Wafer thinning parameter optimization method and system for semiconductor production

Abstract

The invention discloses a wafer thinning parameter optimization method and system for semiconductor production, and relates to the technical field of semiconductor manufacturing. The method comprises the following steps: collecting wafer characteristics of a wafer to be thinned, and setting a grinding efficiency weight and a grinding damage weight; carrying out iterative optimization by utilizing a grinding predictor to obtain an optimal grinding parameter, an optimal efficiency parameter and an optimal damage layer parameter prediction value; according to efficiency parameter distribution and damage layer parameter distribution generated in the grinding optimization process, a polishing damage weight is obtained, and a polishing precision weight is configured; in combination with a polishing predictor, polishing parameter optimization is carried out with the damage removal rate and the thickness precision as targets, and optimal polishing parameters are obtained; and finally, wafer thinning processing is executed according to the optimized grinding and polishing parameter sequence. According to the method, collaborative optimization of grinding and polishing two-stage parameters is achieved, the machining efficiency is improved, meanwhile, a damaged layer is effectively controlled, the thickness precision is guaranteed, and the overall performance and stability of the process are enhanced.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and more specifically to a method and system for optimizing wafer thinning parameters in semiconductor production. Background Technology

[0002] As semiconductor devices continue to evolve towards smaller sizes, higher integration, and stronger performance, chip manufacturing places increasingly stringent demands on wafer thickness control. Wafer thinning, a critical step in back-end processes, uses a combination of grinding and polishing to reduce the wafer substrate thickness to a target value to meet packaging requirements. This process typically begins with grinding to achieve rapid and substantial material removal, followed by polishing to refine the surface and eliminate subsurface damage layers introduced by grinding. However, effectively controlling the depth of the damage layer and achieving a precise final thickness while maintaining high removal efficiency remains a core challenge in wafer thinning processes.

[0003] Currently, wafer thinning parameter settings in industrial production largely rely on fixed process formulations or operator experience, typically treating grinding and polishing as two isolated stages with fixed process parameters. However, wafer materials vary in characteristics such as initial thickness and target thickness, making fixed parameter combinations unsuitable for all situations. This can lead to low grinding efficiency, excessively deep or unevenly distributed damage layers. Furthermore, due to a lack of consideration for the correlation between grinding results and subsequent polishing requirements, fixed polishing parameters may fail to effectively remove specific damage from previous processes, or over-polishing may result in uncontrolled thickness, ultimately affecting the wafer's mechanical strength, electrical properties, and overall yield. Summary of the Invention

[0004] This invention addresses the technical problem that fixed wafer thinning process parameters and fragmented optimization during grinding and polishing stages make it difficult to synergistically improve processing quality and efficiency. It provides a method and system for optimizing wafer thinning parameters in semiconductor manufacturing.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for optimizing wafer thinning parameters for semiconductor manufacturing, comprising: Collect the wafer features of the wafer to be thinned, and set grinding efficiency weight and grinding damage weight according to the wafer features; Based on the wafer characteristics, grinding efficiency weight, and grinding damage weight, grinding parameters are optimized to obtain the optimal grinding parameters, as well as the predicted optimal efficiency parameters and optimal damage layer parameters. Based on the distribution of efficiency parameters and damage layer parameters during the grinding parameter optimization process, the grinding damage weight is adjusted to obtain the polishing damage weight, and the polishing accuracy weight is configured to optimize the polishing parameters and obtain the optimal polishing parameters. The wafer is thinned according to the optimal grinding and polishing parameters.

[0006] Secondly, the present invention provides a wafer thinning parameter optimization system for semiconductor manufacturing, comprising: The parameter weight configuration module is used to collect the wafer features of the wafer to be thinned, and set the grinding efficiency weight and grinding damage weight according to the wafer features. The grinding parameter optimization module is used to optimize grinding parameters based on the wafer characteristics, grinding efficiency weight, and grinding damage weight to obtain the optimal grinding parameters, as well as the predicted optimal efficiency parameters and optimal damage layer parameters. The polishing parameter optimization module is used to adjust the grinding damage weight based on the efficiency parameter distribution and damage layer parameter distribution during the grinding parameter optimization process, obtain the polishing damage weight, configure the polishing accuracy weight, optimize the polishing parameters, and obtain the optimal polishing parameters. The thinning process execution module is used to perform thinning processes on the wafer according to the optimal grinding parameters and optimal polishing parameters.

[0007] The beneficial effects of this invention are: Compared to existing technologies, this invention first dynamically sets initial optimization target weights based on the individual characteristics of the wafer to be processed, enabling the parameter optimization process to have personalized adaptability and overcoming the shortcomings of fixed parameters that are difficult to match different processing requirements. Secondly, it uses a machine learning model to accurately predict the efficiency and damage of the grinding stage and find the optimal parameters, achieving a scientific trade-off between material removal rate and damage control in the grinding process. Thirdly, it innovatively utilizes the parameter distribution information generated during the grinding optimization process to adaptively adjust the optimization target weights of the subsequent polishing stage, establishing a synergistic relationship between the grinding and polishing stages, ensuring that the polishing process can effectively respond to and compensate for the specific damage conditions caused by the preceding grinding. Finally, by constructing a full-process parameter optimization and execution framework covering grinding and polishing, it improves the performance of single-step processes while enhancing the global optimality and consistency of the overall wafer thinning process chain, thereby simultaneously improving processing efficiency and effectively controlling subsurface damage while ensuring final thickness accuracy. Attached Figure Description

[0008] Figure 1 A schematic flowchart of a wafer thinning parameter optimization method for semiconductor manufacturing provided by the present invention; Figure 2This is a schematic diagram of a wafer thinning parameter optimization system for semiconductor manufacturing provided by the present invention.

[0009] In the attached diagram, the components represented by each number are as follows: Parameter weight configuration module 11, grinding parameter optimization module 12, polishing parameter optimization module 13, and thinning process execution module 14. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0012] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0013] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for optimizing wafer thinning parameters in semiconductor manufacturing, including: S10: Collect the wafer features of the wafer to be thinned, and set the grinding efficiency weight and grinding damage weight according to the wafer features; Specifically, the wafer features of the wafer to be thinned are collected, and based on these wafer features, grinding efficiency weights and grinding damage weights are set, including: The wafer features to be thinned are collected, including the original thickness and the target thickness. The thinning thickness is calculated based on the original thickness and the target thickness. Based on the thinning thickness, calculate the configuration grinding efficiency weight and grinding damage weight.

[0014] First, the wafer features to be thinned are collected. Specifically, the wafer features include the original thickness of the wafer and the target thickness to be achieved. Second, based on the obtained original and target thicknesses, mathematical calculations are performed to obtain the thinning thickness. The thinning thickness is the difference between the original thickness and the target thickness, and this value directly reflects the total amount of material to be removed during the grinding process.

[0015] Furthermore, based on the calculated thinning thickness, the grinding efficiency weight and grinding damage weight are dynamically configured. The grinding efficiency weight and grinding damage weight represent quantification coefficients indicating the relative importance of material removal efficiency and subsurface damage control during grinding parameter optimization, respectively. Their configuration logic follows these process requirements: when the thinning thickness is large, it means a heavy material removal task, and processing efficiency should be prioritized; therefore, a higher grinding efficiency weight is assigned. Simultaneously, since damage control is relatively less important at this stage, a lower grinding damage weight is assigned. This weight allocation method ensures that the parameter optimization process can adaptively adjust the emphasis on efficiency and damage control objectives according to the actual processing volume.

[0016] Specifically, based on the thinning thickness, the grinding efficiency weight and grinding damage weight are calculated and configured, including: Obtain the maximum thinning thickness within the wafer thinning process log; The ratio of the thinning thickness to the maximum thinning thickness is calculated and used as a weight for grinding efficiency; The grinding damage weight is calculated based on the grinding efficiency weight.

[0017] First, the maximum recorded thinning thickness is retrieved from the historical wafer thinning process logs. This maximum thinning thickness represents the case in which the process system previously handled the largest amount of material removal required, and can be used as a benchmark to measure the relative scale of the current processing task.

[0018] Secondly, the current thinning thickness of the wafer to be processed is compared with the maximum thinning thickness, and the ratio between the current thinning thickness and the maximum thinning thickness is calculated. This ratio is then determined as the value of the grinding efficiency weight. This calculation makes the grinding efficiency weight a normalized value between 0 and 1, intuitively reflecting the relative workload of the current processing task within the historical experience range. The higher the ratio, the closer the amount of material to be removed is to the historical limit, and the larger the grinding efficiency weight is. This guides the algorithm to focus more on improving the material removal rate in the subsequent grinding parameter optimization process to cope with the heavy material removal task.

[0019] Finally, based on the determined grinding efficiency weight, specifically, the difference between 1 and the grinding efficiency weight is defined as the grinding damage weight. This calculation ensures that the sum of the grinding efficiency weight and the grinding damage weight remains constant, equal to 1, thus forming a pair of mutually balancing weight factors in the optimization model. When the grinding efficiency weight approaches 1 due to a large reduction in thickness, the grinding damage weight correspondingly approaches 0, clearly indicating that the optimization algorithm should strongly favor improving efficiency; conversely, when the reduction in thickness is small, the grinding efficiency weight decreases while the grinding damage weight increases, guiding the optimization process to focus more on suppressing the damage layer.

[0020] S20: Based on the wafer characteristics, grinding efficiency weight, and grinding damage weight, optimize the grinding parameters to obtain the optimal grinding parameters, as well as the predicted optimal efficiency parameters and optimal damage layer parameters. Specifically, based on the wafer characteristics, grinding efficiency weight, and grinding damage weight, grinding parameters are optimized to obtain optimal grinding parameters, as well as predicted optimal efficiency parameters and optimal damage layer parameters, including: Obtain the grinding parameter space; The first grinding parameter is randomly set within the grinding parameter space; Based on the first grinding parameters and wafer characteristics, a first efficiency parameter and a first damage layer parameter are predicted and obtained, wherein the first efficiency parameter includes normalized grinding completion time data and the first damage layer parameter includes normalized damage layer depth data. Based on the grinding efficiency weight and grinding damage weight, combined with the first efficiency parameter and the first damage layer parameter, the first grinding cost of the first grinding parameter is calculated. The grinding parameters are iteratively optimized to obtain the optimal grinding parameters with the minimum grinding cost, and the optimal efficiency parameters and optimal damage layer parameters are obtained from the optimal grinding parameters and wafer feature prediction.

[0021] First, a predefined grinding parameter space is obtained. This space encompasses the range of all adjustable and optimizable parameters in the grinding process, such as key variables like wheel speed, feed rate, and coolant flow rate, used to define the search boundaries for the optimization algorithm. Specifically, this grinding parameter space is set based on the safety operating specifications of the process equipment, the physical limitations of the grinding wheel and wafer materials, and historical process experience data. For example, the upper and lower limits of the grinding wheel speed are determined based on the spindle motor performance and the strength of the grinding wheel structure, set to 3000 to 6000 revolutions per minute; the feed rate range must avoid breakage due to excessive speed or surface burns due to excessive slowness, set to 100 to 500 micrometers per minute; and the coolant flow rate must ensure the minimum requirements for sufficient heat dissipation and chip removal, as well as the maximum capacity of the supply system, set to 5 to 20 liters per minute.

[0022] Secondly, within the acquired grinding parameter space, a set of grinding parameters is randomly set, denoted as the first grinding parameter, as the starting point for the iterative optimization process. Then, the first grinding parameter and wafer features are input into a pre-trained grinding prediction model. This grinding prediction model, constructed based on historical processing data and machine learning methods, can simulate the input-output relationship of the grinding process. Its output includes a first efficiency parameter and a first damage layer parameter. The first efficiency parameter refers to the normalized grinding completion time data, reflecting the processing efficiency under this first grinding parameter; the first damage layer parameter refers to the normalized damage layer depth data, characterizing the degree of subsurface damage that may be introduced under this first grinding parameter.

[0023] It should be noted that the normalization process aims to eliminate the influence of different physical dimensions and numerical magnitudes on model training and optimization calculations. Preferably, the min-max normalization method can be used to linearly transform the original data to the 0-1 interval.

[0024] Specifically, based on the first grinding parameters and wafer characteristics, the first efficiency parameter and the first damage layer parameter are predicted and obtained, including: Based on the wafer grinding process logs over a historical period, a set of sample grinding parameters and a set of sample wafer features were collected. Grinding efficiency parameters, grinding damage layer parameters, and grinding wafer thickness under different sample grinding parameters and sample wafer features were obtained, and the set of sample efficiency parameters, sample damage layer parameters, and sample grinding wafer thickness were labeled to obtain the set of sample grinding parameters. A grinding predictor is built based on machine learning. The grinding predictor is trained under supervision using the sample grinding parameter set, sample wafer feature set, sample efficiency parameter set, sample damage layer parameter set, and sample grinding wafer thickness set. The training is completed after the test converges. The first grinding parameters and wafer features are input into the grinding predictor, and the first efficiency parameter, the first damage layer parameter, and the first grinding wafer thickness are output.

[0025] First, data was systematically collected from historical wafer grinding logs to construct training samples. Specifically, the collected data mainly consisted of three parts: 1) a set of actual grinding parameters used in the samples, such as the combination of grinding wheel speed, feed rate, and coolant flow rate in each processing cycle; 2) a set of corresponding sample wafer characteristics, such as the original thickness and material type of each wafer; and 3) the actual process results measured under different sample grinding parameters and sample wafer characteristics, including grinding efficiency parameters, grinding damage layer parameters, and the finished wafer thickness. Among these, the grinding efficiency parameter is an indicator used to quantify the material removal rate, typically expressed as the time required to complete grinding per unit thickness; the grinding damage layer parameter is an indicator used to characterize the depth of physical damage introduced into the wafer subsurface during grinding, such as the damage layer thickness measured by a specific detection method; and the finished wafer thickness represents the remaining physical thickness of the wafer after a single grinding cycle, a key dimension for calculating material removal and evaluating whether the process meets standards. The actual process results are organized and labeled to form a set of sample efficiency parameters, a set of sample damage layer parameters, and a set of sample grinding wafer thicknesses for supervised learning.

[0026] Secondly, an appropriate machine learning algorithm architecture is selected to construct a grinding predictor. This predictor is used to establish a complex mapping relationship between input variables and output variables. Specifically, the input variables are grinding parameters and wafer characteristics, while the output variables are the efficiency parameters, damage layer parameters, and the thickness of the ground wafer to be predicted. Machine learning algorithms refer to computational models and methods that can learn patterns from given data and make predictions or decisions; algorithms such as deep neural networks, gradient boosting decision trees, or support vector regression are all applicable.

[0027] Furthermore, the prepared sets of sample grinding parameters and sample wafer features are used as input features, while the sets of sample efficiency parameters, sample damage layer parameters, and sample ground wafer thicknesses are used as corresponding target labels, all of which are input into the grinding predictor. Through supervised training, the parameters of the internal model of the grinding predictor are continuously adjusted so that its output predictions gradually approach the true target label values. The training process continues until the prediction performance of the grinding predictor on the independent validation set reaches a stable convergence state, at which point the grinding predictor training is considered complete. The convergence condition is set according to preset model performance evaluation metrics and iteration stopping criteria. For example, when the mean squared error loss function value on the validation set decreases by less than one ten-thousandth over ten consecutive training epochs, or when the average absolute percentage error of the model's predictions for all output variables on the validation set is less than five percent, the model is considered to have converged.

[0028] For example, considering the complex nonlinear physical relationship between grinding process parameters, wafer material characteristics and grinding results, and the excellent ability of deep neural network models to fit high-dimensional nonlinear mappings and abstract potential laws, deep neural network models can be selected as the core technical architecture for building this grinding predictor.

[0029] Specifically, this grinding predictor employs a multi-input multi-output feedforward neural network structure, primarily composed of an input layer, a feature abstraction layer, and a parallel output layer. The input layer receives a standardized input feature vector, which is jointly composed of a set of sample grinding parameters and a set of sample wafer features. Specifically, it may include process parameters such as wheel speed, feed rate, and coolant flow rate, as well as feature dimensions such as the original wafer thickness and material type. The feature abstraction layer uses a fully connected neural network structure with two hidden layers. The number of neurons in each hidden layer is adaptively configured according to the total dimension of the input features and the task complexity, for example, 128 and 64 neurons respectively. Each neural network layer uses the ReLU activation function to introduce nonlinear transformation capability, and a Dropout layer is embedded after both hidden layers with a dropout rate of 0.3 to effectively suppress model overfitting and improve its generalization performance. The parallel output layer contains three independent linear output nodes, corresponding to the three target variables to be predicted: efficiency parameter, damage layer parameter, and ground wafer thickness.

[0030] During training, the key hyperparameters were set to a learning rate of 0.0005, a training epoch count of 200, and a batch size of 32. The learning rate was set to ensure stable convergence during training and avoid oscillations; the number of training epochs ensured the model had sufficient iteration opportunities to learn the complex mapping relationships in the process data; and the batch size balanced training efficiency with the stability of gradient updates. Specifically, a supervised learning training method was adopted, and the sample grinding parameter set, sample wafer feature set, sample efficiency parameter set, sample damage layer parameter set, and sample ground wafer thickness set were randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0031] Furthermore, the sample grinding parameters and wafer features are combined as the input feature vector, while the corresponding sample efficiency parameters, sample damage layer parameters, and sample ground wafer thickness are used as the supervision signal target vector. Using the backpropagation algorithm in conjunction with the Adam optimizer, all weight parameters of the network are iteratively optimized. The mean squared error loss function is used to comprehensively measure the overall deviation between the predicted three output variables and the actual target values. The training process is monitored in real time using a validation set. When the average absolute percentage error of the model's prediction results for the three output variables on the validation set is less than 5%, the model is considered to have converged, training is terminated, and the trained grinding predictor is obtained. This grinding predictor can effectively establish the complex nonlinear relationship between process parameters and wafer features and various processing results, providing accurate and rapid prediction capabilities for subsequent parameter optimization.

[0032] Finally, when evaluating the first grinding parameters, they are input along with the wafer features of the wafer to be processed into the pre-trained grinding predictor. The grinding predictor calculates and outputs the corresponding prediction results, namely the first efficiency parameter, the first damage layer parameter, and the first grinding wafer thickness.

[0033] Furthermore, based on the aforementioned grinding efficiency weight and grinding damage weight, the predicted first efficiency parameter and first damage layer parameter are comprehensively evaluated to calculate the first grinding cost corresponding to the first grinding parameter. The value of this grinding cost comprehensively reflects the overall performance of the set of grinding parameters in terms of both efficiency and damage control under the current weight preference. A higher grinding cost value indicates a worse overall performance of the set of grinding parameters under the given weight evaluation system, meaning that both the predicted efficiency and damage control effects are unsatisfactory or severely unbalanced. Conversely, a lower grinding cost value indicates a better performance of the set of grinding parameters in comprehensively balancing efficiency and damage control objectives, and a closer approximation to the ideal state under the current weight setting. Specifically, the first grinding cost = grinding efficiency weight × first efficiency parameter + grinding damage weight × first damage layer parameter.

[0034] Finally, an iterative optimization algorithm is initiated. This algorithm aims to minimize the grinding cost by continuously generating new combinations of grinding parameters within the grinding parameter space, repeatedly performing the prediction and cost calculation process. Through multiple iterations and comparisons, the set of grinding parameters that minimizes the grinding cost is finally found; this set of grinding parameters is determined as the optimal grinding parameters. Simultaneously, the optimal grinding parameters and wafer features are input again into the grinding prediction model to obtain the optimal efficiency parameters and optimal damage layer parameters. These optimal efficiency parameters and optimal damage layer parameters represent the theoretically achievable best efficiency performance and minimum damage level under given weights and wafer features.

[0035] S30: Based on the distribution of efficiency parameters and damage layer parameters during the grinding parameter optimization process, the grinding damage weight is adjusted to obtain the polishing damage weight, and the polishing accuracy weight is configured to optimize the polishing parameters and obtain the optimal polishing parameters. Specifically, based on the efficiency parameter distribution and damage layer parameter distribution during the grinding parameter optimization process, the grinding damage weight is adjusted to obtain the polishing damage weight, and a polishing accuracy weight is configured to optimize the polishing parameters and obtain the optimal polishing parameters, including: All efficiency parameters and damage layer parameters in the grinding parameter optimization process are obtained and arranged in ascending order to obtain the distribution of efficiency parameters and damage layer parameters. Based on the efficiency parameter distribution and the damage layer parameter distribution, the grinding damage weight is adjusted to obtain the polishing damage weight; Calculate the polishing accuracy weight based on the polishing damage weight; Based on the polishing damage weight and polishing accuracy weight, the polishing parameters are optimized to obtain the optimal polishing parameters.

[0036] First, all efficiency parameters and damage layer parameters generated during the iterative optimization of grinding parameters are obtained. These parameters are the predicted values ​​generated by the grinding prediction model when evaluating different combinations of grinding parameters. The efficiency and damage layer parameters are then arranged in ascending order of value to obtain their respective distributions. These distributions clearly reveal the theoretical ranges and central tendency of the two objectives—efficiency and damage—that can be achieved during grinding optimization.

[0037] Secondly, based on the distribution of efficiency parameters and damage layer parameters, the initially set grinding damage weights are adjusted to calculate polishing damage weights suitable for the optimization objectives of the subsequent polishing stage. The damage layer parameter distribution explored during the grinding parameter optimization process reflects the theoretical range from minimum to maximum damage that the grinding process may produce; while the efficiency parameter distribution reflects the theoretical range from best to worst efficiency that can be achieved. If the overall damage layer parameter distribution is high and the overall efficiency parameter distribution is low, it means that after grinding optimization, the residual damage may be severe while the space for efficiency improvement is limited, resulting in the subsequent polishing process needing to undertake a heavier damage removal task. Therefore, the adjustment of the polishing damage weights will be based on this distribution information, so that under the condition that the predicted damage is more severe and the efficiency is less than ideal in the grinding stage, the optimization objective of the polishing stage will focus more on damage removal, that is, assigning a larger value to the polishing damage weight.

[0038] Specifically, based on the efficiency parameter distribution and the damage layer parameter distribution, the grinding damage weight is adjusted to obtain the polishing damage weight, including: Calculate the midpoint value of the efficiency parameter distribution and the damage layer parameter distribution to obtain the midpoint efficiency parameter and the midpoint damage layer parameter; The ratio of the number of damage layer parameters greater than the midpoint damage layer parameter to the number of damage layer parameters less than the midpoint damage layer parameter within the damage layer parameter distribution is used as the first damage weight adjustment coefficient. The ratio of the number of efficiency parameters less than the midpoint efficiency parameter to the number of efficiency parameters greater than the midpoint efficiency parameter within the efficiency parameter distribution is calculated and used as the second damage weight adjustment coefficient. The damage weight adjustment coefficient is calculated based on the first damage weight adjustment coefficient and the second damage weight adjustment coefficient. The grinding damage weight is adjusted and calculated using the damage weight adjustment coefficient to obtain the polishing damage weight.

[0039] First, calculate the midpoint values ​​of the efficiency parameter distribution and the damage layer parameter distribution, respectively. For a distribution sequence already sorted by numerical value, the midpoint value is the median of the sequence. The specific calculation rule is as follows: Assume the sorted sequence contains n values. If n is odd, the median is the (n+1) / 2th value in the sequence; if n is even, the median is the arithmetic mean of the (n / 2)th and (n / 2+1)th values ​​in the sequence. For example, for a damage layer parameter distribution sorted in ascending order, with specific values ​​of 2µm, 3µm, 5µm, 6µm, and 8µm, the midpoint value of this damage layer parameter distribution is the value located at the center of the sequence, i.e., 5µm. Through this calculation, the midpoint efficiency parameter and midpoint damage layer parameter representing the center position of the distribution can be obtained.

[0040] Secondly, based on the distribution of damage layer parameters, a first damage weight adjustment coefficient is calculated. Specifically, the number of parameters whose values ​​are greater than the midpoint damage layer parameter is counted, and the number of parameters whose values ​​are less than the midpoint damage layer parameter is also counted. The ratio of the former to the latter is calculated, and this ratio is defined as the first damage weight adjustment coefficient. This first damage weight adjustment coefficient reflects the asymmetry in the distribution of damage layer parameters. If the ratio is greater than 1, it indicates that more damage layer parameter values ​​are located in the high damage range, meaning that the grinding optimization process tends to produce a relatively deep damage layer; conversely, if the ratio is less than 1, it indicates that the optimization process tends to produce a shallower damage layer.

[0041] Simultaneously, based on the efficiency parameter distribution, a second damage weight adjustment coefficient is calculated. Specifically, the number of parameters whose values ​​are less than the midpoint efficiency parameter is counted, and the number of parameters whose values ​​are greater than the midpoint efficiency parameter is also counted. The ratio of the former to the latter is calculated, and this ratio is defined as the second damage weight adjustment coefficient. This second damage weight adjustment coefficient reflects the asymmetry in the distribution of efficiency parameters. If the ratio is greater than 1, it indicates that more efficiency parameter values ​​are located in the low-efficiency range, meaning that the grinding optimization process faces significant difficulties in improving efficiency; conversely, if the ratio is less than 1, it indicates that the optimization process can more easily achieve higher efficiency.

[0042] Further, the final damage weight adjustment coefficient is calculated based on the first damage weight adjustment coefficient and the second damage weight adjustment coefficient. Preferably, the damage weight adjustment coefficient = (first damage weight adjustment coefficient + second damage weight adjustment coefficient) / 2. This damage weight adjustment coefficient simultaneously considers the skewness of the damage layer distribution and the skewness of the efficiency parameter distribution.

[0043] Finally, the calculated damage weight adjustment coefficient is used to adjust the original grinding damage weight, thereby obtaining the polishing damage weight. Specifically, the polishing damage weight = grinding damage weight × damage weight adjustment coefficient. After this adjustment, the polishing damage weight can be dynamically corrected based on the process potential and limitations revealed during the grinding parameter optimization exploration process. When the grinding optimization results show that the damage layer is difficult to control or the efficiency improvement space is limited, the damage weight adjustment coefficient increases, resulting in a corresponding increase in the polishing damage weight. This guides the polishing parameter optimization process to allocate more weight to the damage removal target to compensate for the shortcomings of the previous process.

[0044] Furthermore, based on the calculated polishing damage weight, the polishing accuracy weight is further calculated and configured. The polishing accuracy weight represents the degree of importance attached to the final wafer thickness control accuracy during the polishing parameter optimization process. Typically, there is an inverse relationship between the polishing damage weight and the polishing accuracy weight, with their sum remaining constant. Therefore, the polishing accuracy weight can be directly obtained by subtracting the polishing damage weight from 1, ensuring that the optimization objective of the polishing stage achieves a balance between effectively removing the damaged layer and accurately controlling the final thickness.

[0045] Finally, using the determined polishing damage weight and polishing accuracy weight, the optimization process of polishing parameters is initiated.

[0046] Specifically, polishing parameters are optimized to obtain the optimal polishing parameters, including: Obtain the polishing parameter space and generate the first polishing parameters; Based on the first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters, polishing prediction is performed to obtain the first damage removal rate and the first polished wafer thickness. Calculate the similarity between the thickness of the first polished wafer and the target thickness, and use it as the first polishing accuracy value; Based on the polishing damage weight and polishing precision weight, the first damage removal rate and the first polishing precision value are weighted and calculated to obtain the first polishing adaptability value; Continue iteratively optimizing the polishing parameters until convergence, obtaining the optimal polishing parameters with the largest polishing fitness value.

[0047] First, a polishing parameter space is obtained, which includes the range of all adjustable process parameters in the polishing process, such as polishing pad pressure, polishing slurry flow rate, polishing disc rotation speed, and polishing time. This polishing parameter space is set based on the mechanical and fluid control capabilities of the polishing equipment, the physicochemical properties of the polishing pad and polishing slurry, and an empirical operating window to ensure process stability. For example, the upper and lower limits of the polishing pad pressure are determined by the accuracy of the pressure actuator and the strength of the wafer structure, and are set to 10 kPa to 50 kPa per square centimeter; the polishing slurry flow rate range needs to balance the minimum and maximum values ​​of chemical action and heat dissipation requirements of the supply system, and is set to 100 ml to 500 ml per minute; the polishing disc rotation speed is limited based on the balance of centrifugal force and frictional heat, and is set to 30 rpm to 90 rpm; the polishing time is set within a reasonable range based on the typical material removal rate, and is set to 30 seconds to 300 seconds. Within this polishing parameter space, a set of polishing parameters is randomly generated, denoted as the first polishing parameter, as the starting point for the iterative optimization process.

[0048] Secondly, the first polishing parameters, along with the first ground wafer thickness and optimal damage layer parameters obtained from the grinding stage, are input into a pre-trained polishing prediction model. This polishing prediction model, built based on historical polishing data, can simulate the input-output relationship of the polishing process. Its prediction results include the first damage removal rate and the first polished wafer thickness. The first damage removal rate characterizes the percentage of damage layer thickness removed under the first polishing parameters; the first polished wafer thickness predicts the final remaining thickness of the wafer after the polishing process.

[0049] Specifically, based on the first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters, polishing prediction is performed to obtain the first damage removal rate and the first polished wafer thickness, including: Based on the polishing process logs over a historical period, a set of sample polishing parameters, a set of sample grinding wafer thicknesses, and a set of sample damage layer parameters were collected. Additionally, a set of sample damage removal rates and a set of sample polished wafer thicknesses were collected after polishing under different sample polishing parameters, sample grinding wafer thicknesses, and sample damage layer parameters. Each sample damage removal rate includes the percentage of damage layer removed by polishing. A polishing predictor is built based on machine learning. The polishing predictor is supervised and tested using the set of sample polishing parameters, the set of sample grinding wafer thickness, the set of sample damage layer parameters, the set of sample damage removal rate, and the set of sample polishing wafer thickness until the test is qualified. The first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters are input into the polishing predictor, and the first damage removal rate and the first polished wafer thickness are output.

[0050] First, data is systematically collected from polishing process logs accumulated over historical periods to construct training samples. Specifically, the collected data includes two parts: input features and target labels. Input features include three parts: first, a set of actual polishing parameters used in the samples, such as the combination of polishing pressure, polishing fluid flow rate, polishing disc speed, and polishing time in each polishing process; second, a set of sample grinding wafer thicknesses before polishing; and third, a set of known sample damage layer parameters before polishing. Target labels are two types of process results actually measured after the polishing process under different sample polishing parameters, sample grinding wafer thicknesses, and sample damage layer parameter characteristics: first, a set of sample damage removal rates, where each sample damage removal rate specifically refers to the percentage of known damage layer thickness removed in that polishing process; and second, a set of sample polished wafer thicknesses after polishing, specifically referring to the actual thickness value obtained by physical measurement of the final wafer thickness after the polishing process, directly reflecting the final thickness control result of the polishing process.

[0051] Secondly, a suitable machine learning algorithm framework is selected to construct a polishing predictor. This predictor aims to establish a complex nonlinear mapping relationship between input features and output targets. Specifically, the input features are polishing parameters, grinding wafer thickness, and damage layer parameters, while the output targets are the damage removal rate and polished wafer thickness to be predicted. Machine learning algorithms refer to mathematical models and computational methods that can automatically learn patterns from multi-dimensional data and make predictions; for example, random forest regression, deep feedforward neural networks, or Gaussian process regression are all applicable.

[0052] Specifically, the set of sample polishing parameters, the set of sample grinding wafer thicknesses, and the set of sample damage layer parameters are used as input feature data, while the corresponding set of sample damage removal rates and the set of sample polished wafer thicknesses are used as target label data. Supervised learning is employed to train the polishing predictor. During training, the internal parameters of the model are continuously adjusted through optimization algorithms to minimize the error between the predicted output and the actual label. Training continues until a test is passed. For example, if the mean absolute error of the predicted damage removal rate is less than 5%, and the mean absolute error of the predicted polished wafer thickness is less than 1µm, then the polishing predictor is considered to have completed training and possess reliable predictive capabilities.

[0053] For example, considering the complex physicochemical interactions between polishing parameters, post-grinding wafer state, and polishing results, and the significant nonlinearity of their correlation, and the powerful ability of deep neural network models to learn complex mapping relationships from high-dimensional mixed features, a deep neural network model can be selected as the core architecture for building this polishing predictor.

[0054] Specifically, the polishing predictor employs a multi-input multi-output feedforward neural network structure, primarily consisting of an input layer, a feature abstraction layer, and a parallel output layer. The input layer receives a standardized input feature vector, which is constructed by concatenating the sample polishing parameter set, the sample grinding wafer thickness set, and the sample damage layer parameter set. The feature abstraction layer uses a fully connected neural network structure with three hidden layers. The number of neurons in each hidden layer can be set to 256, 128, and 64, respectively, to abstract high-level features layer by layer. Each neural network layer uses the ReLU activation function to introduce non-linear transformation capability, and a Dropout layer is embedded after each hidden layer with a uniform dropout rate of 0.25 to effectively suppress model overfitting and improve its generalization performance. The parallel output layer contains two independent linear output nodes, corresponding to the two target variables to be predicted: damage removal rate and polished wafer thickness.

[0055] During training, the key hyperparameters were set as follows: learning rate of 0.0003, number of training epochs of 250, and batch size of 48. The learning rate was set to ensure smooth convergence during training; the number of training epochs ensured the model had sufficient iteration opportunities to learn the complex mapping relationships in the process data; and the batch size balanced training efficiency with the stability of gradient updates. Specifically, a supervised learning training method was adopted, and the sample polishing parameter set, sample grinding wafer thickness set, sample damage layer parameter set, sample damage removal rate set, and sample polished wafer thickness set were randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0056] Furthermore, the sample polishing parameters, sample grinding wafer thickness, and sample damage layer parameters are combined as the input feature vector, while the corresponding sample damage removal rate and sample polished wafer thickness are used as the supervision signal target vector. Using a backpropagation algorithm paired with the Adam optimizer, all network weight parameters are iteratively optimized. The mean squared error loss function is used to comprehensively measure the overall deviation between the predicted two output variables and the actual target values. The training process is monitored in real time using a validation set. When the model's mean absolute percentage error for predicting the damage removal rate is less than 5% and the mean absolute error for predicting the polished wafer thickness is less than 1µm, the model is considered to have converged, training is terminated, and the polishing predictor is obtained. This polishing predictor can effectively establish the complex nonlinear relationship between polishing process parameters and preceding states to the polishing result, providing an accurate predictive basis for polishing parameter optimization.

[0057] Finally, the first polishing parameters are evaluated, and their corresponding first grinding wafer thickness and optimal damage layer parameters are input together into the polishing predictor, which has been trained and tested successfully. The polishing predictor calculates and outputs the corresponding prediction results, namely the first damage removal rate and the first polished wafer thickness.

[0058] Furthermore, to evaluate the thickness control accuracy after polishing, it is necessary to calculate the similarity between the first polished wafer thickness and the target thickness, and define this similarity value as the first polishing accuracy value. The target thickness specifically refers to the physical thickness that the wafer is expected to reach after the polishing process, pre-set in the wafer thinning process planning. This similarity calculation can employ various quantification methods; preferably, the first polishing accuracy value = 1 − |(first polished wafer thickness − target thickness) / target thickness|. This calculation maps the relative deviation between thicknesses to a value between 0 and 1. When the polished thickness is completely consistent with the target thickness, the similarity reaches its maximum value of 1. As the absolute value of the relative deviation increases, the similarity monotonically decreases from 1, approaching 0. Therefore, the higher the calculated similarity value, the closer the polished thickness is to the target thickness, indicating higher thickness control accuracy.

[0059] Furthermore, based on the pre-set polishing damage weight and polishing accuracy weight, the obtained first damage removal rate and first polishing accuracy value are weighted and calculated. Specifically, the calculation formula is: First Polishing Adaptability Value = Polishing Damage Weight × First Damage Removal Rate + Polishing Accuracy Weight × First Polishing Accuracy Value. Through this calculation, the damage removal effect and thickness control accuracy are combined into the first polishing adaptability value. The level of this first polishing adaptability value comprehensively reflects the overall performance of this set of polishing parameters in achieving the core objectives of the polishing process under the current weight settings. Specifically, the higher the first polishing adaptability value, the better the overall performance of this set of polishing parameters in both damage removal and thickness control objectives, and the closer it is to the ideal process state under the current weight balance; the lower the first polishing adaptability value, the worse the performance of this set of polishing parameters in one or both of the above objectives, and the worse its overall process efficiency.

[0060] Finally, the iterative optimization algorithm is initiated. This algorithm aims to maximize the polishing fitness value, continuously generating new combinations of polishing parameters within the polishing parameter space, and repeating the aforementioned prediction, accuracy calculation, and fitness value evaluation processes. Through multiple iterative comparisons, the algorithm ultimately finds a set of polishing parameters that maximizes the global polishing fitness value; this set of parameters is then determined as the optimal polishing parameters. The optimization process continues until a preset convergence condition is met. This convergence condition is set based on the stability and accuracy requirements of the optimization process. For example, if the improvement in the polishing fitness value is less than 0.05% in 10 consecutive iterations, the optimization process is considered converged, and the iteration is terminated.

[0061] S40: Thin the wafer according to the optimal grinding parameters and optimal polishing parameters.

[0062] Specifically, the optimal combination of grinding parameters, including specific values ​​such as grinding wheel speed, feed rate, and coolant flow rate, is set to the control system of the wafer grinding equipment. The grinding equipment performs the grinding process on the wafer to be processed according to the optimal grinding parameters, so as to achieve rapid and efficient material removal, and strive to control the depth of the subsurface damage layer introduced by grinding to be near the predicted optimal damage layer parameters.

[0063] After the grinding process is completed, the system automatically switches to the polishing process. At this time, the optimal combination of polishing parameters, including the specific values ​​of polishing pad pressure, polishing fluid flow rate, polishing disc rotation speed, and polishing time, is set to the control system of the wafer polishing equipment. The polishing equipment performs the polishing process on the wafer that has been ground according to the optimal polishing parameters. Its goal is to efficiently remove the grinding damage layer and accurately control the final thickness of the wafer to the process target thickness.

[0064] By sequentially performing grinding and polishing processes, the entire thinning process from the original wafer thickness to the target thickness was completed, ultimately achieving the comprehensive goals of improving overall thinning efficiency, controlling damage, and ensuring accuracy.

[0065] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this application firstly achieves adaptive and precise configuration of optimization objectives for both grinding and polishing stages by dynamically setting and adjusting process weights based on wafer features, overcoming the limitation of fixed parameters failing to match varying process requirements. Secondly, by constructing and utilizing high-precision grinding and polishing prediction models, parameter performance is rapidly evaluated in virtual space, reducing actual trial-and-error costs and optimization cycles. Thirdly, by establishing a collaborative mechanism that drives polishing weight adjustment based on grinding optimization distribution information, intelligent connection and global optimization of upstream and downstream process stages are achieved, effectively improving the synergy and consistency of the overall process chain.

[0066] Finally, the obtained optimal grinding and polishing parameters were applied to the production line. While ensuring the final thickness accuracy, the material removal efficiency was comprehensively improved and subsurface damage was reduced, thereby comprehensively improving the quality, efficiency and stability of the wafer thinning process.

[0067] Example 2, as Figure 2 As shown, based on the same inventive concept as the wafer thinning parameter optimization method for semiconductor manufacturing provided in Embodiment 1, this embodiment of the invention also provides a wafer thinning parameter optimization system for semiconductor manufacturing, comprising: The parameter weight configuration module 11 is used to collect the wafer features of the wafer to be thinned, and set the grinding efficiency weight and grinding damage weight according to the wafer features. The grinding parameter optimization module 12 is used to optimize the grinding parameters based on the wafer characteristics, grinding efficiency weight and grinding damage weight, to obtain the optimal grinding parameters, as well as the predicted optimal efficiency parameters and optimal damage layer parameters. The polishing parameter optimization module 13 is used to adjust the grinding damage weight according to the efficiency parameter distribution and damage layer parameter distribution in the grinding parameter optimization process, obtain the polishing damage weight, configure the polishing accuracy weight, optimize the polishing parameters, and obtain the optimal polishing parameters. Thinning process execution module 14 is used to perform thinning process on the wafer according to the optimal grinding parameters and optimal polishing parameters.

[0068] Specifically, the parameter weight configuration module 11 is used for: The wafer features to be thinned are collected, and grinding efficiency weights and grinding damage weights are set based on these wafer features, including: The wafer features to be thinned are collected, including the original thickness and the target thickness. The thinning thickness is calculated based on the original thickness and the target thickness. Based on the thinning thickness, calculate the configuration grinding efficiency weight and grinding damage weight.

[0069] Specifically, based on the thinning thickness, the grinding efficiency weight and grinding damage weight are calculated and configured, including: Obtain the maximum thinning thickness within the wafer thinning process log; The ratio of the thinning thickness to the maximum thinning thickness is calculated and used as a weight for grinding efficiency; The grinding damage weight is calculated based on the grinding efficiency weight.

[0070] The grinding parameter optimization module 12 is specifically used for: Based on the wafer characteristics, grinding efficiency weight, and grinding damage weight, grinding parameters are optimized to obtain the optimal grinding parameters, as well as the predicted optimal efficiency parameters and optimal damage layer parameters, including: Obtain the grinding parameter space; The first grinding parameter is randomly set within the grinding parameter space; Based on the first grinding parameters and wafer characteristics, a first efficiency parameter and a first damage layer parameter are predicted and obtained, wherein the first efficiency parameter includes normalized grinding completion time data and the first damage layer parameter includes normalized damage layer depth data. Based on the grinding efficiency weight and grinding damage weight, combined with the first efficiency parameter and the first damage layer parameter, the first grinding cost of the first grinding parameter is calculated. The grinding parameters are iteratively optimized to obtain the optimal grinding parameters with the minimum grinding cost, and the optimal efficiency parameters and optimal damage layer parameters are obtained from the optimal grinding parameters and wafer feature prediction.

[0071] Specifically, based on the first grinding parameters and wafer characteristics, the first efficiency parameter and the first damage layer parameter are predicted and obtained, including: Based on the wafer grinding process logs over a historical period, a set of sample grinding parameters and a set of sample wafer features were collected. Grinding efficiency parameters, grinding damage layer parameters, and grinding wafer thickness under different sample grinding parameters and sample wafer features were obtained, and the set of sample efficiency parameters, sample damage layer parameters, and sample grinding wafer thickness were labeled to obtain the set of sample grinding parameters. A grinding predictor is built based on machine learning. The grinding predictor is trained under supervision using the sample grinding parameter set, sample wafer feature set, sample efficiency parameter set, sample damage layer parameter set, and sample grinding wafer thickness set. The training is completed after the test converges. The first grinding parameters and wafer features are input into the grinding predictor, and the first efficiency parameter, the first damage layer parameter, and the first grinding wafer thickness are output.

[0072] Specifically, the polishing parameter optimization module 13 is used for: Based on the efficiency parameter distribution and damage layer parameter distribution during the grinding parameter optimization process, the grinding damage weight is adjusted to obtain the polishing damage weight, and a polishing accuracy weight is configured to optimize the polishing parameters and obtain the optimal polishing parameters, including: All efficiency parameters and damage layer parameters in the grinding parameter optimization process are obtained and arranged in ascending order to obtain the distribution of efficiency parameters and damage layer parameters. Based on the efficiency parameter distribution and the damage layer parameter distribution, the grinding damage weight is adjusted to obtain the polishing damage weight; Calculate the polishing accuracy weight based on the polishing damage weight; Based on the polishing damage weight and polishing accuracy weight, the polishing parameters are optimized to obtain the optimal polishing parameters.

[0073] Specifically, based on the efficiency parameter distribution and the damage layer parameter distribution, the grinding damage weight is adjusted to obtain the polishing damage weight, including: Calculate the midpoint value of the efficiency parameter distribution and the damage layer parameter distribution to obtain the midpoint efficiency parameter and the midpoint damage layer parameter; The ratio of the number of damage layer parameters greater than the midpoint damage layer parameter to the number of damage layer parameters less than the midpoint damage layer parameter within the damage layer parameter distribution is used as the first damage weight adjustment coefficient. The ratio of the number of efficiency parameters less than the midpoint efficiency parameter to the number of efficiency parameters greater than the midpoint efficiency parameter within the efficiency parameter distribution is calculated and used as the second damage weight adjustment coefficient. The damage weight adjustment coefficient is calculated based on the first damage weight adjustment coefficient and the second damage weight adjustment coefficient. The grinding damage weight is adjusted and calculated using the damage weight adjustment coefficient to obtain the polishing damage weight.

[0074] Furthermore, the polishing parameters are optimized to obtain the optimal polishing parameters, including: Obtain the polishing parameter space and generate the first polishing parameters; Based on the first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters, polishing prediction is performed to obtain the first damage removal rate and the first polished wafer thickness. Calculate the similarity between the thickness of the first polished wafer and the target thickness, and use it as the first polishing accuracy value; Based on the polishing damage weight and polishing precision weight, the first damage removal rate and the first polishing precision value are weighted and calculated to obtain the first polishing adaptability value; Continue iteratively optimizing the polishing parameters until convergence, obtaining the optimal polishing parameters with the largest polishing fitness value.

[0075] Specifically, based on the first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters, polishing prediction is performed to obtain the first damage removal rate and the first polished wafer thickness, including: Based on the polishing process logs over a historical period, a set of sample polishing parameters, a set of sample grinding wafer thicknesses, and a set of sample damage layer parameters were collected. Additionally, a set of sample damage removal rates and a set of sample polished wafer thicknesses were collected after polishing under different sample polishing parameters, sample grinding wafer thicknesses, and sample damage layer parameters. Each sample damage removal rate includes the percentage of damage layer removed by polishing. A polishing predictor is built based on machine learning. The polishing predictor is supervised and tested using the set of sample polishing parameters, the set of sample grinding wafer thickness, the set of sample damage layer parameters, the set of sample damage removal rate, and the set of sample polishing wafer thickness until the test is qualified. The first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters are input into the polishing predictor, and the first damage removal rate and the first polished wafer thickness are output.

[0076] The thinning process execution module 14 is specifically used for: The wafer is thinned according to the optimal grinding and polishing parameters.

[0077] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0078] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0079] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for optimizing wafer thinning parameters in semiconductor manufacturing, characterized in that, The method includes: Collect the wafer features of the wafer to be thinned, and set grinding efficiency weight and grinding damage weight according to the wafer features; Based on the wafer characteristics, grinding efficiency weight, and grinding damage weight, grinding parameters are optimized to obtain the optimal grinding parameters, as well as the predicted optimal efficiency parameters and optimal damage layer parameters. Based on the distribution of efficiency parameters and damage layer parameters during the grinding parameter optimization process, the grinding damage weight is adjusted to obtain the polishing damage weight, and the polishing accuracy weight is configured to optimize the polishing parameters and obtain the optimal polishing parameters. The wafer is thinned according to the optimal grinding and polishing parameters.

2. The method for optimizing wafer thinning parameters for semiconductor manufacturing according to claim 1, characterized in that, The wafer features to be thinned are collected, and grinding efficiency weights and grinding damage weights are set based on these wafer features, including: The wafer features to be thinned are collected, including the original thickness and the target thickness. The thinning thickness is calculated based on the original thickness and the target thickness. Based on the thinning thickness, calculate the configuration grinding efficiency weight and grinding damage weight.

3. The method for optimizing wafer thinning parameters for semiconductor manufacturing according to claim 2, characterized in that, Based on the thinning thickness, calculate the grinding efficiency weight and grinding damage weight, including: Obtain the maximum thinning thickness within the wafer thinning process log; The ratio of the reduced thickness to the maximum reduced thickness is calculated and used as a weight for grinding efficiency; The grinding damage weight is calculated based on the grinding efficiency weight.

4. The method for optimizing wafer thinning parameters for semiconductor manufacturing according to claim 1, characterized in that, Based on the wafer characteristics, grinding efficiency weight, and grinding damage weight, grinding parameters are optimized to obtain the optimal grinding parameters, as well as the predicted optimal efficiency parameters and optimal damage layer parameters, including: Obtain the grinding parameter space; The first grinding parameter is randomly set within the grinding parameter space; Based on the first grinding parameters and wafer characteristics, a first efficiency parameter and a first damage layer parameter are predicted and obtained, wherein the first efficiency parameter includes normalized grinding completion time data and the first damage layer parameter includes normalized damage layer depth data. Based on the grinding efficiency weight and grinding damage weight, combined with the first efficiency parameter and the first damage layer parameter, the first grinding cost of the first grinding parameter is calculated. The grinding parameters are iteratively optimized to obtain the optimal grinding parameters with the minimum grinding cost, and the optimal efficiency parameters and optimal damage layer parameters are obtained from the optimal grinding parameters and wafer feature prediction.

5. The wafer thinning parameter optimization method for semiconductor manufacturing according to claim 4, characterized in that, Based on the first grinding parameters and wafer characteristics, the first efficiency parameter and the first damage layer parameter are predicted and obtained, including: Based on the wafer grinding process logs over a historical period, a set of sample grinding parameters and a set of sample wafer features were collected. Grinding efficiency parameters, grinding damage layer parameters, and grinding wafer thickness under different sample grinding parameters and sample wafer features were obtained, and the set of sample efficiency parameters, sample damage layer parameters, and sample grinding wafer thickness were labeled to obtain the set of sample grinding parameters. A grinding predictor is built based on machine learning. The grinding predictor is trained under supervision using the sample grinding parameter set, sample wafer feature set, sample efficiency parameter set, sample damage layer parameter set, and sample grinding wafer thickness set. The training is completed after the test converges. The first grinding parameters and wafer features are input into the grinding predictor, and the first efficiency parameter, the first damage layer parameter, and the first grinding wafer thickness are output.

6. The method for optimizing wafer thinning parameters for semiconductor manufacturing according to claim 1, characterized in that, Based on the efficiency parameter distribution and damage layer parameter distribution during the grinding parameter optimization process, the grinding damage weight is adjusted to obtain the polishing damage weight, and a polishing accuracy weight is configured to optimize the polishing parameters and obtain the optimal polishing parameters, including: All efficiency parameters and damage layer parameters in the grinding parameter optimization process are obtained and arranged in ascending order to obtain the distribution of efficiency parameters and damage layer parameters. Based on the efficiency parameter distribution and the damage layer parameter distribution, the grinding damage weight is adjusted to obtain the polishing damage weight; Calculate the polishing accuracy weight based on the polishing damage weight; Based on the polishing damage weight and polishing accuracy weight, the polishing parameters are optimized to obtain the optimal polishing parameters.

7. The method for optimizing wafer thinning parameters for semiconductor manufacturing according to claim 6, characterized in that, Based on the efficiency parameter distribution and the damage layer parameter distribution, the grinding damage weight is adjusted to obtain the polishing damage weight, including: Calculate the midpoint value of the efficiency parameter distribution and the damage layer parameter distribution to obtain the midpoint efficiency parameter and the midpoint damage layer parameter; The ratio of the number of damage layer parameters greater than the midpoint damage layer parameter to the number of damage layer parameters less than the midpoint damage layer parameter within the damage layer parameter distribution is used as the first damage weight adjustment coefficient. The ratio of the number of efficiency parameters less than the midpoint efficiency parameter to the number of efficiency parameters greater than the midpoint efficiency parameter within the efficiency parameter distribution is calculated and used as the second damage weight adjustment coefficient. The damage weight adjustment coefficient is calculated based on the first damage weight adjustment coefficient and the second damage weight adjustment coefficient. The grinding damage weight is adjusted and calculated using the damage weight adjustment coefficient to obtain the polishing damage weight.

8. The method for optimizing wafer thinning parameters for semiconductor manufacturing according to claim 1, characterized in that, Optimize polishing parameters to obtain the optimal polishing parameters, including: Obtain the polishing parameter space and generate the first polishing parameters; Based on the first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters, polishing prediction is performed to obtain the first damage removal rate and the first polished wafer thickness. Calculate the similarity between the thickness of the first polished wafer and the target thickness, and use it as the first polishing accuracy value; Based on the polishing damage weight and polishing precision weight, the first damage removal rate and the first polishing precision value are weighted and calculated to obtain the first polishing adaptability value; Continue iteratively optimizing the polishing parameters until convergence, obtaining the optimal polishing parameters with the largest polishing fitness value.

9. The method for optimizing wafer thinning parameters for semiconductor manufacturing according to claim 8, characterized in that, Based on the first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters, polishing prediction is performed to obtain the first damage removal rate and the first polished wafer thickness, including: Based on the polishing process logs over a historical period, a set of sample polishing parameters, a set of sample grinding wafer thicknesses, and a set of sample damage layer parameters were collected. Additionally, a set of sample damage removal rates and a set of sample polished wafer thicknesses were collected after polishing under different sample polishing parameters, sample grinding wafer thicknesses, and sample damage layer parameters. Each sample damage removal rate includes the percentage of damage layer removed by polishing. A polishing predictor is built based on machine learning. The polishing predictor is supervised and tested using the set of sample polishing parameters, the set of sample grinding wafer thickness, the set of sample damage layer parameters, the set of sample damage removal rate, and the set of sample polishing wafer thickness until the test is qualified. The first polishing parameters, the first grinding wafer thickness, and the optimal damage layer parameters are input into the polishing predictor, and the first damage removal rate and the first polished wafer thickness are output.

10. A wafer thinning parameter optimization system for semiconductor manufacturing, characterized in that, A method for optimizing wafer thinning parameters for semiconductor manufacturing according to any one of claims 1-9 includes: The parameter weight configuration module is used to collect the wafer features of the wafer to be thinned, and set the grinding efficiency weight and grinding damage weight according to the wafer features. The grinding parameter optimization module is used to optimize grinding parameters based on the wafer characteristics, grinding efficiency weight, and grinding damage weight to obtain the optimal grinding parameters, as well as the predicted optimal efficiency parameters and optimal damage layer parameters. The polishing parameter optimization module is used to adjust the grinding damage weight based on the efficiency parameter distribution and damage layer parameter distribution during the grinding parameter optimization process, obtain the polishing damage weight, configure the polishing accuracy weight, optimize the polishing parameters, and obtain the optimal polishing parameters. The thinning process execution module is used to perform thinning processes on the wafer according to the optimal grinding parameters and optimal polishing parameters.

Images