A control method for hot pressing of wafer adhesive coating based on RSM and neural network

Abstract

The present invention provides a control method for hot pressing of wafer coating glue based on RSM and neural network, belonging to the field of intelligent control technology for semiconductor manufacturing equipment. The method comprises: constructing a TTV mathematical model based on the execution parameters of a double-layer vacuum hot pressing molding device through response surface methodology; outputting preliminary execution parameters according to the target TTV using a genetic algorithm; obtaining a predicted TTV using the preliminary execution parameters based on the mathematical model in combination with a neural network; iterating the genetic algorithm based on the predicted TTV to output optimal execution parameters; and the double-layer vacuum hot pressing molding device executing the optimal execution parameters to complete hot pressing of the wafer coating glue. The method integrates response surface methodology, genetic algorithm, and neural network to achieve efficient optimization of hot pressing molding parameters under physical constraints, solving the technical problem of parameter optimization difficulties in traditional hot pressing molding processes.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of semiconductor manufacturing equipment, and in particular to a control method for hot pressing of wafer coating glue based on RSM and neural network. Background Art

[0002] In the semiconductor manufacturing process, wafer protection is a crucial step in ensuring device performance and reliability. Effective protection measures are required for both front-end wafer processes (such as coating, etching, and thin film deposition) and back-end processes (such as thinning, dicing, and packaging) to prevent wafer damage caused by mechanical stress, chemical contamination, or thermal effects. The wafer hot press coating process, a key manufacturing process, has core indicators including the uniformity of the formed film, adhesion strength, and total thickness variation (TTV). TTV directly impacts the accuracy of subsequent processes and device yield, so maintaining TTV within industry standards is key to improving semiconductor product reliability.

[0003] However, in the existing technology, the control of the execution parameters of the double-layer vacuum hot pressing molding equipment in the wafer hot pressing coating process relies on manual experience, which has problems such as being unable to monitor small changes in real time, easily missing the best adjustment time, extending the production cycle, and increasing costs.

[0004] Therefore, a control method for wafer adhesive hot pressing molding based on RSM and neural network is needed to improve the target TTV accuracy. Summary of the Invention

[0005] In view of this, the present invention provides a control method for hot pressing molding of wafer coating glue based on RSM and neural network, which integrates response surface methodology, genetic algorithm (GA) and neural network (NN) to achieve efficient optimization of hot pressing molding parameters under physical constraints, and solve the technical problem of parameter optimization difficulty in traditional hot pressing molding process.

[0006] To this end, the present invention provides the following technical solutions:

[0007] A control method for hot pressing of wafer adhesive coating based on RSM and neural network, characterized by comprising:

[0008] A TTV mathematical model based on the execution parameters of the double-layer vacuum hot pressing equipment was constructed using the response surface methodology.

[0009] Genetic algorithm is used to output preliminary execution parameters according to the target TTV;

[0010] Obtaining a predicted TTV based on the mathematical model in combination with a neural network using the preliminary execution parameters;

[0011] Iterating the genetic algorithm based on the predicted TTV and outputting optimal execution parameters;

[0012] The double-layer vacuum hot pressing molding equipment executes the optimal execution parameters to complete the hot pressing molding of the wafer coating glue.

[0013] Furthermore, the execution parameters of the double-layer vacuum hot pressing molding equipment include:

[0014] Temperature, pressure, hot pressing time and process type.

[0015] Furthermore, the mathematical model includes:

[0016]

[0017] in, represents the intercept term, represents the target TTV, represents random error;

[0018] represents the linear main effect of a single independent variable on the target TTV, is the corresponding independent variable The coefficient of

[0019] Indicates the nonlinear effect of a single independent variable on the target TTV yes Term coefficient;

[0020] represents the interaction term, reflecting the synergistic or antagonistic effect between the two independent variables on the target TTV. and is the mutual coefficient.

[0021] Furthermore, the fitness function of the genetic algorithm includes:

[0022]

[0023] in, To predict TTV, Calculated based on the degree of exceeding equipment limits or not meeting production specifications, is the preset penalty coefficient, represents the target TTV, represents the accuracy bonus item, represents the constraint penalty term.

[0024] Furthermore, the neural network model includes:

[0025] The number of nodes in the input layer is 9, including: 7 execution parameters of the double-layer vacuum hot pressing molding equipment as well as quadratic terms and interaction terms;

[0026] The number of nodes in the first hidden layer is 12, and the activation function is LeakyReLU, which is used to perform preliminary nonlinear mapping and feature extraction on the input features;

[0027] The second hidden layer has 7 nodes and the activation function is LeakyReLU, which is used to further compress and abstract the high-dimensional features output by the first hidden layer.

[0028] The number of nodes in the output layer is 1, and the activation function is Sigmoid, which is used to output the normalized predicted TTV.

[0029] Furthermore, the physical constraints of the mathematical model include:

[0030] The pressure range is set to 20-50T and the temperature range is 120-160°C.

[0031] Advantages and positive effects of the present invention:

[0032] This method uses the response surface methodology to conduct experimental design and data analysis, construct a TTV mathematical model, and reveal the relationship between process parameters and response variables.

[0033] The neural network model is combined with the TTV mathematical model to output the predicted TTV. Based on the predicted TTV, the fitness function of the genetic algorithm is constructed for iteration, and the optimal execution parameters of the equipment are output to achieve efficient parameter optimization.

[0034] By incorporating physical constraints into the response surface methodology and neural network model, the optimal combination of hot pressing molding parameters can be found quickly and accurately while ensuring process feasibility, thereby achieving protective film molding. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0036] Figure 1 Flowchart of a control method for hot-pressing wafer adhesive coating based on RSM and neural network in an embodiment of the present invention;

[0037] Figure 2This is a framework diagram of a control method for wafer adhesive coating and hot pressing based on RSM and neural network in an embodiment of the present invention;

[0038] Figure 3 A structural diagram of a neural network model in an embodiment of the present invention;

[0039] Figure 4 This is a visualization page diagram of the control method for wafer adhesive coating hot pressing molding based on RSM and neural network in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0041] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0042] This invention provides a control method for hot-pressing wafer adhesive coating based on RSM and neural networks. Using response surface methodology (RSM) and neural networks (NN), the TTV is predicted based on execution parameters. The predicted TTV is then iteratively optimized using a genetic algorithm (GA) to output the optimal execution parameters, guiding the equipment to complete hot-pressing. Physical constraints are introduced into the RSM to ensure that parameters are optimized within a practical range. This method significantly improves the uniformity of the target TTV and process stability.

[0043] Combine Figure 1 , the method steps include:

[0044] S1. Build a TTV mathematical model based on the execution parameters of the double-layer vacuum hot pressing molding equipment through response surface methodology;

[0045] S2, using genetic algorithm to output preliminary execution parameters according to target TTV;

[0046] S3, using the preliminary execution parameters to obtain the predicted TTV based on the TTV mathematical model combined with the neural network;

[0047] S4, iterate the genetic algorithm based on the predicted TTV and output the optimal execution parameters;

[0048] S5. The double-layer vacuum hot pressing molding equipment executes the optimal execution parameters to complete the hot pressing molding of the wafer coating glue.

[0049] Example 1

[0050] Combine Figure 2 The steps of this method include:

[0051] S1. Build a TTV mathematical model based on the execution parameters of the double-layer vacuum hot pressing molding equipment through response surface methodology;

[0052] The Box-Behnken experimental design method was used to estimate the quadratic effects of parameters with a relatively small number of experiments, and 49 sets of initial experimental data were generated to ensure that the practical feasible range of the process parameters was covered.

[0053] In this embodiment, 49 initial experiments were conducted with the following seven device execution parameters:

[0054] (Temperature 1, range: 130-160°C), (Temperature 2, range: 120-150°C), (Temperature 3, range: 125-135°C), (Temperature 4, range: 135-155°C), (pressure, range: 30-60T), (time, range: 20-40 seconds) and (Process type, represented by an integer value 0, 1, or 2).

[0055] Response surface methodology (RSM) was used to perform statistical analysis and mathematical modeling on the preprocessed experimental data to describe the global relationship between execution parameters and target TTV.

[0056] 1) For seven independent variables According to the Box–Behnken design, a series of experimental points were arranged.

[0057] 2) Based on the TTV obtained from the experiment, a mathematical model was constructed through multivariate quadratic regression:

[0058]

[0059] in, : intercept term, when all independent variables are at their initial values ​​(here we assume 0), the dependent variable (Target TTV) baseline value.

[0060] : A linear term that reflects the effect of a single independent variable (temperature 1-4, pressure value, pressure holding time, process) on The linear main effect of The coefficient of .

[0061] : Quadratic term, reflecting the effect of a single independent variable on Nonlinear effects, such as excessively high or low temperatures, may not linearly affect the TTV of the adhesive target. yes Term coefficient.

[0062] : Interaction term, reflecting the synergistic or antagonistic effect between two independent variables Effects such as temperature and pressure duration may work together to affect the adhesive target TTV. and is the mutual coefficient.

[0063] : Random error, representing the part that cannot be explained by the model, including measurement errors, unconsidered minor factors, etc.

[0064] Through limited experimental data, the general trend and main influencing factors between input variables and output responses are found, providing the model with a preliminary analytical expression of the relationship between execution parameters and TTV based on statistical analysis.

[0065] 3) Through goodness of fit, regression coefficient significance test and residual analysis, an analytical expression describing the TTV trend and dominant factors was obtained, providing preliminary predictions for subsequent neural network calibration.

[0066] S2. Construct a neural network based on the mathematical model of TTV and perform training.

[0067] In order to further capture the high-order nonlinear relationships that are difficult to describe in the RSM layer, this embodiment introduces a multi-layer feedforward neural network to correct and optimize the initial output of the RSM.

[0068] To capture the complex interactions between parameters, the input feature space is expanded from 7 dimensions to 9 dimensions, and two new interaction terms are added: squared The product of x5 and x6 Through preliminary analysis or domain knowledge, we identify new items that have a significant impact on TTV. Experimental data is generated using a double-layer vacuum hot press and the corresponding TTV is recorded.

[0069] 1) Network structure: Network input includes RSM output and the original parameters , and new composite features and , a total of 9 nodes; there are two hidden layers, with 12 and 7 nodes respectively; the output layer has 1 node, corresponding to the predicted TTV.

[0070] 2) Activation function: The hidden layer uses the LeakyReLU activation function ( =0.03 ), the output layer uses the Sigmoid function to map the network output to the 0-1 interval and then denormalize to obtain the final target TTV prediction.

[0071] Combine Figure 3 As shown, the neural network includes:

[0072] 1) ResNeSt101 backbone network:

[0073] Multi-channel residual block: The network consists of 101 residual convolutional layers, and each residual block includes:

[0074] Grouped Conv, in the hot pressing process, divides the input execution parameters into multiple groups, and performs independent convolution operations on each group, thereby capturing the local relationship in different parameter subspaces. In this embodiment, the temperature-related parameters ( ) are divided into a group, pressure related parameters ( ) are divided into another group, and features are extracted separately and then fused.

[0075] The intra-group features are convolved to locally extract the intra-group subspace relationship; in this embodiment, 3×3 convolution or 1×1 convolution is performed.

[0076] 2) Split Attention Mechanism:

[0077] Feature splitting: split the group convolution output into several branches;

[0078] Global Average Pooling: Perform GAP (Global Average Pooling) on ​​each branch to extract global statistical information;

[0079] Attention weight generation: Calculate the importance weight of each branch through two layers of full connection + activation function;

[0080] Reweighted fusion: The features of each branch are weighted and fused to highlight the key interaction features.

[0081] Residual connection: Jump-connect the input to the output to ensure smooth gradient backpropagation and prevent gradient vanishing in deep networks.

[0082] Hierarchical stacking: The above residual blocks are repeatedly stacked 101 times to form a deep multi-channel feature extraction skeleton.

[0083] Through the group attention mechanism, the nonlinear interaction effects of parameters such as temperature and pressure in the glue coating process can be captured more accurately. Combined with the global search capability of the genetic algorithm, an efficient solution for reverse optimization of seven process parameters from the target TTV can be achieved.

[0084] 2) Physical Constraint Embedding

[0085] Insert a constrained layer after each residual block:

[0086] Detect whether the current feature (after mapping back to parameter space) exceeds the physical limit (such as temperature > 160°C);

[0087] If the boundary is crossed, the degree of violation is added to the backpropagation loss through the penalty gradient mechanism, forcing the network to fall back to the legal space.

[0088] 3) Global fusion and prediction output:

[0089] Perform GAP on the output feature map of the last residual block to obtain a one-dimensional vector.

[0090] The one-dimensional vector is input into several fully connected layers and finally mapped to a single output node.

[0091] 4) Loss function and training strategy:

[0092] Compound loss:

[0093]

[0094] in, To predict TTV, Calculated based on the degree of exceeding equipment limits or not meeting production specifications, is the preset penalty coefficient, represents the target TTV, represents the accuracy bonus item, represents the constraint penalty term.

[0095] Optimizer: Adam, initial learning rate 10-3, using cosine annealing decay;

[0096] Training details: batch size 16, maximum 200 epochs; early stopping on the validation set: termination if validation loss does not decrease after 10 epochs.

[0097] Dataset partitioning: Split the training / validation / test sets into an 8:1:1 ratio to ensure model generalization capabilities.

[0098] 5) Reasoning and parameter inversion:

[0099] Forward reasoning: The preliminary parameter combination of the target TTV obtained by genetic algorithm search is used as input, the forward network is executed, and the predicted TTV is output;

[0100] Feedback iteration: Genetic algorithm The prediction error and constraint penalty in are used as fitness, and the parameter population is selected, crossed, and mutated to continuously approach the optimal process parameter combination.

[0101] S3. Use a genetic algorithm to output preliminary execution parameters according to the target TTV. Use the preliminary execution parameters to obtain a predicted TTV based on the TTV mathematical model combined with a neural network. Iterate the genetic algorithm based on the predicted TTV to output the optimal execution parameters.

[0102] The preset target TTV is expressed as , a numerical optimization combining genetic algorithm and reinforcement learning strategy is introduced to obtain the optimal parameter combination that meets the physical constraints.

[0103] 1) Use real number execution parameters Execute parameters with subtype (The values ​​are 0, 1, and 2, representing three process types) The chromosome sequence is composed of real number coding and integer coding respectively.

[0104] 2) The fitness function is defined as:

[0105]

[0106] in, is the predicted TTV output by the neural network, Calculated based on the degree of exceeding equipment limits or not meeting production specifications, The penalty coefficient set by the user controls the penalty intensity of "constraint violation" The bigger it is, the heavier the punishment.

[0107] 3) Genetic manipulation:

[0108] Population size: 120 individuals to ensure diversity of solution space;

[0109] Selection operator: tournament selection (k=5);

[0110] Crossover rate: simulated binary crossover (SBX), crossover rate 0.85;

[0111] Mutation rate: Gaussian mutation, mutation rate 0.12;

[0112] Stop condition: Fitness improvement < 0.1% for 20 consecutive generations.

[0113] During the model training phase, the RSM parsing layer provides initial predictions for the neural network, which then performs nonlinear corrections. During the optimization phase, the genetic algorithm uses the neural network to make predictions and selects, crosses over, and mutates chromosomes based on the fitness function to achieve a global search of the parameter space.

[0114] Once the target TTV is determined, execute the following steps in sequence:

[0115] Genetic algorithm initializes the population;

[0116] Decode each individual and input it into the neural network to predict TTV;

[0117] Evaluate individuals based on a fitness function;

[0118] Perform selection, crossover, and mutation in sequence, and iteratively update the population;

[0119] After reaching the termination condition, the best parameter combination is output , guiding the actual hot pressing molding.

[0120] Example 2

[0121] In this implementation, this method is implemented based on the interactive interface:

[0122] 1) First, click the "Train Model" button on the interface to complete model initialization;

[0123] 2) In the "Target TTV ( )" input box, enter: 7.5 ;

[0124] 3) Click "Calculate Parameters" and the system starts the optimization process. The results are as follows: Figure 4 As shown:

[0125] The neural network model predicts TTV to be 16 , corresponding to the bar chart display of "forecast parameter combination" in the interface.

[0126] After several generations of evolution, GA outputs the following parameters: =155.26℃, =139.32℃, =125.47℃, =152.67℃, =50.62T, = 37.59 seconds, =1. The equipment is operated according to this parameter and the actual TTV is measured after hot pressing;

[0127] 4) Click "History" to review past data and view previous predicted TTV, target TTV, and execution parameters.

[0128] This method achieves precise target TTV control on a double-layer vacuum hot pressing system. Combining the statistical robustness of RSM, the nonlinear modeling capabilities of NN, and the global optimization advantages of GA, this method ensures that parameter optimization balances accuracy and practicality, providing an efficient and reliable solution for semiconductor manufacturing.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for hot pressing of wafer coating adhesive based on RSM and neural network, characterized in that: include: A TTV mathematical model based on the execution parameters of the double-layer vacuum hot pressing molding equipment was constructed using the response surface methodology; the mathematical model: in, represents the intercept term, represents the target TTV, represents random error; represents the linear main effect of a single independent variable on the target TTV, is the corresponding independent variable The coefficient of represents the nonlinear effect of a single independent variable on the target TTV, yes Term coefficient; represents the interaction term, reflecting the synergistic or antagonistic effect between the two independent variables on the target TTV. and is the mutual coefficient; A genetic algorithm is used to output preliminary execution parameters according to the target TTV; the fitness function of the genetic algorithm is: in, To predict TTV, Calculated based on the degree of exceeding equipment limits or not meeting production specifications, is the preset penalty coefficient, represents the target TTV, represents the accuracy bonus item, represents the constraint penalty term; Obtaining a predicted TTV based on the mathematical model in combination with a neural network using the preliminary execution parameters; Iterating the genetic algorithm based on the predicted TTV and outputting optimal execution parameters; The double-layer vacuum hot pressing molding equipment executes the optimal execution parameters to complete the hot pressing molding of the wafer coating glue.

2. The method according to claim 1, characterized in that The execution parameters of the double-layer vacuum hot pressing molding equipment include: Temperature, pressure, hot pressing time and process type.

3. The method according to claim 1, characterized in that The neural network model includes: The number of nodes in the input layer is 9, including: 7 execution parameters of the double-layer vacuum hot pressing molding equipment as well as quadratic terms and interaction terms; The number of nodes in the first hidden layer is 12, and the activation function is LeakyReLU, which is used to perform preliminary nonlinear mapping and feature extraction on the input features; The second hidden layer has 7 nodes and the activation function is LeakyReLU, which is used to further compress and abstract the high-dimensional features output by the first hidden layer. The number of nodes in the output layer is 1, and the activation function is Sigmoid, which is used to output the normalized predicted TTV.

4. The method according to claim 1, wherein The physical constraints of the mathematical model include: The pressure range is set to 20-50T and the temperature range is 120-160°C.

Images