Wafer polishing monitoring control method based on multi-parameter coupling

Abstract

The application discloses a wafer grinding monitoring control method based on multi-parameter coupling and relates to the technical field of wafer grinding monitoring. The wafer grinding monitoring control method comprises the following steps: collecting, in real time, grinding liquid parameters, grinding surface parameters and grinding head operation parameters in a wafer grinding process; obtaining a dynamic matching degree between the grinding liquid, the grinding surface and the grinding head based on the grinding liquid parameters, the grinding surface parameters and the grinding head operation parameters; inputting the dynamic matching degree into a pre-constructed grinding damage prediction model to obtain a grinding damage prediction risk; analyzing an optimal grinding head moving speed based on the grinding damage prediction risk; and adjusting a current grinding moving speed based on the optimal grinding head moving speed. The wafer grinding monitoring control method can dynamically calculate an optimal speed according to the damage risk, can actively inhibit surface damage accumulation while ensuring grinding efficiency, can realize the collaborative optimization of grinding quality and processing efficiency, and can improve the monitoring efficiency and process stability of the wafer grinding process.

Description

Technical Field

[0001] This invention relates to the field of wafer grinding monitoring technology, and in particular to a wafer grinding monitoring and control method based on multi-parameter coupling. Background Technology

[0002] Wafer grinding is a critical process in semiconductor manufacturing. Its processing quality directly affects the surface flatness, defect density, and subsequent yield of the wafer. During the grinding process, any mismatch in any link can lead to grinding abnormalities. In particular, the moving speed of the grinding head relative to the wafer directly determines the intensity of the mechanical action, which can lead to local over-grinding, surface scratches, or accumulation of surface damage, ultimately affecting wafer performance. Therefore, monitoring the wafer grinding process is extremely important.

[0003] Existing technologies for monitoring the wafer grinding process mainly involve determining the fixed moving speed for each grinding stage through numerous experiments during the process development phase, forming a process formula, and executing the process at a constant speed according to the formula during production. However, in actual grinding, the consumption and dilution of the grinding slurry, the wear and glazing of the grinding pad, and the continuous changes in the wafer surface morphology during the planarization process are all factors that this approach treats as independent of each other. It ignores the comprehensive impact of the dynamic coupling between the grinding slurry, surface condition, and grinding head movement on the grinding quality, which can easily lead to misjudgments when operating conditions fluctuate. In addition, most existing methods focus on post-event detection of damage that has already occurred, lacking the ability to predict damage trends in real time and intervene in advance. This results in predictions lagging behind changes in process conditions, making it impossible to optimize grinding parameters in a timely manner before damage accumulation becomes significant. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a wafer grinding monitoring and control method based on multi-parameter coupling, comprising the following specific steps: Step 1: Real-time acquisition of polishing slurry parameters, polishing surface parameters, and polishing head operating parameters during the wafer polishing process; Step 2: Obtain the dynamic matching degree between the polishing slurry, polishing surface, and polishing head based on the polishing slurry parameters, polishing surface parameters, and polishing head operating parameters; Step 3: Input the dynamic matching degree into the pre-built grinding damage prediction model to obtain the grinding damage prediction risk; Step 4: Analyze the optimal grinding head movement speed based on the predicted grinding damage risk, and adjust the current grinding speed accordingly.

[0005] Preferably, step one includes the following specific steps: Real-time acquisition of polishing slurry parameters, polishing surface parameters, and polishing head operating parameters during the wafer polishing process; The grinding fluid parameters include grinding fluid flow rate, grinding fluid temperature, abrasive material hardness, and abrasive distribution state, wherein the abrasive distribution state includes the average abrasive particle size and the abrasive particle size distribution width. The grinding surface parameters include wafer surface crack distribution and wafer surface roughness, wherein the wafer surface crack distribution includes crack spacing, crack area and crack depth; The grinding head operating parameters include grinding head pressure, grinding head rotation speed, grinding head chuck hardness, and grinding head vibration speed.

[0006] Preferably, step two includes the following specific steps: The absolute values ​​of the differences between the slurry flow rate, slurry temperature, average abrasive particle size, and abrasive particle size distribution width and their corresponding standard values ​​are divided by the corresponding standard values ​​to obtain the slurry flow rate deviation value, slurry temperature deviation value, average abrasive particle size deviation value, and abrasive particle size distribution width deviation value. The first threat value is obtained by weighted summation of the slurry flow rate deviation value, slurry temperature deviation value, average abrasive particle size deviation value, and abrasive particle size distribution width deviation value. The absolute values ​​of the differences between the grinding head pressure, grinding head speed, and grinding head vibration speed and their corresponding standard values ​​are divided by the corresponding standard values ​​to obtain the grinding head pressure deviation value, grinding head speed deviation value, and grinding head vibration speed deviation value. The second threat value is obtained by weighted summation of the grinding head pressure deviation value, grinding head speed deviation value, and grinding head vibration speed deviation value. The overall threat value is obtained by weighted summation of the first and second threat values; The crack depth anomaly value and crack area anomaly value are obtained by dividing the crack depth and crack area by the corresponding standard values, respectively. The crack spacing anomaly value is obtained by dividing the crack spacing standard value by the crack spacing. The crack anomaly value is obtained by weighted summation of the crack depth anomaly value, crack area anomaly value and crack spacing anomaly value. The wafer surface roughness anomaly value is obtained by dividing the wafer surface roughness by the standard value of the wafer surface roughness, and the wafer surface fragility value is obtained by weighted summation of the wafer surface roughness anomaly value and the crack anomaly value. The wafer vulnerability factor is obtained by multiplying the wafer surface vulnerability value by the vulnerability impact coefficient and adding a value of 1. The wafer comprehensive anomaly value is obtained by multiplying the wafer vulnerability factor by the comprehensive threat value. The wafer grinding matching degree is obtained by subtracting the wafer comprehensive anomaly value from the value 1.

[0007] Preferably, step three includes the following specific steps: A grinding damage prediction model is constructed. The construction steps are as follows: the historical wafer grinding matching degree during the historical wafer grinding batch process is collected, and paired with the historical wafer surface damage index detected after the batch is completed to form a paired sample. The historical wafer surface damage index includes scratch density, indentation depth and thickness deviation value. The grinding damage prediction model adopts a fully connected feedforward neural network structure, including an input layer, a first hidden layer, a second hidden layer, and an output layer. The inputs of the grinding damage prediction model include wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness. The input layer contains 3 neurons, corresponding to wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness, respectively. The first hidden layer contains 16 neurons, the second hidden layer contains 8 neurons, and the output layer contains 1 neuron, which is used to output the grinding damage risk value. The wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness are input into the grinding damage prediction model. After activation of each hidden layer, the grinding damage risk value is output.

[0008] Preferably, step four includes the following specific steps: Obtain a preset grinding damage risk threshold. When the grinding damage risk value is greater than or equal to the preset grinding damage risk threshold, the excess damage space is obtained by subtracting the preset grinding damage risk threshold from the grinding damage risk value. The optimal grinding head movement speed is obtained by subtracting the product of the damage space and the damage influence coefficient from the value of 1, and then multiplying it by the maximum safe movement speed of the grinding head. The speed difference is obtained by subtracting the current actual speed from the optimal grinding head speed. The speed difference is then limited by a single cycle. The limited speed difference is used as the grinding speed adjustment amount and input into the grinding head drive unit to adjust the current actual speed to the optimal grinding head speed.

[0009] Secondly, the present invention provides a wafer grinding monitoring and control system based on multi-parameter coupling, comprising: The parameter acquisition module is used to acquire the parameters of the polishing slurry, the polishing surface, and the polishing head operating parameters in real time during the wafer polishing process; The dynamic matching analysis module is used to obtain the dynamic matching degree between the polishing slurry, polishing surface and polishing head based on the polishing slurry parameters, polishing surface parameters and polishing head operating parameters. The damage risk prediction module is used to input the dynamic matching degree into a pre-built grinding damage prediction model to obtain the grinding damage prediction risk. The grinding speed adjustment module is used to adjust the current grinding speed based on the grinding damage prediction risk analysis and the optimal grinding head speed.

[0010] Thirdly, the present invention provides a storage medium comprising stored instructions, wherein, when the instructions are executed, the device in which the storage medium is located is controlled to perform the wafer grinding monitoring and control method based on multi-parameter coupling as described above.

[0011] Fourthly, the present invention provides an electronic device, including a memory and one or more instructions, wherein one or more instructions are stored in the memory and configured to be executed by one or more processors as described above in the wafer grinding monitoring and control method based on multi-parameter coupling.

[0012] The beneficial effects of this invention are: by obtaining the dynamic matching degree between the grinding fluid, the grinding surface and the grinding head through real-time analysis, it overcomes the limitations of the prior art that only relies on a single parameter or a simple linear combination, and can more realistically reflect the actual grinding conditions. By inputting the dynamic matching degree into a pre-built grinding damage prediction model and taking the evolution trend of the matching degree as input, the risk level is predicted before the damage becomes explicit, so that the adjustment action occurs before the damage forms, thus improving the prediction accuracy. Based on the risk analysis of grinding damage prediction, the optimal grinding head movement speed is determined, and the current grinding speed is adjusted in real time according to the optimal speed. The optimal speed can be dynamically calculated according to the damage risk, which can actively suppress the accumulation of surface damage while ensuring grinding efficiency. This achieves synergistic optimization of grinding quality and processing efficiency, and improves the monitoring efficiency and process stability of the wafer grinding process. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A schematic diagram of the wafer grinding monitoring and control method based on multi-parameter coupling provided in an embodiment of the present invention; Figure 2 A schematic diagram of the wafer grinding monitoring and control method based on multi-parameter coupling provided in an embodiment of the present invention; Figure 3 A schematic diagram of the wafer grinding monitoring and control method based on multi-parameter coupling provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the wafer grinding monitoring and control system based on multi-parameter coupling provided in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 This invention provides a wafer grinding monitoring and control method based on multi-parameter coupling, comprising the following specific steps: Step 1: Real-time acquisition of polishing slurry parameters, polishing surface parameters, and polishing head operating parameters during the wafer polishing process; In this embodiment, step one includes the following specific steps: Real-time acquisition of polishing slurry parameters, polishing surface parameters, and polishing head operating parameters during the wafer polishing process; The parameters of the polishing fluid include the polishing fluid flow rate, polishing fluid temperature, abrasive material hardness and abrasive distribution state. The abrasive distribution state includes the average abrasive particle size and the abrasive particle size distribution width. The grinding surface parameters include the wafer surface crack distribution and wafer surface roughness. The wafer surface crack distribution includes crack spacing, crack area and crack depth. The operating parameters of the grinding head include grinding head pressure, grinding head rotation speed, grinding head chuck hardness, and grinding head vibration speed.

[0017] In this embodiment, the grinding fluid flow rate is collected in real time by an electromagnetic flow meter installed in series in the grinding fluid supply pipeline; The temperature of the polishing slurry is collected in real time by a PT100 platinum resistance temperature sensor placed at the outlet of the polishing slurry nozzle. The hardness of the grinding head chuck is determined by testing the surface of the grinding head chuck with a hardness tester (Rockwell hardness) during the grinding head installation stage. The average abrasive particle size and the abrasive particle size distribution width are collected in real time by an online laser particle size analyzer connected in the side of the grinding fluid return line. The abrasive particle size distribution width is defined as the difference between D90 and D10 of the particle size distribution curve. This definition reflects the dispersion of the abrasive particle size. The larger the width, the more uneven the particle size. The distribution of cracks and the surface roughness of the wafer are obtained by optically scanning the wafer surface during the intervals of the grinding head lifting using a high-resolution industrial camera and a confocal displacement sensor arranged next to the grinding station. The crack spacing is the average spacing between adjacent cracks per unit area, the crack area is the sum of the areas of all cracks per unit area, the crack depth is obtained by scanning along the crack contour using the confocal displacement sensor to obtain the maximum depth value, and the surface roughness of the wafer is calculated by the arithmetic mean deviation of the contour of the confocal displacement sensor in the sampling area. The pressure of the grinding head is collected in real time by a pressure sensor inside the pressurizing cylinder of the grinding head. The grinding head rotation speed is acquired in real time via the grinding head spindle encoder; The vibration velocity of the grinding head is obtained by integration using a triaxial accelerometer mounted on the grinding head housing; The hardness of abrasive materials is determined by checking the Mohs hardness of the material during the grinding fluid preparation stage, based on the type of abrasive material (silicon dioxide, alumina, diamond, etc.). This step collects multi-dimensional parameters to comprehensively characterize the state during the grinding process, improving the sensitivity of abnormality identification and the accuracy of damage prediction.

[0018] Step 2: Obtain the dynamic matching degree between the polishing slurry, polishing surface, and polishing head based on the polishing slurry parameters, polishing surface parameters, and polishing head operating parameters; Please see Figure 2 In this embodiment, step two includes the following specific steps: The absolute values ​​of the differences between the slurry flow rate, slurry temperature, average abrasive particle size, and abrasive particle size distribution width and their corresponding standard values ​​are divided by the corresponding standard values ​​to obtain the slurry flow rate deviation value, slurry temperature deviation value, average abrasive particle size deviation value, and abrasive particle size distribution width deviation value. The first threat value is obtained by weighted summation of the slurry flow rate deviation value, slurry temperature deviation value, average abrasive particle size deviation value, and abrasive particle size distribution width deviation value. The absolute values ​​of the differences between the grinding head pressure, grinding head speed, and grinding head vibration speed and their corresponding standard values ​​are divided by the corresponding standard values ​​to obtain the grinding head pressure deviation value, grinding head speed deviation value, and grinding head vibration speed deviation value. The second threat value is obtained by weighted summation of the grinding head pressure deviation value, grinding head speed deviation value, and grinding head vibration speed deviation value. The overall threat value is obtained by weighted summation of the first and second threat values; The crack depth anomaly value and crack area anomaly value are obtained by dividing the crack depth and crack area by the corresponding standard values, respectively. The crack spacing anomaly value is obtained by dividing the crack spacing standard value by the crack spacing. The crack anomaly value is obtained by weighted summation of the crack depth anomaly value, crack area anomaly value and crack spacing anomaly value. It should be noted that the deeper the crack and the larger the crack area, the more severe the wafer damage and the higher the degree of anomaly. Conversely, the larger the crack spacing and the sparser the crack distribution, the lighter the wafer damage and the lower the degree of anomaly. When the cracks are so dense that the spacing cannot be distinguished, the measured crack spacing is zero. At this time, the measured crack spacing is forcibly taken as the preset lower limit value (0.1 in this embodiment) before calculating the crack spacing anomaly value. An upper limit value is set for the crack spacing anomaly value (10 in this embodiment) to limit the excessive amplification effect of extremely dense cracks on the results. When the upper limit is exceeded, the upper limit value is taken. The wafer surface roughness anomaly value is obtained by dividing the wafer surface roughness by the standard value of the wafer surface roughness, and the wafer surface fragility value is obtained by weighted summation of the wafer surface roughness anomaly value and the crack anomaly value. The wafer vulnerability factor is obtained by multiplying the wafer surface vulnerability value by the vulnerability impact coefficient and adding a value of 1. The wafer comprehensive anomaly value is obtained by multiplying the wafer vulnerability factor by the comprehensive threat value. The wafer grinding matching degree is obtained by subtracting the wafer comprehensive anomaly value from the value 1. The wafer grinding matching degree ranges from [0,1]. A matching degree of 0 indicates a complete mismatch, and a matching degree of 1 indicates a complete match.

[0019] In this embodiment, the standard values ​​(grinding slurry flow rate, grinding slurry temperature, average abrasive particle size, abrasive particle size distribution width, grinding head pressure, grinding head rotation speed, grinding head vibration speed, crack depth, crack area, crack spacing, and wafer surface roughness) are obtained as follows: During the wafer grinding process trial stage, under the conditions of the same type of wafer, the same grinding slurry formula, and the same grinding head configuration, no less than thirty batches of grinding tests are conducted. The time-series data of each parameter during the grinding process of each batch are collected, and the average value of each parameter in the stable grinding stage within the batch is taken as the representative value of that batch. Then, the arithmetic mean of the representative values ​​of the thirty batches is taken as the corresponding standard value. This step can more accurately reflect the positive feedback effect of surface damage accumulation and subsequent damage aggravation during the grinding process.

[0020] Step 3: Input the dynamic matching degree into the pre-built grinding damage prediction model to obtain the grinding damage prediction risk; Please see Figure 3 In this embodiment, step three includes the following specific steps: To construct a grinding damage prediction model, the construction steps are as follows: collect historical wafer grinding matching degree during the historical wafer grinding batch process, and form a paired sample with the historical wafer surface damage index detected after the batch ends. The historical wafer surface damage index includes scratch density, indentation depth and thickness deviation value. The grinding damage prediction model adopts a fully connected feedforward neural network structure, including an input layer, a first hidden layer, a second hidden layer, and an output layer. The inputs of the grinding damage prediction model include wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness. The input layer contains 3 neurons, corresponding to wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness, respectively. The first hidden layer contains 16 neurons, the second hidden layer contains 8 neurons, and the output layer contains 1 neuron, which is used to output the grinding damage risk value. In this embodiment, the grinding damage prediction model adopts a fully connected feedforward neural network (MLP) structure, and the specific steps are as follows: The input layer contains 3 neurons, corresponding to wafer grinding matching degree (value range [0,1]), abrasive material hardness (normalized to [0,1]), and grinding head chuck hardness (normalized to [0,1]), respectively; the first hidden layer contains 16 neurons with ReLU activation function, and is connected to the BatchNormalization layer and Dropout layer (Dropout rate 0.2); the second hidden layer contains 8 neurons with ReLU activation function, and is connected to the BatchNormalization layer; the output layer contains 1 neuron with Sigmoid activation function, outputting the grinding damage risk value, with a value range [0,1], where a larger value indicates a higher damage risk; The loss function of the grinding damage prediction model uses a combination of mean squared error loss and L2 regularization term, and the calculation formula is as follows: ,in, The value of the loss function. For the sample size, Let be the predicted grinding damage risk value of the model for the i-th sample. For the true damage label of the i-th sample, is the L2 regularization coefficient, with a value of 0.001. Let j be the trainable weight in the neural network. The L2 norm is used, i ranges from 1 to N, j traverses all trainable weights in the neural network, the mean squared error term in the loss function has a weight of 1, and the L2 regularization coefficient is 0.001. The two are combined by addition. The weight of the regularization term is much smaller than the weight of the error term, which ensures the model fitting accuracy and prevents overfitting. Retrieve complete process data for no less than 500 historical grinding batches from the historical production database of the wafer grinding production line. Each batch includes: the wafer grinding matching degree time sequence data calculated according to the method in step two during the grinding process of that batch, with the matching degree of the last sampling period before the end of the batch as the matching degree feature of that batch; the hardness of the abrasive material and the hardness of the grinding head chuck used in that batch (retrieved from the process formula database); and the wafer surface damage index obtained by testing equipment (white light interferometer to measure scratch density and indentation depth, and thickness gauge to measure thickness deviation value) after the grinding of that batch is completed. The labeling method for training samples (true value of grinding damage risk) is as follows: divide the scratch density, indentation depth and thickness deviation values ​​in the historical wafer surface damage index by their respective preset damage thresholds to obtain three normalized damage scores, and then sum the three normalized damage scores by weight to obtain the comprehensive damage index of the batch. The comprehensive damage index is mapped to the [0,1] interval through Min-Max normalization and used as the true label of the batch of samples. When the comprehensive damage index is greater than or equal to 1, the label is 1 (severe damage), and when the comprehensive damage index is equal to 0, the label is 0 (no damage). The specific training process of the grinding damage prediction model is as follows: 500 batch samples are divided into training, validation, and test sets in a 7:2:1 ratio. Stratified sampling is used to ensure a consistent proportion of samples for each damage level. The Adam optimizer is used with an initial learning rate of 0.001 and a learning rate decay strategy of ReduceLROnPlateau. The learning rate is multiplied by 0.5 when the validation set loss does not decrease for 5 consecutive epochs, with a minimum learning rate limit of 1e-6. The batch size is set to 32. The maximum number of training epochs is set to 200, and an early stopping mechanism is set. Training is terminated early when the validation set loss does not decrease for 15 consecutive epochs, and the model weights with the lowest validation set loss are saved as the final model. The weights of each hidden layer are initialized using the He normal initialization method, and the bias term is initialized to 0. The training environment is based on the PyTorch 2.0 framework, with GPU-accelerated training. A single complete training session takes approximately 30 minutes. After training, the mean squared error (MSE) and coefficient of determination (R²) are calculated on the test set. 2 The requirement is that MSE is less than 0.01 (corresponding to an average deviation of less than 10% between the predicted risk value and the actual label, meeting the accuracy requirements for engineering applications) and R 2 A value greater than 0.85 (meeting the requirements for regression model validity) is required for deployment. This standard has been verified in the backtesting of historical data from the wafer grinding production line, ensuring that the damage prediction accuracy is not less than 90% after the model is deployed. Otherwise, the network structure should be adjusted or the training data expanded and the model should be retrained. The selection criteria for neural network structure parameters are as follows: the number of neurons in the input layer is consistent with the dimension of the input features; the selection of 16 neurons in the first hidden layer and 8 neurons in the second hidden layer is determined by using a grid search method among candidate structures (8-4, 16-8, 32-16, 64-32) with the lowest MSE on the validation set as the standard. Experiments have verified that the 16-8 structure achieves the optimal balance between fitting accuracy and generalization ability, and the output layer has 1 neuron corresponding to a single risk value output. The wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness are input into the grinding damage prediction model. After nonlinear activation of each hidden layer, the grinding damage risk value is output, with a value range of [0,1]. The larger the value, the higher the damage risk.

[0021] This step uses the wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness as inputs to a neural network. Through a data-driven approach, it learns the nonlinear mapping relationship between matching degree, material hardness, and damage risk. Compared to the linear threshold judgment of existing technologies, this step utilizes the nonlinear fitting capability of neural networks to capture the interaction effect between matching degree and material hardness (for example, under the same matching degree, the higher the abrasive hardness, the greater the damage risk), thus achieving accurate online damage risk prediction during the grinding process.

[0022] Step 4: Analyze the optimal grinding head movement speed based on the grinding damage prediction risk, and adjust the current grinding speed accordingly.

[0023] In this embodiment, step four includes the following specific steps: Obtain a preset grinding damage risk threshold. When the grinding damage risk value is greater than or equal to the preset grinding damage risk threshold, the excess damage space is obtained by subtracting the preset grinding damage risk threshold from the grinding damage risk value. The optimal grinding head movement speed is obtained by subtracting the product of the damage space and the damage influence coefficient from the value of 1, and then multiplying it by the maximum safe movement speed of the grinding head. During the grinding process, the grinding head simultaneously undergoes rotational motion (grinding head speed) and downward feed motion (grinding head speed). The maximum safe moving speed of the grinding head is obtained as follows: During the trial production stage of the grinding process, a stepped moving speed test is conducted using wafers of the same model. The moving speed is gradually increased from low to high. Three wafers are ground at each moving speed and the damage index is detected. The highest moving speed at which no damage exceeds the standard is taken as the maximum safe moving speed of the grinding head. In this embodiment, when the grinding damage risk value is equal to the preset grinding damage risk threshold (zero for exceeding the damage space), the optimal moving speed is equal to the maximum safe moving speed. When the damage space increases, the optimal moving speed decreases proportionally. It should be noted that when the calculated result of the optimal grinding head moving speed is less than the preset minimum moving speed (taken as 10% of the maximum safe moving speed, determined through grinding efficiency tests; when the moving speed is lower than 10% of the maximum safe moving speed, the grinding time for a single batch increases by more than three times, and the grinding efficiency drops severely, so 10% is taken as the lower limit), the preset minimum moving speed is taken as the optimal grinding head moving speed to avoid the grinding efficiency dropping severely due to the moving speed being too low. When the grinding damage risk value is less than the preset grinding damage risk threshold, it indicates that the current grinding state is safe. At this time, the optimal grinding head moving speed is the maximum safe moving speed of the grinding head, that is, no speed reduction adjustment is made, and the current efficient grinding state is maintained. The speed difference is obtained by subtracting the current actual speed from the optimal grinding head speed. The speed difference is then limited by a single cycle. The limited speed difference is used as the grinding speed adjustment amount and input into the grinding head drive unit to adjust the current actual speed to the optimal grinding head speed.

[0024] In this embodiment, the rule for limiting the single-cycle variation is as follows: a preset maximum single-cycle speed change is defined. When the absolute value of the speed difference does not exceed the preset maximum single-cycle speed change, the speed adjustment is equal to the speed difference. When the absolute value of the speed difference exceeds the preset maximum single-cycle speed change, the speed adjustment is taken as a positive / negative preset maximum single-cycle speed change (the sign is consistent with the speed difference). This limiting mechanism can avoid sudden speed changes within a single cycle from impacting the mechanical structure of the grinding head and the wafer surface, achieving a smooth transition. The method for obtaining the maximum change in single-cycle speed is as follows: take 5% of the maximum safe speed of the grinding head as the initial value, perform a step response test on the grinding head drive unit, gradually increase the maximum change in single-cycle speed until the vibration speed of the grinding head exceeds the standard or the wafer surface is damaged by impact, and take the maximum value that does not exceed the standard as the final value of the maximum change in single-cycle speed. This step reduces the moving speed before damage actually occurs to ensure the grinding effect. At the same time, the limiting mechanism balances the timeliness of adjustment and mechanical safety, avoiding the secondary damage problem caused by sudden changes in moving speed in traditional methods. In this embodiment, the unified method for obtaining weights and thresholds is as follows: obtain historical batch wafer grinding data, including the grinding fluid parameters, grinding surface parameters, grinding head operating parameters, abrasive material hardness and grinding head chuck hardness of each historical batch, as well as a binary judgment result of whether damage exceeds the standard after wafer grinding. The damage exceeding the standard is judged based on whether the scratch density, indentation depth and thickness deviation values ​​detected offline exceed their respective preset damage thresholds. 1 indicates damage exceeding the standard and 0 indicates no exceeding the standard. Substitute the historical batch data into the method in step two to calculate the following values ​​for each batch: slurry flow rate deviation, slurry temperature deviation, abrasive particle size mean deviation, abrasive particle size distribution width deviation, grinding head pressure deviation, grinding head rotation speed deviation, grinding head vibration velocity deviation, crack depth anomaly, crack area anomaly, crack spacing anomaly, wafer surface roughness anomaly, first threat value, second threat value, overall threat value, wafer surface fragility value, wafer fragility factor, overall wafer anomaly value, and wafer grinding matching degree. Substitute the data from each historical batch into the method in step three to calculate the grinding damage risk value for each batch. The calculated deviation values, outliers, threat values, vulnerability values, and matching degrees of each component, as well as the calculated grinding damage risk value, are used as independent variable features, and whether the damage exceeds the standard is used as the dependent variable target to construct a logistic regression model. The parameters of the logistic regression model are solved by the maximum likelihood estimation method, which maximizes the log-likelihood between the predicted probability of damage exceeding the standard and the actual damage exceeding the standard label, and obtains the regression coefficients and intercept terms corresponding to the independent variable features. The decision boundary is extracted from the logistic regression model, and the regression coefficients are normalized and used as the weighting coefficients for the corresponding sub-items. The specific steps are as follows: After normalizing the regression coefficients corresponding to the first threat value, the four weighting weights for the first threat value are obtained by allocating them according to the internal proportions of the grinding fluid flow deviation value, grinding fluid temperature deviation value, abrasive particle size mean deviation value, and abrasive particle size distribution width deviation value; similarly, the three weighting weights for the second threat value, the two weighting weights for the comprehensive threat value, the three weighting weights for the crack anomaly value, the two weighting weights for the wafer surface fragility value, and the three weighting weights for the comprehensive damage index are obtained. The vulnerability impact coefficient and damage impact coefficient are set as the normalized values ​​of the regression coefficients corresponding to the wafer surface vulnerability value and the wafer comprehensive anomaly value in the logistic regression model; Set an optimal probability cut-off point (0.5 in this embodiment), convert the model's predicted probability into a category label, and the constant term of the linear equation corresponding to the optimal probability cut-off point is the preset grinding damage risk threshold. By adjusting this cut-off point on the validation set, the model accuracy is maximized, thereby determining the final preset grinding damage risk threshold and the final values ​​of each weighting coefficient and influence coefficient. The preset damage threshold is obtained as follows: During the wafer grinding process trial stage, no less than 30 batches of grinding tests are conducted using wafers of the same model. The scratch density, indentation depth and thickness deviation values ​​are collected after each batch of grinding is completed. The maximum value of each index within the batch is taken, and the 95th percentile of the maximum value of the 30 batches is taken as the preset damage threshold of the corresponding index.

[0025] Please see Figure 4The present invention also provides a wafer grinding monitoring and control system based on multi-parameter coupling, including: The parameter acquisition module is used to acquire the parameters of the polishing slurry, the polishing surface, and the polishing head operating parameters in real time during the wafer polishing process; The dynamic matching analysis module is used to obtain the degree of dynamic matching between the grinding fluid, the grinding surface, and the grinding head based on the grinding fluid parameters, grinding surface parameters, and grinding head operating parameters. The damage risk prediction module is used to input the dynamic matching degree into a pre-built grinding damage prediction model to obtain the grinding damage prediction risk. The grinding speed adjustment module is used to optimize the grinding head speed based on the grinding damage prediction risk analysis, and then adjust the current grinding speed based on the optimal grinding head speed.

[0026] This invention also provides a storage medium, which includes stored instructions, wherein, when the instructions are executed, the device where the storage medium is located is controlled to perform the wafer grinding monitoring and control method based on multi-parameter coupling as described above.

[0027] This invention also provides an electronic device, specifically including a memory and one or more instructions, wherein one or more instructions are stored in the memory and configured to be executed by one or more processors as described above in the wafer grinding monitoring and control method based on multi-parameter coupling.

[0028] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined in this invention may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A wafer grinding monitoring and control method based on multi-parameter coupling, characterized in that, The specific steps include the following: Step 1: Real-time acquisition of polishing slurry parameters, polishing surface parameters, and polishing head operating parameters during the wafer polishing process; Step 2: Obtain the dynamic matching degree between the polishing slurry, polishing surface, and polishing head based on the polishing slurry parameters, polishing surface parameters, and polishing head operating parameters; Step 3: Input the dynamic matching degree into the pre-built grinding damage prediction model to obtain the grinding damage prediction risk; Step 4: Based on the predicted grinding damage risk analysis, determine the optimal grinding head movement speed, and adjust the current grinding speed accordingly. Step 2 includes the following specific steps: The absolute values ​​of the differences between the slurry flow rate, slurry temperature, average abrasive particle size, and abrasive particle size distribution width and their corresponding standard values ​​are divided by the corresponding standard values ​​to obtain the slurry flow rate deviation value, slurry temperature deviation value, average abrasive particle size deviation value, and abrasive particle size distribution width deviation value. The first threat value is obtained by weighted summation of the slurry flow rate deviation value, slurry temperature deviation value, average abrasive particle size deviation value, and abrasive particle size distribution width deviation value. The absolute values ​​of the differences between the grinding head pressure, grinding head speed, and grinding head vibration speed and their corresponding standard values ​​are divided by the corresponding standard values ​​to obtain the grinding head pressure deviation value, grinding head speed deviation value, and grinding head vibration speed deviation value. The second threat value is obtained by weighted summation of the grinding head pressure deviation value, grinding head speed deviation value, and grinding head vibration speed deviation value. The overall threat value is obtained by weighted summation of the first and second threat values; The crack depth anomaly value and crack area anomaly value are obtained by dividing the crack depth and crack area by the corresponding standard values, respectively. The crack spacing anomaly value is obtained by dividing the crack spacing standard value by the crack spacing. The crack anomaly value is obtained by weighted summation of the crack depth anomaly value, crack area anomaly value and crack spacing anomaly value. The wafer surface roughness anomaly value is obtained by dividing the wafer surface roughness by the standard value of the wafer surface roughness, and the wafer surface fragility value is obtained by weighted summation of the wafer surface roughness anomaly value and the crack anomaly value. The wafer vulnerability factor is obtained by multiplying the wafer surface vulnerability value by the vulnerability impact coefficient and adding a value of 1. The wafer comprehensive anomaly value is obtained by multiplying the wafer vulnerability factor by the comprehensive threat value. The wafer grinding matching degree is obtained by subtracting the wafer comprehensive anomaly value from the value 1.

2. The wafer grinding monitoring and control method based on multi-parameter coupling according to claim 1, characterized in that, Step three includes the following specific steps: A grinding damage prediction model is constructed. The construction steps are as follows: the historical wafer grinding matching degree during the historical wafer grinding batch process is collected, and paired with the historical wafer surface damage index detected after the batch is completed to form a paired sample. The historical wafer surface damage index includes scratch density, indentation depth and thickness deviation value. The grinding damage prediction model adopts a fully connected feedforward neural network structure, including an input layer, a first hidden layer, a second hidden layer, and an output layer. The inputs of the grinding damage prediction model include wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness. The input layer contains 3 neurons, corresponding to wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness, respectively. The first hidden layer contains 16 neurons, the second hidden layer contains 8 neurons, and the output layer contains 1 neuron. The wafer grinding matching degree, abrasive material hardness, and grinding head chuck hardness are input into the grinding damage prediction model. After activation of each hidden layer, the grinding damage risk value is output.

3. The wafer grinding monitoring and control method based on multi-parameter coupling according to claim 2, characterized in that, Step four includes the following specific steps: Obtain a preset grinding damage risk threshold. When the grinding damage risk value is greater than or equal to the preset grinding damage risk threshold, the excess damage space is obtained by subtracting the preset grinding damage risk threshold from the grinding damage risk value. The optimal grinding head movement speed is obtained by subtracting the product of the damage space and the damage influence coefficient from the value of 1, and then multiplying it by the maximum safe movement speed of the grinding head. The speed difference is obtained by subtracting the current actual speed from the optimal grinding head speed. The speed difference is then limited by a single cycle. The limited speed difference is used as the grinding speed adjustment amount and input into the grinding head drive unit to adjust the current actual speed to the optimal grinding head speed.

4. The wafer grinding monitoring and control method based on multi-parameter coupling according to claim 3, characterized in that, Step one includes the following specific steps: Real-time acquisition of polishing slurry parameters, polishing surface parameters, and polishing head operating parameters during the wafer polishing process; The grinding fluid parameters include grinding fluid flow rate, grinding fluid temperature, abrasive material hardness, and abrasive distribution state, wherein the abrasive distribution state includes the average abrasive particle size and the abrasive particle size distribution width. The grinding surface parameters include wafer surface crack distribution and wafer surface roughness, wherein the wafer surface crack distribution includes crack spacing, crack area and crack depth; The grinding head operating parameters include grinding head pressure, grinding head rotation speed, grinding head chuck hardness, and grinding head vibration speed.

5. A wafer grinding monitoring and control system based on multi-parameter coupling, used to implement the wafer grinding monitoring and control method based on multi-parameter coupling as described in any one of claims 1-4, characterized in that, include: The parameter acquisition module is used to acquire the parameters of the polishing slurry, the polishing surface, and the polishing head operating parameters in real time during the wafer polishing process. The dynamic matching analysis module is used to obtain the dynamic matching degree between the polishing slurry, polishing surface and polishing head based on the polishing slurry parameters, polishing surface parameters and polishing head operating parameters. The damage risk prediction module is used to input the dynamic matching degree into a pre-built grinding damage prediction model to obtain the grinding damage prediction risk. The grinding speed adjustment module is used to adjust the current grinding speed based on the grinding damage prediction risk analysis and the optimal grinding head speed.

6. A storage medium, characterized in that, The storage medium includes stored instructions, wherein, when the instructions are executed, the device containing the storage medium is controlled to perform the wafer grinding monitoring and control method based on multi-parameter coupling as described in any one of claims 1-4.

7. An electronic device, characterized in that, It includes a memory and one or more instructions, wherein one or more instructions are stored in the memory and configured to be executed by one or more processors as described in any one of claims 1-4.

Images