A method and device for predicting the removal rate of silicon carbide substrate grinding material

Abstract

The present application provides a method and device for predicting the grinding material removal rate of a silicon carbide substrate. The method comprises: fitting the acquired historical process test data to construct an empirical material removal rate prediction model; collecting dynamic data of the grinding process in response to a correction request; using a trained XGBoost dynamic correction model to predict the characteristic parameters to be measured corresponding to the dynamic data of the grinding process to obtain a dynamic correction amount; updating the empirical coefficient in the empirical material removal rate prediction model based on the dynamic correction amount to obtain a corrected empirical material removal rate prediction model; and calculating the removal rate prediction value corresponding to the dynamic data of the grinding process using the corrected empirical material removal rate prediction model. The present invention predicts the dynamic correction amount by using the XGBoost model to update the empirical coefficient, realizes the prediction of the removal rate prediction value, and improves the prediction accuracy and process control level of the grinding process.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor material precision processing, and in particular to a method and device for predicting the removal rate of abrasive material of a silicon carbide substrate. Background Art

[0002] Silicon carbide (SiC), due to its excellent thermal stability, electrical properties, and mechanical strength, has become a core representative of third-generation semiconductor materials and is widely used in key areas such as high-power electronic devices, radio frequency communications, and rail transportation. To ensure the processing accuracy and functional integrity of its devices, SiC substrates undergo multiple rounds of precision grinding and polishing in the back-end process to achieve a highly flat wafer surface with few surface defects.

[0003] During silicon carbide grinding, the material removal rate (MRR) serves as a key indicator of machining efficiency and surface quality control. Its prediction accuracy directly impacts process parameter settings and the stability of machining quality. Traditional MRR prediction methods are primarily based on empirical or physical models, deriving an expression by fitting a power function relationship between process parameters such as pressure, speed, and abrasive particle size and the MRR. While these models are simple in structure and highly physically interpretable, they struggle to address the nonlinear dynamic behavior of actual machining processes caused by factors such as equipment fluctuations, material differences, and varying operating conditions.

[0004] On the other hand, although the data-driven modeling methods that have emerged in recent years (such as neural networks and ensemble learning) have strong fitting and generalization capabilities, they generally lack physical constraints and interpretability, and are prone to overfitting or insufficient robustness in the absence of stage perception and structural guidance.

[0005] In summary, existing technologies for predicting material removal rates during silicon carbide polishing still face the following key challenges: First, the model's weak response to dynamic changes prevents adaptive process adjustment; second, the lack of stage identification and feature scheduling mechanisms results in poor generalization of the prediction model under different working conditions; and third, the ineffective integration of data-driven and empirical mechanisms creates a trade-off between model accuracy and interpretability. Therefore, an MRR prediction method that integrates physical models with dynamic data, adaptive scheduling, and high-precision correction capabilities is urgently needed to improve modeling accuracy and intelligent control of the silicon carbide substrate polishing process. Summary of the Invention

[0006] The present invention provides a method and device for predicting the grinding material removal rate of a silicon carbide substrate, which are used to solve the problems that traditional prediction methods are difficult to cope with nonlinear dynamic changes under actual processing conditions, lack process stage identification and feature scheduling mechanisms, and have difficulty in balancing model prediction accuracy and interpretability.

[0007] In a first aspect, the present invention provides a method for predicting a silicon carbide substrate grinding material removal rate, comprising:

[0008] Fit the acquired historical process test data to build an empirical material removal rate prediction model;

[0009] In response to the correction request, collecting dynamic data of the grinding process;

[0010] Using the trained XGBoost dynamic correction model, prediction is performed based on the characteristic parameters to be measured corresponding to the dynamic data of the grinding process to obtain a dynamic correction value;

[0011] Based on the dynamic correction amount, updating the empirical coefficient in the empirical material removal rate prediction model to obtain a corrected empirical material removal rate prediction model;

[0012] The removal rate prediction value corresponding to the dynamic data of the grinding process is calculated using the revised empirical material removal rate prediction model.

[0013] Optionally, a trained XGBoost dynamic correction model is used to predict the characteristic parameters to be measured corresponding to the dynamic data of the grinding process to obtain a dynamic correction amount, including:

[0014] Based on the historical process test data and its historical removal rate, a matching relationship between the grinding stage and the sensitive feature set is established;

[0015] Determining a sensitive feature set to be measured for the dynamic data of the grinding process at the grinding stage to be measured based on a matching relationship between the grinding stage and the sensitive feature set;

[0016] The feature parameters to be measured corresponding to the sensitive feature set to be measured in the dynamic data of the grinding process are input into the trained XGBoost dynamic correction model to obtain the dynamic correction amount.

[0017] Optionally, based on the historical process test data and its historical removal rate, a matching relationship between the grinding stage and the sensitive feature set is established, including:

[0018] Grinding and dividing the historical process test data according to a preset physical threshold rule;

[0019] A weighted approach using the Pearson correlation coefficient and mutual information analysis is employed to construct a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate.

[0020] Optionally, before the step of determining the sensitive feature set to be measured for the grinding stage to be measured of the grinding process dynamic data based on the matching relationship between the grinding stage and the sensitive feature set, the method further includes:

[0021] According to the physical threshold rule, the dynamic data of the grinding process is ground and divided to determine the grinding stage to be measured.

[0022] Optionally, a weighted method of Pearson correlation coefficient and mutual information analysis is used to construct a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate, including:

[0023] respectively calculating the Pearson coefficient and the mutual information value between the historical characteristic parameters of each feature at different grinding stages and the historical removal rate in the historical process test data;

[0024] Normalizing all the mutual information values, and calculating a weighted sensitivity score of each of the features at different grinding stages based on the Pearson coefficient and the normalized mutual information values;

[0025] According to the magnitude relationship of the sensitivity scores, a preset number of features are selected to form corresponding sensitive feature sets, so as to form a matching relationship between the grinding stage and the sensitive feature set.

[0026] Optionally, the trained XGBoost dynamic correction model includes multiple dynamic correction sub-models at different stages; inputting the measured feature parameters corresponding to the measured sensitive feature set in the dynamic data of the grinding process into the trained XGBoost dynamic correction model to obtain the dynamic correction amount includes:

[0027] Based on the grinding stage to be tested, extracting the feature parameters to be tested corresponding to the sensitive feature set to be tested from the dynamic data of the grinding process;

[0028] The characteristic parameter to be measured is input into the dynamic correction sub-model of the corresponding stage to obtain the dynamic correction amount.

[0029] Optionally, the empirical material removal rate prediction model is:

[0030] ；

[0031] in, is the predicted removal rate, is the first empirical coefficient, is the second empirical coefficient, is the third empirical coefficient, is the fourth empirical coefficient, Grinding pressure, is the grinding speed, is the abrasive particle size.

[0032] Optionally, the training steps of the trained XGBoost dynamic correction model include:

[0033] Determining a sample removal rate prediction value corresponding to the dynamic training sample data of the grinding process by using the empirical material removal rate prediction model;

[0034] The ratio of the actual sample removal rate of the dynamic training sample data of the grinding process to the corresponding sample removal rate prediction value is used as the removal rate ratio;

[0035] The sensitive features of the training samples are used as input, and the difference between the removal rate ratio and the first empirical coefficient is used as a label to train the constructed XGBoost dynamic correction model to obtain a target XGBoost dynamic correction model.

[0036] In a second aspect, the present invention provides a device for predicting abrasive material removal rate of a silicon carbide substrate, comprising:

[0037] The model building module is used to fit the acquired historical process test data and build an empirical material removal rate prediction model;

[0038] an acquisition module, configured to acquire dynamic data of the grinding process in response to a correction request;

[0039] A dynamic correction amount determination module is used to use the trained XGBoost dynamic correction model to predict the measured characteristic parameters corresponding to the dynamic data of the grinding process to obtain a dynamic correction amount;

[0040] a correction module, configured to update the empirical coefficients in the empirical material removal rate prediction model based on the dynamic correction amount to obtain a corrected empirical material removal rate prediction model;

[0041] The prediction module is used to calculate the removal rate prediction value corresponding to the dynamic data of the grinding process through the modified empirical material removal rate prediction model.

[0042] Optionally, the dynamic correction amount determination module includes:

[0043] a relationship building submodule for building a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate;

[0044] a submodule for determining a sensitive feature set to be measured, configured to determine a sensitive feature set to be measured for the dynamic data of the grinding process in the grinding stage to be measured based on a matching relationship between the grinding stage and the sensitive feature set;

[0045] The dynamic correction amount determination submodule is used to input the feature parameters to be measured corresponding to the sensitive feature set to be measured in the dynamic data of the grinding process into the trained XGBoost dynamic correction model to obtain the dynamic correction amount.

[0046] Optionally, the relationship building submodule includes:

[0047] A division unit, configured to perform grinding division on the historical process test data according to a preset physical threshold rule;

[0048] The relationship building unit is used to build a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate by using a weighted method of Pearson correlation coefficient and mutual information analysis.

[0049] Optionally, the dynamic correction amount determination module further includes:

[0050] The division submodule is used to divide the dynamic data of the grinding process into grinding stages according to the physical threshold rule, and determine the grinding stage to be measured.

[0051] Optionally, the relationship building submodule includes:

[0052] a relationship value calculation unit, configured to respectively calculate the Pearson coefficient and the mutual information value between the historical characteristic parameters of each feature at different grinding stages and the historical removal rate in the historical process test data;

[0053] a weighted sensitivity score calculation unit, configured to normalize all the mutual information values ​​and calculate a weighted sensitivity score of each of the features at different grinding stages based on the Pearson coefficient and the normalized mutual information values;

[0054] The relationship building unit is used to select a preset number of features to form a corresponding sensitive feature set according to the size relationship of the sensitivity scores, so as to form a matching relationship between the grinding stage and the sensitive feature set.

[0055] Optionally, the trained XGBoost dynamic correction model includes multiple dynamic correction sub-models at different stages; the dynamic correction amount determination sub-module includes:

[0056] A characteristic parameter extraction unit to be measured is used to extract the characteristic parameters to be measured corresponding to the sensitive feature set to be measured from the dynamic data of the grinding process based on the grinding stage to be measured;

[0057] The dynamic correction value acquisition unit is used to input the characteristic parameter to be measured into the dynamic correction sub-model of the corresponding stage to obtain the dynamic correction value.

[0058] Optionally, the training of the trained XGBoost dynamic correction model is implemented by the following modules:

[0059] A sample prediction module, configured to determine a sample removal rate prediction value corresponding to the dynamic training sample data of the grinding process by using the empirical material removal rate prediction model;

[0060] a removal rate ratio determination module, configured to use the ratio of the actual sample removal rate of the dynamic training sample data of the grinding process to the corresponding sample removal rate prediction value as the removal rate ratio;

[0061] The training module is used to take the sensitive features of the training samples as input and the difference between the removal rate ratio and the first empirical coefficient as a label to train the constructed XGBoost dynamic correction model to obtain a target XGBoost dynamic correction model.

[0062] In a third aspect, the present invention provides an electronic device comprising a processor and a memory, wherein the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the steps in the method provided in the first aspect are executed.

[0063] In a fourth aspect, the present invention provides a storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, runs the steps of the method provided in the first aspect above.

[0064] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, runs the steps of the method provided in the first aspect.

[0065] It can be seen from the above technical solutions that the present invention has the following advantages:

[0066] The present invention provides a method and device for predicting the material removal rate during silicon carbide substrate grinding. This method constructs an empirical model by fitting historical process test data, collects dynamic data in real time, and divides the grinding phases into regular intervals. The XGBoost model is used to predict dynamic corrections and update empirical coefficients, achieving high-precision removal rate predictions. This significantly improves the prediction accuracy and process control level of the silicon carbide substrate grinding process. This method addresses the difficulties faced by traditional prediction methods in coping with nonlinear dynamic changes under actual machining conditions, lacks process phase identification and feature scheduling mechanisms, and struggles to balance model prediction accuracy and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 This is a flowchart of a method for predicting abrasive material removal rate of a silicon carbide substrate according to a first embodiment of the present invention;

[0069] Figure 2 This is a flowchart of a second embodiment of a method for predicting abrasive material removal rate of a silicon carbide substrate according to the present invention;

[0070] Figure 3 This is a structural block diagram of an embodiment of a device for predicting the removal rate of polishing material on a silicon carbide substrate according to the present invention. DETAILED DESCRIPTION

[0071] An embodiment of the present invention provides a method and device for predicting the grinding material removal rate of a silicon carbide substrate, which is used to solve the problems that traditional prediction methods are difficult to cope with nonlinear dynamic changes under actual processing conditions, lack process stage identification and feature scheduling mechanisms, and have difficulty in balancing model prediction accuracy and interpretability.

[0072] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0073] For example 1, please refer to Figure 1 , Figure 1 This is a flow chart of a first embodiment of a method for predicting abrasive material removal rate of a silicon carbide substrate according to the present invention. The method includes:

[0074] Step S101, fitting the acquired historical process test data to construct an empirical material removal rate prediction model;

[0075] In this example, historical process test data related to the grinding process was collected, covering factors that may affect the material removal rate, such as grinding pressure, grinding speed, grinding time, and abrasive particle size. The collected data was then preprocessed, including data cleaning and data normalization, and then fitted to construct an empirical material removal rate prediction model.

[0076] Step S102, in response to the correction request, collecting dynamic data of the grinding process;

[0077] In the embodiment of the present application, the pre-set grinding pressure, grinding speed and abrasive particle size are obtained.

[0078] Step S103, using the trained XGBoost dynamic correction model, predicting the measured characteristic parameters corresponding to the dynamic data of the grinding process to obtain a dynamic correction value;

[0079] In an embodiment of the present application, the characteristic parameters to be measured corresponding to the grinding stage to be measured are extracted from the dynamic data of the grinding process, and then the extracted characteristic parameters to be measured are input into the trained XGBoost dynamic correction model, and the dynamic correction amount is obtained using the prediction ability of the model.

[0080] Step S104, updating the empirical coefficients in the empirical material removal rate prediction model based on the dynamic correction amount to obtain a corrected empirical material removal rate prediction model;

[0081] In the embodiment of the present application, the empirical coefficients in the empirical material removal rate prediction model are updated based on the obtained dynamic correction amount to improve the accuracy and adaptability of the model. The updated empirical coefficients are then substituted into the empirical material removal rate prediction model to obtain a revised empirical material removal rate prediction model.

[0082] Step S105 , calculating a removal rate prediction value corresponding to the dynamic data of the grinding process by using the modified empirical material removal rate prediction model.

[0083] In an embodiment of the present application, the dynamic data of the grinding process is input into the revised empirical material removal rate prediction model, and the corresponding material removal rate prediction value is calculated based on the input data and the updated empirical coefficient.

[0084] The present invention provides a method for predicting the material removal rate of silicon carbide substrate grinding. This method constructs an empirical model by fitting historical process test data, collects dynamic data in real time, divides the grinding phases according to rules, and uses the XGBoost model to predict dynamic corrections to update empirical coefficients. This method achieves high-precision removal rate predictions, significantly improving the prediction accuracy and process control level of the silicon carbide substrate grinding process. This method addresses the problems that traditional prediction methods have difficulty in coping with nonlinear dynamic changes under actual processing conditions, lack process phase identification and feature scheduling mechanisms, and have difficulty balancing model prediction accuracy and interpretability.

[0085] For example 2, please refer to Figure 2 , Figure 2This is a flow chart of a second embodiment of a method for predicting a silicon carbide substrate grinding material removal rate according to the present invention, the steps comprising:

[0086] Step S201: Fit the acquired historical process test data to construct an empirical material removal rate prediction model; the empirical material removal rate prediction model is:

[0087] ；

[0088] in, is the predicted removal rate, is the first empirical coefficient, is the second empirical coefficient, is the third empirical coefficient, is the fourth empirical coefficient, Grinding pressure, is the grinding speed, is the abrasive particle size.

[0089] In an embodiment of the present application, an empirical material removal rate prediction model is constructed using historical process test data, and the material removal rate prediction value is expressed as a function of grinding pressure, grinding speed, and abrasive particle size.

[0090] It should be noted that this model is used as the subsequent baseline model to represent the material removal rate under “ideal or static conditions”.

[0091] Step S202, in response to the correction request, collecting dynamic data of the grinding process;

[0092] Step S203, dividing the dynamic data of the grinding process into grinding process stages according to a preset physical threshold rule, and determining the grinding stage to be measured;

[0093] In this embodiment, the collected dynamic grinding process data is divided into grinding process stages according to preset physical threshold rules to determine the grinding stage to be measured. Based on the physical change characteristics of temperature rise rate, vibration intensity, and acoustic emission energy, the grinding process stages are divided into initial rough grinding stage (Stage I), stable grinding stage (Stage II), and final finishing stage (Stage III).

[0094] In practice, at the beginning of grinding, due to the relatively rough substrate surface, the contact area between the abrasive particles and the substrate is large, resulting in high friction and significant heat generation. Furthermore, the equipment will experience significant vibrations due to initial operation and surface unevenness. Therefore, the initial rough grinding stage is determined by a temperature rise rate greater than 0.2 and a vibration intensity greater than 2.0. The primary goal at this stage is to quickly remove a large amount of material and initially reduce the substrate surface roughness.

[0095] In this embodiment, the collected dynamic data of multiple physical quantities during the grinding process is first segmented and analyzed according to preset physical threshold rules to identify the grinding stage to be measured. Combining three indicators—temperature rise rate, vibration intensity, and acoustic emission (AE) energy—the grinding process is divided into: initial rough grinding (Stage I), stable grinding (Stage II), and final finishing (Stage III).

[0096] The judgment conditions for the initial rough grinding stage include: (1) temperature rise rate > 1.0 °C s⁻¹; (2) vibration intensity > 2.0 gRMS (10 Hz-5 kHz bandwidth, three-axis acceleration after 600 Hz low-pass filtering); (3) AE energy continuously higher than the baseline of the steady grinding stage by ≥ 3 dB.

[0097] These thresholds were determined based on historical silicon carbide substrate samples obtained through an experimental design and statistically analyzed with a 95% confidence interval: During rough grinding, contact friction work increases significantly, with a typical temperature rise rate ranging from 1.0-5.0°C / s⁻¹; vibration amplitude often exceeds 2.0 gRMS; and AE energy increases by 3-5 dB relative to the steady grinding baseline, reflecting an increase in abrasive impact and embrittlement events. If any two of these conditions are met, the process enters Stage I. The primary goal of this stage is to efficiently remove surface defects and rapidly reduce initial roughness, creating conditions for stable grinding in Stage II.

[0098] The experimental design is shown in Table 1 below:

[0099] Table 1 Experimental design

[0100]

[0101] As shown in the table, the number of experimental groups for the model training / calibration phase and the threshold calibration phase requires a total of 50-60 groups, and an independent independent validation set is reserved for the validation / blank phase in the experimental phase. The reasons are as follows: (1) Regression robustness: If the model contains 4 parameters, 30 groups can only barely meet the "10× parameter" rule; increasing to 40 groups can significantly reduce the parameter variance. (2) Confidence interval width: The 95% quantile error is larger than the confidence interval width. Convergence, for example, when n = 36, the error band is ±0.27σ, and when n = 50, the error band shrinks to ±0.22σ, where n is the number of experimental groups. (3) Stage sample balance: The coarse grinding stage is often less than 25% of the total length. Doing more experiments can avoid threshold deviation caused by insufficient data in Stage I.

[0102] The thresholds can be recalibrated monthly in a mass production environment to ensure compatibility with substrate and abrasive batches. The baseline is the 30-second average of the signal during the machine's no-load and steady-grinding phases. The vibration bandwidth is 10 Hz to 5 kHz, and the RMS value is taken after a 600 Hz low-pass filter. The AE energy measurement window is 100-400 kHz, with a sliding window length of 0.1 seconds. Furthermore, the judgment logic in the table states that if all conditions are met, the corresponding stage is entered. If the signal fluctuates across the threshold, a 2-second hysteresis is used to confirm the result to avoid jitter.

[0103] As polishing progresses, the substrate surface gradually becomes flatter, and the polishing process stabilizes. At this point, the temperature rise rate remains constant, and the material removal rate exhibits minimal fluctuation. This is because the contact between the abrasive particles and the substrate surface is relatively stable during this phase, and various process parameters are also relatively stable, resulting in a relatively stable material removal rate.

[0104] To achieve higher surface quality requirements near the end of grinding, fine finishing is required. During this stage, the acoustic emission energy decreases significantly, and the material removal rate decreases. This is because during the final finishing phase, the grinding force and speed are appropriately reduced to avoid over-grinding, resulting in a decrease in acoustic emission energy and material removal rate.

[0105] Step S204: constructing a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate;

[0106] The specific steps include: dividing the historical process test data into grinding stages according to a preset physical threshold rule; and constructing a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate using a weighted method of Pearson correlation coefficient and mutual information analysis.

[0107] In an embodiment of the present application, the Pearson coefficient and mutual information value between the historical feature parameters of each feature in different grinding stages and the historical removal rate in the historical process test data are calculated respectively; all the mutual information values ​​are normalized, and based on the Pearson coefficient and the normalized mutual information value, the weighted sensitivity score of each feature in different grinding stages is calculated; according to the size relationship of the sensitivity scores, a preset number of features are selected to form a corresponding sensitive feature set to form a matching relationship between the grinding stage and the sensitive feature set.

[0108] It should be noted that the Pearson correlation coefficient is used to measure the degree of linear correlation between two continuous variables, and mutual information is derived from information theory and is used to measure any dependency (including linear and nonlinear) between two random variables.

[0109] In practice, in order to accurately identify the key factors affecting the material removal rate at each stage of the silicon carbide substrate grinding process, this application calculates the correlation scores between all candidate features and the target variable in each stage, and selects a subset of sensitive features that distinguish between stages accordingly. The specific steps are as follows:

[0110] (1) Assume It is a collection of three process stages; each stage Contains samples; a total of The original features are recorded as ; The target variable is material removal rate .

[0111] (2) For each stage , calculate each feature separately and Pearson correlation coefficient and mutual information value , the calculation formula is as follows:

[0112] ;

[0113] ;

[0114] in, For the Phase I Samples in the feature The value of ; Characterized by The mean value at this stage; For the Phase I Material removal rate corresponding to the sample; For the The mean of the material removal rates of all samples within the stage; Get the value of a variable The marginal probability density of ; Get the value of a variable The marginal probability density of ; Get the value of a variable and The joint probability density of 、 and Estimated by histogram or kernel density estimation.

[0115] (3) A weighted fusion scoring function is introduced to define the comprehensive sensitivity score for each feature at each stage as:

[0116] ;

[0117] in: is the Pearson weight, with a size of 0.5; is the mutual information weight, which is 0.5; Indicates the Stage characteristics This weighted approach allows for comprehensive consideration of both linear and nonlinear correlations, allowing for a more comprehensive assessment of the impact of each feature on the material removal rate.

[0118] (4) Sort the sensitivity scores of all features in descending order and select the top The characteristics of the first Sensitive feature subset of the stage :

[0119] ;

[0120] Finally, the following feature set is formed: : Stage I sensitive feature subset; : Stage II sensitive feature subset; : Stage III sensitive feature subset.

[0121] Step S205, determining a sensitive feature set to be measured for the dynamic data of the grinding process in the grinding stage to be measured based on the matching relationship between the grinding stage and the sensitive feature set;

[0122] In the embodiment of the present application, the sensitive feature set of the corresponding stage is dynamically called according to the real-time stage recognition result. The grinding stage to be determined is judged based on the current sensor data stream, and the corresponding sensitive feature set to be measured is determined.

[0123] In actual application, the system uses real-time collected dynamic data from the grinding process, including temperature rise rate, vibration RMS, and acoustic emission energy, combined with preset physical threshold rules, to determine the grinding stage in real time. A subset of sensitive features corresponding to the corresponding stage is then used as model input, and the matching sub-model is scheduled to predict the dynamic correction amount. This joint feature-model scheduling mechanism ensures that the prediction model uses the optimal input features and regression structure at different times, thereby improving prediction accuracy and system adaptability.

[0124] Step S206, inputting the measured feature parameters corresponding to the measured sensitive feature set in the dynamic data of the grinding process into the trained XGBoost dynamic correction model to obtain the dynamic correction amount; the trained XGBoost dynamic correction model includes multiple dynamic correction sub-models at different stages;

[0125] In an embodiment of the present application, based on the grinding stage to be measured, the characteristic parameters to be measured corresponding to the sensitive feature set to be measured are extracted from the dynamic data of the grinding process; the characteristic parameters to be measured are input into the dynamic correction sub-model of the corresponding stage to obtain the dynamic correction amount.

[0126] The training steps of the XGBoost dynamic correction model include: determining the sample removal rate prediction value corresponding to the dynamic training sample data of the grinding process through the empirical material removal rate prediction model; taking the ratio of the actual sample removal rate of the dynamic training sample data of the grinding process to the corresponding sample removal rate prediction value as the removal rate ratio; taking the sensitive features of the training sample as input and the difference between the removal rate ratio and the first empirical coefficient as the label, the constructed XGBoost dynamic correction model is trained to obtain the target XGBoost dynamic correction model.

[0127] In the specific implementation, all sub-models in each stage need to complete offline training using historical data before model deployment. The labels of the training samples are determined by the following formula:

[0128] ;

[0129] in, is the true value of the removal rate measured by the instrument, is the correction amount.

[0130] Step S207, based on the dynamic correction amount, updating the empirical coefficient in the empirical material removal rate prediction model to obtain a corrected empirical material removal rate prediction model;

[0131] In the actual grinding process, due to the influence of various factors, the empirical coefficients in the empirical model may no longer be applicable to the current situation. Therefore, updating the empirical coefficients through dynamic corrections can make the model better adapt to the dynamic changes in the actual grinding process. The embodiment of the present application updates the empirical coefficients in the empirical material removal rate prediction model based on the obtained dynamic corrections to obtain a revised empirical material removal rate prediction model. Its formula is expressed as:

[0132] ;

[0133] in, is the revised predicted removal rate.

[0134] Step S208 , calculating a removal rate prediction value corresponding to the dynamic data of the grinding process using the modified empirical material removal rate prediction model.

[0135] In addition, if the abrasive particle size influence index needs to be corrected, the model is expanded to:

[0136] ;

[0137] in, is the dynamic particle size correction.

[0138] The embodiment of the present invention discloses a method for predicting the removal rate of the grinding material of a silicon carbide substrate, which calculates the removal rate prediction value corresponding to the dynamic data of the grinding process through the revised empirical material removal rate prediction model. By fitting the historical process test data to construct an empirical model, collecting dynamic data in real time and dividing the grinding stages according to rules, and using the XGBoost model to predict the dynamic correction amount to update the empirical coefficient, a high-precision prediction effect of the removal rate prediction value is achieved, which significantly improves the prediction accuracy and process control level of the silicon carbide substrate grinding process. It solves the problems that traditional prediction methods are difficult to cope with nonlinear dynamic changes under actual processing conditions, lack process stage identification and feature scheduling mechanisms, and it is difficult to balance model prediction accuracy and interpretability.

[0139] In order to facilitate those skilled in the art to understand the beneficial effects of the present invention, an example of a method for predicting a material removal rate of a silicon carbide substrate polishing according to the present invention is described below.

[0140] This example uses a Logitech PM6 precision grinder and polisher as the lapping equipment. This machine has a 254mm platen diameter, a rotation speed range of 5 to 100 rpm, and programmable pressure. The data acquisition system uses the NI PXIe1085 chassis and PXIe 5122 digitizer. The digitizer offers 14-bit resolution and a sampling rate of up to 100 MS / s, ensuring accurate and real-time data acquisition.

[0141] In order to monitor various parameters during the grinding process in real time, the experiment installed a variety of sensors. The vibration sensor uses a piezoelectric accelerometer with a frequency response range of 1Hz to 10kHz and a sensitivity of 100mV / g, which can accurately capture vibration information during the grinding process. The temperature sensor uses a K-type thermocouple with a measurement range of -200 to 1250°C and a response speed of less than 100ms, ensuring the real-time and accuracy of temperature data. In addition, a broadband piezoelectric acoustic emission sensor with a center frequency of 150kHz and a 40dB preamplifier was used to effectively improve the detection sensitivity of the acoustic emission signal.

[0142] In terms of removal rate measurement, the experiment used a high-precision OHAUS PMK224 electronic balance with an accuracy of up to 0.01mg, and a Zygo NewView white light interferometer to ensure the accuracy of the removal rate measurement.

[0143] The prediction method steps include:

[0144] (1) Experimental process and data collection, including:

[0145] ① Sample preparation: A 4-inch diameter, 350-micron thick 4H silicon carbide wafer was selected as the experimental sample. After the pre-grinding treatment, the surface roughness Ra of the sample was approximately 200 nanometers.

[0146] ② Process Settings: The grinding pressure was set to 10 kPa, the grinding disc speed was set to 60 rpm, and the workpiece speed was set to 45 rpm. A 3-micron diamond free abrasive was used as the grinding fluid, and the flow rate was controlled at 50 ml / min.

[0147] ③ Dynamic signal acquisition: Set the sampling frequency of the vibration and acoustic emission channels to 100 kilosamples per second, and the sampling frequency of the temperature channel to 10 samples per second. During the 200-second grinding process, vibration, acoustic emission, and temperature data are continuously recorded and stored synchronously in real time.

[0148] ④ Obtaining the true value of MRR: measuring the mass difference of the sample before and after grinding, and combining it with the density of silicon carbide The volume removed was calculated by combining the grinding area and the removal volume. Furthermore, the actual material removal rate was calculated by dividing the removal volume by the grinding time. To verify the accuracy of the gravimetric method, a white light interferometer was used to measure the average thickness difference on the surface of the polished sample. This was cross-validated with the material removal rate obtained by the gravimetric method to ensure the accuracy of the measurement results, with an error margin of less than 2%.

[0149] (2) Empirical model construction:

[0150] Based on the data collection, an empirical model was constructed to preliminarily predict the material removal rate. The experimental data from the 30 groups of tests conducted in the early stage according to Table 1 were used to fit the data using the least squares method to obtain the empirical coefficients and construct an empirical material removal rate prediction model, specifically:

[0151] ;

[0152] in, is the grinding pressure, in units of ; is the grinding speed, in units of ; is the abrasive particle size, in units of This model serves as a static baseline.

[0153] (3) Automatic identification of process stages:

[0154] During the grinding process, key parameters such as the temperature rise rate, vibration root mean square value, and acoustic emission energy are calculated in real time. Combined with the preset threshold rules, the 200-second time series data is judged. The judgment results are shown in the stage judgment table in Table 2 below:

[0155] Table 2 Stage discrimination table

[0156]

[0157] The baseline is the 30-second average of the machine tool's signal during the no-load and steady-run phases; the vibration bandwidth is 10 Hz to 5 kHz, and the RMS value is obtained after a 600 Hz low-pass filter; the AE energy measurement window is 100 to 400 kHz, with a sliding window length of 0.1 seconds. Furthermore, the judgment logic in the table states that if all conditions are met, the corresponding phase is entered. If the signal fluctuates across a threshold, a 2-second hysteresis is used to confirm the result to avoid jitter.

[0158] By writing stage labels, an index is provided for subsequent feature scheduling, ensuring that the model can be dynamically adjusted according to the characteristics of different stages.

[0159] (4) Feature extraction and sensitive feature screening:

[0160] Using a sliding window technique, time domain, frequency domain, and derivative features were extracted from the time series data using a 30-second window and a 10-second step size. A comprehensive measure of the correlation between each feature and the material removal rate was performed using a weighted Pearson correlation coefficient and mutual information analysis. Based on the relative weighted sensitivity scores, the features with the highest scores in each stage were selected as the sensitive feature set. These sensitive feature sets more effectively reflect the characteristics of different grinding stages and provide important input for subsequent model training.

[0161] (5) Stage sub-model training and dynamic scheduling:

[0162] Using the sensitive features of each stage as input and dynamic corrections as output, the XGBoost algorithm was used to train three sub-models. These sub-models correspond to the initial rough grinding stage, the stable grinding stage, and the final finishing stage, respectively, and can dynamically adjust to the characteristics of each stage. During online scheduling, the scheduling controller reads the stage labels in real time and dynamically calls the corresponding feature subset and sub-model based on the current stage. By generating corrections, the model can adapt to changes in the grinding process in real time, thereby improving prediction accuracy and model practicality.

[0163] (6) Dynamic correction and results:

[0164] Based on the dynamically generated correction amount, the empirical material removal rate prediction model is corrected in real time. The corrected empirical material removal rate prediction model is:

[0165] .

[0166] The modified model was used to calculate the predicted removal rate corresponding to the dynamic data of the grinding process. By comparing the mean absolute error of the 200-second grinding data, it was found that the error was large when using only the empirical model, while the error was significantly reduced after adopting the dynamic correction method of the present invention, with a relative error reduction of 73.0%. The specific comparison results are shown in the result comparison table in Table 3 below:

[0167] Table 3 Comparison of results

[0168]

[0169] This result fully verifies the significant advantages of the method of the present invention in improving the MRR prediction accuracy, and provides strong support for the optimization of the silicon carbide substrate grinding process.

[0170] For example three, please refer to Figure 3 , Figure 3 This is a structural block diagram of an embodiment of a device for predicting the removal rate of polished material on a silicon carbide substrate according to the present invention. The device includes:

[0171] The model building module 301 is used to fit the acquired historical process test data to build an empirical material removal rate prediction model;

[0172] The collection module 302 is used to collect dynamic data of the grinding process in response to the correction request;

[0173] A dynamic correction amount determination module 303 is configured to use the trained XGBoost dynamic correction model to predict the measured characteristic parameters corresponding to the dynamic data of the grinding process to obtain a dynamic correction amount;

[0174] A correction module 304 is configured to update the empirical coefficients in the empirical material removal rate prediction model based on the dynamic correction amount to obtain a corrected empirical material removal rate prediction model;

[0175] The prediction module 305 is configured to calculate a removal rate prediction value corresponding to the dynamic data of the grinding process by using the modified empirical material removal rate prediction model.

[0176] In an optional embodiment, the dynamic correction amount determination module 303 includes:

[0177] a relationship building submodule for building a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate;

[0178] a submodule for determining a sensitive feature set to be measured, configured to determine a sensitive feature set to be measured for the dynamic data of the grinding process in the grinding stage to be measured based on a matching relationship between the grinding stage and the sensitive feature set;

[0179] The dynamic correction amount determination submodule is used to input the feature parameters to be measured corresponding to the sensitive feature set to be measured in the dynamic data of the grinding process into the trained XGBoost dynamic correction model to obtain the dynamic correction amount.

[0180] In an optional embodiment, the relationship building submodule includes:

[0181] A division unit, configured to perform grinding division on the historical process test data according to a preset physical threshold rule;

[0182] A relationship building unit is used to build a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate by using a weighted method of Pearson correlation coefficient and mutual information analysis.

[0183] In an optional embodiment, the dynamic correction amount determination module 303 further includes:

[0184] The division submodule is used to divide the dynamic data of the grinding process into grinding stages according to the physical threshold rule, and determine the grinding stage to be measured.

[0185] In an optional embodiment, the relationship building submodule includes:

[0186] a relationship value calculation unit, configured to respectively calculate the Pearson coefficient and the mutual information value between the historical characteristic parameters of each feature at different grinding stages and the historical removal rate in the historical process test data;

[0187] a weighted sensitivity score calculation unit, configured to normalize all the mutual information values ​​and calculate a weighted sensitivity score of each of the features at different grinding stages based on the Pearson coefficient and the normalized mutual information values;

[0188] The relationship building unit is used to select a preset number of features to form a corresponding sensitive feature set according to the size relationship of the sensitivity scores, so as to form a matching relationship between the grinding stage and the sensitive feature set.

[0189] In an optional embodiment, the trained XGBoost dynamic correction model includes multiple dynamic correction sub-models at different stages; the dynamic correction amount determination sub-module includes:

[0190] A characteristic parameter extraction unit to be measured is used to extract the characteristic parameters to be measured corresponding to the sensitive feature set to be measured from the dynamic data of the grinding process based on the grinding stage to be measured;

[0191] The dynamic correction value acquisition unit is used to input the characteristic parameter to be measured into the dynamic correction sub-model of the corresponding stage to obtain the dynamic correction value.

[0192] In an optional embodiment, the empirical material removal rate prediction model is:

[0193] ;

[0194] in, is the predicted removal rate, is the first empirical coefficient, is the second empirical coefficient, is the third empirical coefficient, is the fourth empirical coefficient, Grinding pressure, is the grinding speed, is the abrasive particle size.

[0195] In an optional embodiment, the training of the trained XGBoost dynamic correction model is implemented by the following modules:

[0196] A sample prediction module, configured to determine a sample removal rate prediction value corresponding to the dynamic training sample data of the grinding process by using the empirical material removal rate prediction model;

[0197] a removal rate ratio determination module, configured to use the ratio of the actual sample removal rate of the dynamic training sample data of the grinding process to the corresponding sample removal rate prediction value as the removal rate ratio;

[0198] The training module is used to take the sensitive features of the training samples as input and the difference between the removal rate ratio and the first empirical coefficient as a label to train the constructed XGBoost dynamic correction model to obtain a target XGBoost dynamic correction model.

[0199] Embodiment 4. An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a method for predicting the removal rate of polishing material of a silicon carbide substrate in any embodiment.

[0200] In a fifth embodiment, the present invention further provides a computer storage medium having a computer program stored thereon. When the computer program is executed by the processor, the computer program implements the steps of a method for predicting a silicon carbide substrate grinding material removal rate according to any one of the embodiments.

[0201] In a sixth embodiment, the present invention further provides a computer program product having a computer program stored thereon. When the computer program is executed by the processor, the computer program implements the steps of the method for predicting the grinding material removal rate of a silicon carbide substrate in any embodiment.

[0202] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0203] In the several embodiments provided in this application, it should be understood that the methods, devices, electronic devices and storage media disclosed in the present invention can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0204] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0205] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0206] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned readable storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0207] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the material removal rate of a silicon carbide substrate during grinding, characterized in that: include: Fit the acquired historical process test data to build an empirical material removal rate prediction model; In response to the correction request, collecting dynamic data of the grinding process; Based on the historical process test data and its historical removal rate, a matching relationship between the grinding stage and the sensitive feature set is established; Based on the matching relationship between the grinding stage and the sensitive feature set, determining the sensitive feature set to be measured of the grinding process dynamic data in its grinding stage to be measured; inputting the feature parameters to be measured corresponding to the sensitive feature set to be measured in the grinding process dynamic data into the trained XGBoost dynamic correction model to obtain a dynamic correction amount; Based on the dynamic correction amount, updating the empirical coefficient in the empirical material removal rate prediction model to obtain a corrected empirical material removal rate prediction model; The removal rate prediction value corresponding to the dynamic data of the grinding process is calculated using the revised empirical material removal rate prediction model.

2. The method for predicting the material removal rate of a silicon carbide substrate according to claim 1, wherein: Based on the historical process test data and its historical removal rate, a matching relationship between the grinding stage and the sensitive feature set is established, including: Grinding and dividing the historical process test data according to a preset physical threshold rule; A weighted approach using the Pearson correlation coefficient and mutual information analysis is employed to construct a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate.

3. The method for predicting the material removal rate of a silicon carbide substrate according to claim 2, wherein: Before the step of determining the sensitive feature set to be measured of the grinding stage to be measured of the grinding process dynamic data based on the matching relationship between the grinding stage and the sensitive feature set, the method further includes: According to the physical threshold rule, the dynamic data of the grinding process is ground and divided to determine the grinding stage to be measured.

4. The method for predicting the material removal rate of a silicon carbide substrate according to claim 1, wherein: Based on the historical process test data and its historical removal rate, a matching relationship between the grinding stage and the sensitive feature set is established, including: respectively calculating the Pearson coefficient and the mutual information value between the historical characteristic parameters of each feature at different grinding stages and the historical removal rate in the historical process test data; Normalizing all the mutual information values, and calculating a weighted sensitivity score of each of the features at different grinding stages based on the Pearson coefficient and the normalized mutual information values; According to the magnitude relationship of the sensitivity scores, a preset number of features are selected to form corresponding sensitive feature sets, so as to form a matching relationship between the grinding stage and the sensitive feature set.

5. The method for predicting the grinding material removal rate of a silicon carbide substrate according to claim 1, wherein: The trained XGBoost dynamic correction model includes multiple dynamic correction sub-models at different stages; inputting the measured feature parameters corresponding to the measured sensitive feature set in the dynamic data of the grinding process into the trained XGBoost dynamic correction model to obtain the dynamic correction amount, including: Based on the grinding stage to be tested, extracting the feature parameters to be tested corresponding to the sensitive feature set to be tested from the dynamic data of the grinding process; The characteristic parameter to be measured is input into the dynamic correction sub-model of the corresponding stage to obtain the dynamic correction amount.

6. The method for predicting the grinding material removal rate of a silicon carbide substrate according to claim 1, wherein: The empirical material removal rate prediction model is: ； in, is the predicted removal rate, is the first empirical coefficient, is the second empirical coefficient, is the third empirical coefficient, is the fourth empirical coefficient, Grinding pressure, is the grinding speed, is the abrasive particle size.

7. The method for predicting the material removal rate of a silicon carbide substrate according to claim 6, wherein: The training steps of the trained XGBoost dynamic correction model include: Determining a sample removal rate prediction value corresponding to the dynamic training sample data of the grinding process by using the empirical material removal rate prediction model; The ratio of the actual sample removal rate of the dynamic training sample data of the grinding process to the corresponding sample removal rate prediction value is used as the removal rate ratio; The sensitive features of the training samples are used as input, and the difference between the removal rate ratio and the first empirical coefficient is used as a label to train the constructed XGBoost dynamic correction model to obtain a target XGBoost dynamic correction model.

8. A device for predicting the removal rate of polishing material of a silicon carbide substrate, characterized in that: include: The model building module is used to fit the acquired historical process test data and build an empirical material removal rate prediction model; an acquisition module, configured to acquire dynamic data of the grinding process in response to a correction request; A dynamic correction amount determination module constructs a matching relationship between the grinding stage and the sensitive feature set based on the historical process test data and its historical removal rate; Based on the matching relationship between the grinding stage and the sensitive feature set, determining the sensitive feature set to be measured of the grinding process dynamic data in its grinding stage to be measured; inputting the feature parameters to be measured corresponding to the sensitive feature set to be measured in the grinding process dynamic data into the trained XGBoost dynamic correction model to obtain a dynamic correction amount; a correction module, configured to update the empirical coefficients in the empirical material removal rate prediction model based on the dynamic correction amount to obtain a corrected empirical material removal rate prediction model; The prediction module is used to calculate the removal rate prediction value corresponding to the dynamic data of the grinding process through the modified empirical material removal rate prediction model.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 7 is executed.

Images