicable to semiconductor wafer cutting. The aim is to accurately predict the viscosity of the protective fluid and, by combining this prediction with the corresponding rotation speed and the Meyerhofer theoretical model, determine the corresponding protective fluid film thickness. This improves the cutting accuracy of the protective fluid, reduces thermal damage, and ensures the integrity of the wafer surface after cutting. Background Technology

[0002] As chip technology advances towards higher precision, chip packaging is becoming increasingly miniaturized and thinner, with wafer thickness shrinking to the tens of micrometer level and linewidth continuously breaking through to the nanometer level. This places stringent demands on the precision and stability of subsequent processing technologies. Due to the significant decrease in mechanical strength, coupled with the high porosity and low flexural strength of the internal low-k dielectric material, ultrathin wafers are highly susceptible to cracking or delamination under external forces during the dicing process, making the challenges faced by traditional processing technologies increasingly prominent.

[0003] Traditional diamond blade cutting relies on mechanical contact to separate materials. The mechanical stress generated during the cutting process is transmitted into the wafer interior. For ultra-thin wafers, this stress can easily cause edge chipping and breakage, and may also lead to the separation of the Low-K dielectric layer from the metal wiring, severely affecting chip yield. In contrast, laser cutting, with its non-contact processing characteristics, can effectively reduce the damage caused by mechanical stress. Its advantages of narrow kerf (down to a few micrometers) and high cutting speed have made it a mainstream technology in advanced packaging and ultra-thin wafer processing, especially suitable for high-precision cutting of 12-inch and larger wafers. Before laser cutting, spin-coating a water-soluble protective liquid onto the wafer surface is a critical step. This protective liquid, after film formation, must have good laser energy absorption capabilities (especially for commonly used ultraviolet laser wavelengths) to reduce thermal damage to the wafer substrate by the laser, while blocking debris generated during the cutting process to avoid secondary contamination. Its water-soluble properties ensure that it can be easily removed by cleaning after cutting without affecting subsequent processes.

[0004] The performance of laser cutting protective fluid directly affects cutting quality and wafer yield, yet its formulation development has long faced multiple bottlenecks. First, there is the problem of high-dimensional combinatorial explosion. Protective fluids are typically composed of water-soluble resins, organic solvents, and functional additives (such as plasticizers and defoamers). Variables such as the molecular weight of the resin, the type and proportion of the solvent alone can create hundreds or even thousands of combinations. Traditional trial-and-error methods for screening these combinations are not only time-consuming but also fail to cover all potential optimal formulations. Second, there are significant nonlinear effects. The viscosity of the protective fluid is not only related to the proportions of its components but is also affected by environmental factors such as temperature and shear rate. Furthermore, film thickness is the result of the combined effects of viscosity, spin coating speed, time, and even ambient humidity. The interactions between these variables exhibit complex nonlinear relationships that are difficult to describe using simple linear models. Furthermore, the ambiguity of the micro-to-macro correlation exacerbates the difficulty of formula optimization. The microscopic features of molecules in the protective liquid, such as their geometry, electronic distribution, and topology, affect the macroscopic viscosity and film-forming properties through intermolecular forces (such as hydrogen bonds and van der Waals forces). However, this transfer mechanism from the molecular scale to macroscopic performance is not yet fully understood, making it difficult to precisely control the performance of the protective liquid at the molecular design level.

[0005] Traditional experimental methods rely on accumulated experience, requiring a complete process of preparation, testing, and analysis for each formulation adjustment. This often results in development cycles lasting weeks or even months, and struggles to overcome the limitations of local optima. Existing theoretical models, such as the Meyerhofer equation describing spin-coating film formation, while relating viscosity, rotation speed, and film thickness to some extent, are based on ideal fluid assumptions and fail to consider practical factors such as component evaporation, phase separation, and drying shrinkage during spin-coating. This leads to limited accuracy in predicting film thickness for complex formulations of protective solutions, making it difficult to directly guide formulation optimization. Therefore, overcoming the limitations of existing technologies and efficiently achieving performance control and formulation optimization of laser cutting protective solutions has become a pressing issue in the semiconductor manufacturing field. Summary of the Invention

[0006] To address the shortcomings and deficiencies of existing technologies, this invention provides a machine learning-based method for predicting and optimizing the performance of laser cutting protective fluids and its application. This method aims to solve problems such as low experimental efficiency, difficulty in modeling nonlinear relationships, and unclear micro-macro correlations in traditional protective fluid formulation development. This method collects formulation components, viscosity, film thickness, and spin-coating parameters of 139 PVA and PVP protective liquids to construct a cross-scale fusion dataset containing formulation component ratios, water-soluble resin molecular weights, molecular descriptors (covering geometric, electronic, and other multi-dimensional features), and spin-coating process parameters. Feature selection retains eight key features significantly related to viscosity. Based on the selected dataset, multiple machine learning models (with gradient boosting decision tree model being optimal, its test set determination coefficient not less than 0.99) are used to train a viscosity prediction model, and hyperparameter optimization improves prediction accuracy. The viscosity prediction value output by the model and the spin-coating speed are input into the Meyerhofer model (film thickness is directly proportional to the cube root of viscosity and inversely proportional to the square root of spin-coating speed, with appropriate empirical constants fitted for PVA and PVP protective liquids respectively) to calculate the film thickness. Finally, the protective liquid formulation is optimized based on the predicted film thickness, especially by adjusting the proportion and molecular weight of water-soluble resins to improve film thickness uniformity. This method can be applied to laser cutting processes in semiconductor manufacturing, photovoltaic cell processing, and other fields, providing an efficient technical means for the intelligent design and performance optimization of protective liquid formulations.

[0007] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0008] A machine learning-based method for predicting and optimizing the performance of laser cutting protective fluid, comprising:

[0009] A multidimensional feature dataset for laser cutting protective liquid is constructed, which includes molecular descriptors of formulation components and cross-scale fusion features of spin coating process parameters;

[0010] Feature selection is performed on the dataset to reduce the number of features and filter out key features related to viscosity;

[0011] Based on the filtered dataset, a viscosity prediction model was trained using a machine learning model.

[0012] Based on the viscosity prediction value and spin coating speed of the model, the film thickness is calculated using the Meyerhofer model, where the film thickness is directly proportional to the cube root of the viscosity and inversely proportional to the square root of the spin coating speed.

[0013] The protective solution formulation is optimized based on the film thickness.

[0014] Furthermore, the formulation components include water-soluble resin, organic solvent, plasticizer, defoamer, ultraviolet absorber, antioxidant and water, wherein the water-soluble resin is PVA type or PVP type.

[0015] Furthermore, the molecular descriptors are obtained by converting the molecular structure of the formulation components into the standard SMILES format using ChemDraw, and then calculating multiple molecular descriptors using RDKit, covering geometric, electronic, topological features and spatial configuration; the multidimensional feature dataset also includes the proportion of the formulation components and the molecular weight of the water-soluble resin.

[0016] Furthermore, the feature selection employs one or more combined methods among low variance filtering, correlation analysis, and PCA dimensionality reduction, ultimately identifying eight key features significantly related to viscosity.

[0017] Furthermore, the machine learning model includes a decision tree ensemble model, which is selected from gradient boosting decision tree, random forest, XGBoost or LightGBM, wherein the optimal model is the gradient boosting decision tree model, and the gradient boosting decision tree model has a determination coefficient R² ≥ 0.99 on the test set.

[0018] Furthermore, when training the viscosity prediction model, the selected dataset is divided into a training set and a test set. A nested K-fold cross-validation strategy is adopted, and the hyperparameters are optimized through grid search. The hyperparameters include the learning rate, the number of trees, and the maximum depth.

[0019] Furthermore, in the Meyerhofer model, corresponding empirical constants k are fitted for PVA-type and PVP-type protective liquids respectively, in order to adapt to the film-forming characteristics of different types of protective liquids.

[0020] Furthermore, the optimization of the protective liquid formulation based on film thickness includes: adjusting the proportion and molecular weight of water-soluble resin, wherein increasing the proportion of water-soluble resin is used to improve film thickness uniformity, so that the predicted film thickness meets the requirements of laser cutting for film thickness uniformity and cutting accuracy.

[0021] And, a machine learning-based system for predicting and optimizing the performance of laser cutting protective fluid, comprising:

[0022] The data acquisition module is used to construct a multidimensional feature dataset of the laser cutting protective liquid, which includes molecular descriptors of the formulation components and cross-scale fusion features of spin coating process parameters.

[0023] The feature processing module is used to perform feature selection on the dataset to compress the number of features and filter out key features related to viscosity.

[0024] The model training module is used to train a viscosity prediction model using a machine learning model based on the filtered dataset.

[0025] The film thickness calculation module is used to calculate the film thickness based on the viscosity prediction value and spin coating speed of the model using the Meyerhofer model, wherein the film thickness is directly proportional to the cube root of the viscosity and inversely proportional to the square root of the spin coating speed.

[0026] The formulation optimization module is used to optimize the protective liquid formulation based on the film thickness.

[0027] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0028] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0029] Compared to existing technologies, this invention and its preferred solution effectively reduce the computational complexity caused by high-dimensional data by constructing a multidimensional dataset that integrates the microscopic features of the formulation components with the macroscopic parameters of the spin coating process, and by combining feature selection technology. This provides a foundation for accurately capturing the correlation between the performance of the protective fluid and the formulation components. Through the training and optimization of various machine learning models, especially the optimized decision tree ensemble model, high-precision prediction of the protective fluid viscosity is achieved, overcoming the limitation of traditional linear models in describing nonlinear relationships, reducing reliance on empirical trial and error, and shortening the formulation development cycle. By combining the viscosity predicted by machine learning with the Meyerhofer model and fitting appropriate empirical parameters for different types of protective fluids, the targeting and accuracy of film thickness prediction are improved. This helps to deduce the optimal formulation from film thickness requirements and achieve targeted control of the protective fluid performance. The overall solution, through a closed-loop design of "data acquisition - model prediction - formulation optimization," correlates microscopic molecular features with macroscopic process performance, providing an efficient and reliable technical means for the intelligent formulation design of laser cutting protective fluids. It is applicable to multiple precision processing fields such as semiconductor manufacturing, and helps improve cutting quality and process stability. Attached Figure Description

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0031] Figure 1 This is a flowchart illustrating the implementation of an embodiment of the present invention;

[0032] Figure 2 This is a comparison chart of the predicted values ​​and experimental values ​​of the dataset by the best model GBDT in this embodiment of the invention;

[0033] Figure 3 This is a viscosity characteristic importance analysis diagram of the optimal model GBDT in the embodiments of the present invention;

[0034] Figure 4 These are fitting diagrams of the Meyerhofer film-forming theoretical models for PVA (left) and PVP (right) type protective liquids;

[0035] Figure 5 The image shows the cutability test results after the protective liquid of Application Example 1 (left) and Comparative Example 1 (right) of the present invention was applied and formed into a film. Detailed Implementation

[0036] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0037] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0038] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0039] To overcome the shortcomings of existing technologies, this invention provides a machine learning-based method for predicting and optimizing the performance of laser cutting protective fluids. This method aims to solve the technical challenges of long experimental cycles, high costs, and the difficulty in establishing the correlation between microscopic molecular structure and macroscopic properties in traditional formulation development. By integrating experimental data, theoretical models, and machine learning algorithms, this method constructs a high-precision model for predicting the viscosity performance of protective fluids. Combined with the Meyerhofer equation theoretical model, it achieves accurate prediction of the viscosity and film thickness of the protective fluid, which can be used for intelligent design and optimization of formulations. It is applicable to the field of laser cutting of semiconductor wafers.

[0040] This invention provides a comprehensive approach that integrates experimental data, theoretical models, and machine learning algorithms. It addresses challenges in developing protective fluid formulations for laser cutting, such as high-dimensional combinatorial explosion, nonlinear effects, and unclear micro-macro correlations. By constructing a multi-dimensional feature space and combining machine learning algorithms with theoretical correction models, this invention achieves accurate prediction of the protective fluid viscosity and film thickness, providing efficient and reliable technical support for laser cutting processes in semiconductor manufacturing.

[0041] This method experimentally measures the viscosity, film thickness, and spin-coating parameters of protective liquids with different formulations, collecting data in the process. A dataset of protective liquid formulation components and viscosities, containing 139 data points, is constructed. Molecular descriptors and process parameters of the formulation components are extracted, a multidimensional feature space is built, and feature selection is performed to obtain a filtered dataset. The dataset is then partitioned, and various machine learning algorithms are used for modeling. Hyperparameters are optimized through nested K-fold cross-validation. Ultimately, the GBDT model performs excellently on the test set (R0). 2 =0.9916, MSE=0.1003*10 -3 This invention demonstrates that the model can accurately predict the viscosity of the protective fluid by predicting formulations with unknown viscosities, and can be used for the optimized design of protective fluid formulations. Based on the Meyerhofer model, the relationship between viscosity, rotation speed, and film thickness is established to predict the corresponding film thickness. The optimized protective fluid formulation is applied to an actual laser cutting process to verify its performance. This invention provides an efficient and reliable technical means for the formulation design and process optimization of laser cutting protective fluids.

[0042] like Figure 1 As shown, the specific technical solution for implementing the present invention includes the following steps:

[0043] Step 1: Collect experimental data on laser cutting protective solutions, including the formulation composition, viscosity, and film thickness. A total of 139 different formulations of laser cutting protective solutions were prepared, including PVA and PVP types. The viscosity and film thickness of each protective solution were measured, and spin-coating parameters (rotation speed, time) and film thickness data were recorded. The original experimental data underwent repeatability testing and outlier removal to ensure data quality.

[0044] Step 2: Construct a dataset of protective fluid formulation components and viscosity for feature selection to obtain a filtered dataset. A dataset of 139 data points was constructed based on the collected laser cutting protective fluid formulation components and their corresponding viscosities. Feature selection was then performed. First, the feature parameters of the dataset were calculated, including the proportions of the components used in the laser cutting protective fluid and the molecular weight (MwA) of the water-soluble resin. Molecular descriptors were extracted using computational chemistry methods. ChemDraw was used to convert the molecular structures to the standard SMILES format. Each molecular structure was used by RDKit to calculate 210 molecular descriptors, covering multidimensional features including geometric, electronic, topological, and spatial configuration features. Each dataset ultimately contained 427 features. Feature selection was then performed, using methods including but not limited to one or more combinations of low-variance filtering, correlation analysis, and PCA dimensionality reduction. Finally, eight key features significantly related to viscosity were retained through screening.

[0045] Step 3: Split the dataset and use multiple machine learning algorithms for modeling and hyperparameter optimization during training. The dataset is divided into a training set (80%) and a test set (20%). Different machine learning algorithms are used to build models for each set. In this invention, two linear regression models are constructed: L1-regularized Lasso regression and L2-regularized Ridge regression. Four ensemble models based on decision trees are also constructed: Gradient Boosting Decision Tree (GBDT), Random Forest (RF), XGBoost (XGB), and LightGBM (LGBM). A nested 5-fold cross-validation strategy is used. The outer loop evaluates the model's generalization ability through data splitting and obtains the average evaluation result. The inner loop optimizes the hyperparameter space through grid search. Optimization parameters include: learning rate; number of trees (n_estimators); and maximum tree depth (max_depth).

[0046] Step 4: Evaluate the predictive performance of the model, select the optimal model, and perform predictions for the optimized design of the protective fluid formulation. This is done by using the coefficient of determination R0. 2 The model performance was evaluated using mean squared error (MSE), and the model prediction mechanism was explained through feature importance analysis. The optimal model was selected to predict the viscosity of unknown laser cutting protective fluid formulations, thereby helping to improve and optimize the performance of the protective fluid.

[0047] Among them, the coefficient of determination R 2 The mean square error (MSE) is calculated as follows:

[0048]

[0049]

[0050] y i Represents the actual value, y' i This represents the model's predicted value, and n represents the number of samples.

[0051] Step 5: Predict the viscosity of the formulation for unknown viscosity, which will be used for the optimization design of the protective fluid formulation. The new protective fluid formulation is reconfigured, and its viscosity is measured to collect viscosity data. Based on the predictive power of the optimal model, the viscosity of the new protective fluid formulation combination is predicted. By comparing the predicted results with experimental results, the generalization ability of the model is analyzed.

[0052] Step 6: Establish the relationship between viscosity, rotation speed, and film thickness based on a theoretical model to predict the corresponding film thickness. The relationship between viscosity, rotation speed, and film thickness is established based on the Meyerhofer model, and the film thickness data is theoretically corrected. By fitting the Meyerhofer model, empirical formulas for PVA and PVP protective fluids are derived, and the film thickness is calculated. The theoretical formula of the Meyerhofer model used is as follows:

[0053]

[0054] Among them, h f υ is the film thickness (in nm), and υ is the kinematic viscosity (in mm). 2 / s), where ω is the spin coating speed (unit: rpm);

[0055] Step 7: Apply the optimized protective solution formulation to the actual laser cutting process to verify its performance. The viscosity of the unknown protective solution is predicted using the optimal viscosity model, and the corresponding film thickness is calculated using the Meyerhofer theoretical model. The unknown protective solution is spin-coated onto the wafer surface at the corresponding rotation speed. After film formation, it is placed in a 7.2W laser cutting machine with a frequency of 1000kHz, a speed of 100mm / s, a wavelength of 355nm, and two blades. After cutting, the cutting effect is observed and recorded under a scanning electron microscope (SEM).

[0056] The above solutions are applicable to PVA and PVP type laser cutting protective fluids. They can predict the viscosity by analyzing the composition of the protective fluid, and then predict the film thickness of the protective fluid by combining it with the Meyerhofer theoretical model. The optimized protective fluid formulation can then be applied to the actual laser cutting process, assisting in the optimization of protective fluid performance development.

[0057] The predictive model of this invention can accurately predict the viscosity and film-forming properties of laser cutting protective fluids under different formulation compositions, component concentrations, and process parameters with a high level of precision. This method can be applied to the formulation design and process optimization of protective fluids in semiconductor manufacturing, photovoltaic cell processing, and precision electronic component cutting.

[0058] Compared with the prior art, the present invention has the following advantages:

[0059] 1. Because the performance of laser cutting protective fluid is affected by multiple factors, traditional linear models are insufficient to accurately describe it. This invention employs the construction of multiple machine learning models to achieve high-precision prediction of the viscosity of laser cutting protective fluid. The optimal model's prediction model R for the dataset is... 2 The value reached 0.9996, and the MSE was 0.1003*10. -3 mN / m. Compared to traditional laser cutting protective fluid development, which relies on extensive experimental trial and error, resulting in high costs and low efficiency, this method can quickly predict the performance of unknown formulations, significantly reducing the number of experiments and shortening the R&D cycle.

[0060] 2. This invention improves prediction accuracy by combining the theoretical model with experimental data. It innovatively combines the Meyerhofer theoretical model with machine learning prediction results, using viscosity prediction to further derive film thickness at different rotational speeds. Furthermore, by applying the optimized formulation to a practical laser cutting process, its cutting effect is verified, demonstrating the practicality and reliability of this invention and providing a solid technical foundation for industrial applications.

[0061] Based on the design scheme provided above by the present invention, the overall scheme design will be further demonstrated and introduced through specific embodiments:

[0062] The method for constructing the prediction model is shown below:

[0063] Experimental data on laser cutting protective fluids were collected, including formulation composition, viscosity, and film thickness. Different formulations of laser cutting protective fluids were prepared experimentally, including PVA and PVP types, totaling 139 types (59 PVA types and 80 PVP types). The formulation design and preparation of the protective fluids are shown in Table 1 below.

[0064] Table 1. Components of Laser Cutting Protective Fluid Formulation

[0065]

[0066] To prepare a uniform film, spin coating first requires spreading the liquid at a low speed (50 rpm, 60 s), then accelerating to a high speed within 1 second. High-speed spin coating is set at five levels (1000 rpm, 1500 rpm, 2000 rpm, 2500 rpm, 3000 rpm, 300 s) to ensure uniform spreading and rapid drying. The viscosity and film thickness of each protective liquid are measured, and spin coating parameters (speed, time) and film thickness data are recorded. The original experimental data are subjected to repeatability checks and outlier removal to ensure data quality.

[0067] Feature selection was performed by constructing a dataset of protective fluid formulation components and viscosities: This embodiment constructed a dataset containing 139 data points based on the collected protective fluid formulation components and corresponding viscosities. Feature parameters of the dataset were calculated, including the proportion of AF (air-based polymerase) for each formulation and the molecular weight (MwA) of the water-soluble resin. A and C were used as variables, and computational chemistry methods were employed to extract molecular descriptors. First, ChemDraw was used to convert the molecular structures to the standard SMILES format. Each molecular structure was calculated using RDKit, yielding 210 molecular descriptors covering geometric, electronic, topological, and spatial configuration features. Each dataset ultimately contained 427 features. Feature selection was then performed, using methods including, but not limited to, low-variance filtering, correlation analysis, and PCA dimensionality reduction, or a combination thereof. Finally, eight key features significantly correlated with viscosity were retained, greatly optimizing the feature space.

[0068] The dataset was divided into training (80%) and test (20%) sets. Different machine learning algorithms were used to build models for each set. This system constructed two linear regression models: L1-regularized Lasso regression and L2-regularized Ridge regression. Four ensemble models based on decision trees were also constructed: Gradient Boosting Decision Tree (GBDT), Random Forest (RF), XGBoost (XGB), and LightGBM (LGBM). A 5-fold cross-validation strategy was adopted. The outer loop evaluated the generalization ability of the model through data splitting, while the inner loop optimized the hyperparameter space, including the learning rate, the number of trees, and the maximum depth, through grid search.

[0069] To evaluate the predictive performance of the models and select the optimal model, the performance metrics for this model include mean squared error (MSE) and coefficient of determination (R²). 2 R 2 The better the model's ability to explain the variance of the target variable, the closer it is to 1. The better the model's ability to explain the variance of the target variable, the closer it is to 0, the closer it is to the expected value.

[0070] Wherein, the determination coefficient R 2 The mean square error (MSE) is calculated as follows:

[0071]

[0072]

[0073] Among them, y i Represents the actual value, y' i This represents the model's predicted value, and n represents the number of samples.

[0074] Using the coefficient of determination R2 The model results are evaluated and visualized using the mean squared error (MSE) (e.g.) Figure 2 As shown in Table 2 below, it can be seen that the dataset in GBDT (test set R) 2 =0.9916, MSE=0.1003*10 -3 The best performing model is GBDT, which can be selected as the optimal model to predict the viscosity of laser cutting protective fluid at unknown concentrations and ratios.

[0075] Table 2. Model Evaluation and Comparison of Datasets

[0076]

[0077] The model prediction mechanism (e.g., through feature importance analysis of the optimal GBDT thin film thickness) is explained. Figure 3 As shown in the figure, the molecular weight and proportion of water-soluble resin have the greatest impact on the film thickness of the laser cutting protective fluid, and they further affect the film thickness by influencing the viscosity.

[0078] Predicting the viscosity of formulations with unknown viscosities for the optimization design of protective fluid formulations: A new protective fluid formulation (20 parts) was reconfigured, and its viscosity was measured to collect viscosity data; based on the predictive power of the optimal model, the viscosity (unit: mm) of the new protective fluid formulation combination was calculated. 2 The generalization ability of the model is analyzed by comparing the prediction results with the experimental results (as shown in Table 3).

[0079] Table 3. Comparison of predicted viscosity results and experimental results for 20 reconfigured protective solutions

[0080] Predicted value Experimental values Formula 1 10.12 9.24 Formula 2 15.03 17.04 Formula 3 14.38 15.41 Formula 4 23.21 24.34 Formula 5 21.56 19.62 Formula 6 34.74 33.80 Formula 7 29.30 29.34 Formula 8 33.20 31.21 Formula 9 32.2750 32.28 Formula 10 52.713 50.78 Formula 11 29.38 29.31 Formula 12 54.86 55.78 Formula 13 62.14 61.23 Formula 14 70.93 73.05 Formula 15 78.37 75.50 Formula 16 101.88 105.87 Formula 17 79.32 80.34 Formula 18 91.96 95.26 Formula 19 114.96 110.13 Formula 20 147.39 145.46

[0081] It can be seen that this method, by integrating experimental data, theoretical models and machine learning algorithms, has constructed a high-precision model for predicting the viscosity performance of protective liquids.

[0082] Based on a theoretical model, the relationship between viscosity, rotational speed, and film thickness is established to predict the corresponding film thickness: The relationship between viscosity, rotational speed, and film thickness is established based on the Meyerhofer model, and the film thickness data is theoretically corrected using the following formula:

[0083] Among them, h f υ is the film thickness (in nm), and υ is the kinematic viscosity (in mm). 2 / s), where ω is the spin coating speed (unit: rpm);

[0084] By experimentally fitting the Meyerhofer model (e.g.) Figure 4As shown in Table 4, empirical formulas for PVA and PVP protective fluids are derived respectively.

[0085] Table 4. Fitting parameters and correlation coefficients of PVA and PVP protective liquid film thickness based on the Meyerhofer model.

[0086] value Standard deviation <![CDATA[R 2 ]]> PVP k1 16855.28792 121.2915 0.986 PVA k2 6922.41728 122.61196 0.920

[0087] By substituting the viscosity and corresponding rotation speed of different formulations, the corresponding film thickness can be calculated.

[0088] Next, the laser cutting protective solution of this application is prepared. The preparation method of the laser cutting protective solution of the comparative example and application example of this application is as follows: First, water-soluble resin is mixed with water and stirred at 1000 rpm for 240 min. Then, organic solvent, plasticizer, defoamer, ultraviolet absorber and antioxidant are added in sequence and mixed and stirred at 1000 rpm for 20 min to obtain laser cutting protective solution.

[0089] Application Example 1

[0090] Weigh out 20 parts of polyvinylpyrrolidone (PVP), 15 parts of propylene glycol methyl ether, 5 parts of polyethylene glycol 800, 0.02 parts of polyether-modified organosilicon, 0.025 parts of ferulic acid, 0.25 parts of L-ascorbic acid, and 59.68 parts of pure water for later use.

[0091] In a cleanroom environment, water-soluble resin polyvinylpyrrolidone (PVP) was first mixed with water at room temperature and stirred at 1000 rpm for 240 min. Then, organic solvent, plasticizer, defoamer, UV absorber and antioxidant were added in sequence and mixed and stirred at 1000 rpm for 20 min to obtain the laser cutting protective liquid of Application Example 1.

[0092] Comparative Example 1

[0093] Weigh out 10 parts of polyvinylpyrrolidone (PVP), 15 parts of propylene glycol methyl ether, 5 parts of polyethylene glycol 800, 0.02 parts of polyether-modified organosilicon, 0.025 parts of ferulic acid, 0.25 parts of L-ascorbic acid, and 69.68 parts of pure water for later use. Other preparation methods and conditions are the same as in Application Example 1, yielding the laser cutting protective solution of Comparative Example 1.

[0094] Comparative Example 2

[0095] The preparation method and conditions were the same as in Application Example 1, and the laser cutting protective liquid of Comparative Example 2 was obtained.

[0096] The laser protective liquid prepared above was spin-coated onto the surface of a wafer using Example 1, Comparative Example 1, and Comparative Example 2. The spin-coating was performed at the corresponding rotation speeds, and after film formation, the wafer was placed in a laser dicing machine with a power of 7.2W, a frequency of 1000kHz, a speed of 100mm / s, a wavelength of 355nm, and two blades for dicing. After dicing, the dicing effect was observed and recorded under a scanning electron microscope (SEM).

[0097] The laser cutting protective liquid provided in Case 1 and Comparative Examples 1 and 2 were coated and tested according to the above configuration method. The composition and effect are shown in Table 5.

[0098] Table 5. Comparison of test results of protective liquid coating and cutting effects in Application Example 1 and Comparative Examples 1 and 2

[0099] Application Example 1 Comparative Example 1 Comparative Example 2 Polyvinylpyrrolidone 20 10 20 Propylene glycol methyl ether 15 15 15 Polyethylene glycol 800 5 5 5 Polyether modified organosilicon 0.02 0.02 0.02 ferulic acid 0.025 0.025 0.025 L-Ascorbic Acid 0.25 0.25 0.25 pure water 59.68 69.68 59.68 <![CDATA[Viscosity (mm 2 / s)]]> 63.42 11.23 63.42 Rotational speed (rpm) 2000 2000 500 Film thickness (nm) 1503.02 844.01 3006.03 Coating and cutting effects The coating is even, and the cuts are relatively smooth. Uneven cutting, resulting in crater-like formations. Uneven film thickness affects cutting consistency.

[0100] Application Example 1 demonstrates uniform coating and relatively smooth cutting results after coating and dicing. This is because the liquid film formed by spin coating has a higher surface tension during contraction, lower viscosity, smaller and narrower grooves, and is more likely to form a smooth and uniform film when it remains stationary during the solvent evaporation period. Application Example 1 can reduce the laser cutting width, thereby reducing chip yield loss. Compared to Application Example 1, Comparative Example 1 reduced polyvinylpyrrolidone from 20 parts to 10 parts and pure water from 59.68 parts to 69.68 parts. The excessively thin protective film formed by spin coating in Comparative Example 1 may allow laser energy to directly act on the wafer surface, causing silicon substrate amorphization or metal layer melting, resulting in uneven cutting and cratering, which does not meet the requirements of subsequent chip processing stages. Compared to Application Example 1, Comparative Example 2 has the same composition, but its low rotation speed results in an excessively thick protective film formed by spin coating. The uneven film formation hinders the rapid heat conduction during laser cutting, leading to a sudden increase in local temperature. This may cause microcracks or thermal stress deformation in the wafer material, affecting the consistency of cutting.

[0101] The laser dicing protective solutions prepared in Application Example 1 and Comparative Example 1 were respectively coated onto wafers to form films, which were then diced using a 355 nm ultraviolet laser. After dicing, the wafers were observed under a scanning electron microscope (SEM). The wafers protected by Application Example 1 and Comparative Example 1 after dicing are shown below. Figure 5 As shown in the figure, the protective liquid coating in Application Example 1 (left) is uniform, and the cut is relatively smooth. No residual melt spatter or abnormal morphology of the heat-affected zone was observed. In Comparative Example 1 (right), the protective liquid coating is uneven, the cut is uneven, crater-like depressions appear at the edge of the cut, and discolored deposits are clearly visible at the edge of the cut.

[0102] In summary, this invention provides a machine learning-based method for predicting and optimizing the performance of laser cutting protective fluids and its application verification in practical industry. This method innovatively integrates multi-dimensional experimental data from 139 PVA and PVP protective fluids, employs machine learning algorithms such as GBDT to construct a prediction model, and establishes a high-precision protective fluid viscosity prediction system. Combined with the Meyerhofer equation theoretical model, it achieves accurate prediction of the protective fluid film thickness, which can be used for intelligent formulation design and optimization. The viscosity prediction model of this invention performs excellently on the enhanced dataset test set, with a viscosity prediction coefficient R0. 2 The value reached 0.9916, and the MSE was 0.1003*10. -3 By integrating multi-dimensional feature engineering with macroscopic process parameters and microscopic molecular descriptors, and employing a nested 5-fold cross-validation strategy, the model can accurately capture the film-forming mechanism of the protective fluid and achieve global hyperparameter optimization. Notably, this invention not only achieves a breakthrough in theoretical prediction models but also further validates the optimized formulation through actual laser cutting, demonstrating its uniform coating and resulting in smoother cuts. This novel "theoretical prediction - process optimization - industrial validation" model overcomes the high-cost bottleneck of traditional "trial and error" methods, providing an efficient solution for the intelligent formulation design of laser cutting protective fluids. It possesses significant engineering application value and commercial potential in fields such as semiconductor manufacturing and the photovoltaic industry.

[0103] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0104] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0105] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0106] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0107] This invention is not limited to the above-described preferred embodiments. Anyone inspired by this invention can derive other forms of machine learning-based laser cutting protective fluid performance prediction and optimization methods. All equivalent variations and modifications made within the scope of the claims of this invention should be included within the scope of this invention.

Claims

1. A method for predicting and optimizing the performance of laser cutting protective fluid based on machine learning, characterized in that, include: A multidimensional feature dataset for laser cutting protective liquid is constructed, which includes molecular descriptors of formulation components and cross-scale fusion features of spin coating process parameters; Feature selection is performed on the dataset to reduce the number of features and filter out key features related to viscosity; Based on the filtered dataset, a viscosity prediction model was trained using a machine learning model. Based on the viscosity prediction value and spin coating speed of the model, the film thickness is calculated using the Meyerhofer model, where the film thickness is directly proportional to the cube root of the viscosity and inversely proportional to the square root of the spin coating speed. The protective solution formulation is optimized based on the film thickness.

2. The method for predicting and optimizing the performance of laser cutting protective fluid based on machine learning according to claim 1, characterized in that: The formulation comprises water-soluble resin, organic solvent, plasticizer, defoamer, ultraviolet absorber, antioxidant and water, wherein the water-soluble resin is PVA type or PVP type.

3. The method for predicting and optimizing the performance of laser cutting protective fluid based on machine learning according to claim 1, characterized in that: The molecular descriptors are obtained by converting the molecular structure of the formulation components into the standard SMILES format using ChemDraw, and then calculating multiple molecular descriptors using RDKit, covering geometric, electronic, topological features and spatial configuration; the multidimensional feature dataset also includes the proportion of the formulation components and the molecular weight of the water-soluble resin.

4. The method for predicting and optimizing the performance of laser cutting protective fluid based on machine learning according to claim 1, characterized in that: The feature selection employed one or more combined methods, including low variance filtering, correlation analysis, and PCA dimensionality reduction, ultimately identifying eight key features significantly related to viscosity.

5. The method for predicting and optimizing the performance of laser cutting protective fluid based on machine learning according to claim 1, characterized in that: The machine learning model includes a decision tree ensemble model, which is selected from gradient boosting decision tree, random forest, XGBoost or LightGBM, wherein the optimal model is the gradient boosting decision tree model, and the gradient boosting decision tree model has a determination coefficient R² ≥ 0.99 on the test set.

6. The method for predicting and optimizing the performance of laser cutting protective fluid based on machine learning according to claim 1, characterized in that: When training the viscosity prediction model, the selected dataset is divided into a training set and a test set. A nested K-fold cross-validation strategy is adopted, and the hyperparameters are optimized by grid search. The hyperparameters include the learning rate, the number of trees, and the maximum depth.

7. The method for predicting and optimizing the performance of laser cutting protective fluid based on machine learning according to claim 1, characterized in that: In the Meyerhofer model, corresponding empirical constants k are fitted for PVA-type and PVP-type protective liquids respectively, in order to adapt to the film-forming characteristics of different types of protective liquids.

8. The method for predicting and optimizing the performance of laser cutting protective fluid based on machine learning according to claim 1, characterized in that: The optimization of the protective liquid formulation based on film thickness includes: adjusting the proportion and molecular weight of water-soluble resin, wherein increasing the proportion of water-soluble resin is used to improve film thickness uniformity, so that the predicted film thickness meets the requirements of laser cutting for film thickness uniformity and cutting accuracy.

9. A machine learning-based system for predicting and optimizing the performance of laser cutting protective fluid, characterized in that, include: The data acquisition module is used to construct a multidimensional feature dataset of the laser cutting protective liquid, which includes molecular descriptors of the formulation components and cross-scale fusion features of spin coating process parameters. The feature processing module is used to perform feature selection on the dataset to compress the number of features and filter out key features related to viscosity. The model training module is used to train a viscosity prediction model using a machine learning model based on the filtered dataset. The film thickness calculation module is used to calculate the film thickness based on the viscosity prediction value and spin coating speed of the model using the Meyerhofer model, wherein the film thickness is directly proportional to the cube root of the viscosity and inversely proportional to the square root of the spin coating speed. The formulation optimization module is used to optimize the protective liquid formulation based on the film thickness.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the method as described in any one of claims 1-8.

Images