Adaptive Optimization Method of Process Parameters for HDI Printed Circuit Board Manufacturing

Abstract

The present invention discloses a method for adaptively optimizing process parameters for manufacturing HDI printed circuit boards. The method relates to the field of printed circuit board manufacturing. The method comprises: obtaining historical micro-blind hole morphology and hole wall state feature image data and corresponding historical micro-blind hole electroplating parameters and quality data; using a convolutional neural network to extract the micro-blind hole morphology and hole wall state feature vectors; training a regression model based on the feature vector and historical data to output electroplating thickness and hole filling rate; and optimizing process parameters using a genetic algorithm with electroplating quality as the goal. The present invention introduces an attention mechanism and a weighted cross entropy loss function to improve the accuracy and discrimination of feature expression, solve the problem of sample imbalance, and realize multi-objective trade-off optimization through a genetic algorithm, thereby improving the adaptability, robustness and efficiency of process parameter optimization.

Description

Technical Field

[0001] The present application relates to the field of printed circuit board manufacturing, and in particular to a method for adaptively optimizing process parameters for HDI printed circuit board manufacturing. Background Art

[0002] As electronic devices continue to develop towards high density, miniaturization and light weight, HDI (High Density Interconnect) printed circuit boards are increasingly used in the electronics manufacturing industry. HDI technology can not only increase the wiring density of circuit boards, but also significantly improve electrical performance. However, the manufacturing process of HDI printed circuit boards is complex, especially the electroplating process of micro blind holes has a direct and far-reaching impact on the quality of the final product.

[0003] In the manufacturing process of HDI printed circuit boards, the formation and electroplating of micro blind holes is one of the core technologies. The morphology and hole wall state of micro blind holes directly affect the electroplating thickness and hole filling rate, and these parameters determine the conductive properties and structural integrity of the circuit board. However, due to the small size and complex morphology of micro blind holes, traditional process parameter optimization methods are difficult to meet the requirements of high precision and high consistency.

[0004] At present, the process parameter optimization method commonly used in the industry mainly relies on experience and trial and error. This method is not only time-consuming and labor-intensive, but also difficult to ensure the stability and reproducibility of the optimal process parameters. With the rapid development of artificial intelligence and machine learning technologies, it has become a trend to apply these advanced technologies to the optimization of micro blind hole electroplating processes. In particular, the outstanding performance of convolutional neural networks (CNNs) in image recognition and feature extraction makes them an ideal tool for processing micro blind hole morphology and hole wall state feature image data.

[0005] However, when processing micro-blind hole image data, existing convolutional neural network models often ignore the complexity and diversity of micro-blind hole morphology and hole wall state, resulting in insufficient accuracy and reliability of feature extraction. In addition, traditional optimization methods are difficult to simultaneously take into account the optimal goals of electroplating thickness and hole filling rate during the optimization of multi-dimensional process parameters. Therefore, there is an urgent need for an adaptive optimization method that can comprehensively consider micro-blind hole morphology, hole wall state and electroplating process parameters to achieve efficient optimization and precise control of HDI printed circuit board manufacturing processes. To this end, a process parameter adaptive optimization method and system for HDI printed circuit board manufacturing are proposed. Summary of the invention

[0006] The present application aims to solve at least one of the technical problems in the related art to a certain extent. To this end, one purpose of the present application is to propose a method and system for adaptively optimizing process parameters for HDI printed circuit board manufacturing, thereby achieving efficient optimization and precise control of the HDI printed circuit board manufacturing process.

[0007] One aspect of the present application provides a method for adaptively optimizing process parameters for HDI printed circuit board manufacturing, including:

[0008] Obtaining the morphology and hole wall state characteristic image data of the historical micro blind hole and the corresponding historical micro blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate;

[0009] The convolutional neural network is used to analyze the morphology and hole wall state characteristic image data of the historical micro blind holes, and the micro blind hole and hole wall characteristic vectors of the morphology and hole wall state characteristic image data of the micro blind holes are extracted;

[0010] A first regression model is trained based on the micro-blind hole and hole wall feature vectors, historical micro-blind hole electroplating numerical parameters, and corresponding historical electroplating thickness and hole filling rate, wherein the output of the first regression model is electroplating thickness and hole filling rate;

[0011] Taking electroplating thickness and hole filling rate as the target, genetic algorithm is used to optimize the process parameters to obtain the optimal combination of process parameters.

[0012] Preferably, the numerical parameters of the micro blind hole electroplating include degumming time and temperature, micro-etching time and temperature, water washing time, plating solution composition, plating solution temperature and plating solution pH value, current density, plating time and conductivity.

[0013] Preferably, the step of optimizing the process parameters using a genetic algorithm to obtain the optimal process parameter combination with the electroplating thickness and the hole filling rate as the target includes:

[0014] generating a plurality of first micro blind hole electroplating numerical parameter combinations;

[0015] Inputting several groups of first micro blind hole electroplating numerical combination parameters into the first regression model respectively, predicting the corresponding electroplating thickness and hole filling rate, and obtaining the corresponding fitness value according to the corresponding electroplating thickness and hole filling rate;

[0016] A genetic algorithm is used to perform selection, crossover and mutation operations on each group of first micro blind hole electroplating numerical parameters to generate optimal micro blind hole electroplating numerical parameters.

[0017] Preferably, the morphology and wall state characteristic image data of the historical micro blind holes are comprehensively screened based on the indexes of the pixel micro blind hole aperture size, the pixel micro blind hole wall roughness, the pixel micro blind hole bottom flatness, and the pixel micro blind hole electroplating uniformity.

[0018] Preferably, the convolutional neural network is an attention-enhanced convolutional neural network model.

[0019] Preferably, the attention-enhanced convolutional neural network model training method is:

[0020] Step S201: using the filtered historical micro blind hole morphology and hole wall state feature image data as training data and representing them;

[0021] Step S202: constructing a spatial attention enhanced convolutional neural network SAE-CNN model, inputting the training data of step S1 into the SAE-CNN model, and extracting high-level features of the micro blind hole morphology and hole wall state through convolution layers, pooling layers and spatial attention mechanisms;

[0022] Step S203: designing a weighted cross entropy loss function to assign different weights to samples of different quality levels;

[0023] Step S204: Use the Adam optimization algorithm to train the SAE-CNN model of step S202, minimize the loss function of step S203, and iteratively update the model parameters;

[0024] Step S205: A K-fold cross-validation strategy is used to evaluate and select the SAE-CNN model trained in step S204 to obtain a final feature extraction model for the micro blind hole morphology and hole wall state.

[0025] Preferably, the weighted cross entropy loss function is:

[0026] ;

[0027] in: Indicates the quality grade of micro blind holes. The weight of Micro blind hole quality grade The number of samples, Indicates a micro blind hole sample Belong to the quality level The true label (0 or 1). Represents the SAE-CNN model predicting the micro blind hole sample Belong to the quality level probability.

[0028] One aspect of the present application provides a process parameter adaptive optimization system for HDI printed circuit board manufacturing, including: a data acquisition module, which acquires historical micro blind hole morphology and hole wall state characteristic image data and corresponding historical micro blind hole electroplating numerical parameters and corresponding historical electroplating thickness and hole filling rate;

[0029] A convolutional neural network model processing module uses a convolutional neural network to analyze the morphology and hole wall state feature image data of historical micro blind holes, and extracts the micro blind hole and hole wall feature vectors of the morphology and hole wall state feature image data of the micro blind holes;

[0030] A first regression model training module is used to train a first regression model based on the micro-blind hole and hole wall feature vectors, historical micro-blind hole electroplating numerical parameters, and corresponding historical electroplating thickness and hole filling rate, wherein the output of the first regression model is electroplating thickness and hole filling rate;

[0031] The optimization module uses genetic algorithm to optimize the process parameters with electroplating thickness and hole filling rate as the target to obtain the optimal process parameter combination.

[0032] One aspect of the present application provides an electronic device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the process parameter adaptive optimization method are implemented.

[0033] One aspect of the present application provides a readable storage medium, characterized in that the readable storage medium stores a computer program, and the computer program is suitable for loading by a processor to execute the steps in the process parameter adaptive optimization method.

[0034] Beneficial Effects

[0035] The process parameter adaptive optimization method for HDI printed circuit board manufacturing proposed in this application obtains the morphological and wall state feature image data of historical micro blind holes by screening, and uses the attention-enhanced convolutional neural network model for analysis and feature extraction. The spatial attention mechanism is introduced to adaptively focus on the key areas in the micro blind hole image, which can accurately characterize the geometric features and surface state of the micro blind holes, improve the accuracy and discriminability of feature expression, and provide a reliable data basis for subsequent process parameter optimization. At the same time, the feature screening is carried out by comprehensively considering multiple indicators such as the aperture size of the micro blind hole, the roughness of the hole wall, the flatness of the hole bottom and the uniformity of electroplating, which ensures the accuracy of the input data and improves the adaptability and robustness of the optimization method. In terms of modeling, a regression model is used to establish the quantitative relationship between the micro blind hole characteristics, electroplating parameters and electroplating quality, and the accurate mapping and prediction of process parameters to electroplating quality are realized, providing a theoretical basis and decision support for process optimization. In the process of model training, a weighted cross entropy loss function is designed to assign different weights to samples of different quality levels, which solves the problem of sample imbalance and improves the effect and generalization ability of model training. At the same time, the Adam optimization algorithm and K-fold cross-validation strategy are used to train and select the model, ensuring the efficiency of model learning and the reliability of performance, reducing the risk of overfitting. Finally, the genetic algorithm is introduced to perform global optimization of process parameters, which can balance multiple objectives and quickly search for the optimal parameter combination, improving optimization efficiency and result quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1A schematic diagram of a method for adaptively optimizing process parameters for manufacturing HDI printed circuit boards provided in Example 1 of the present application;

[0037] Figure 2 A method for training a convolutional neural network model with enhanced attention in this application;

[0038] Figure 3 Schematic diagram of a process parameter adaptive optimization system for HDI printed circuit board manufacturing provided in Example 2 of the present application;

[0039] Figure 4 A schematic diagram of the structure of an electronic device provided in Example 3 of the present application;

[0040] Figure 5 It is a schematic diagram of the computer-readable storage medium structure provided in Example 4 of the present application. DETAILED DESCRIPTION

[0041] In order to better understand the present application, a more detailed description will be made of various aspects of the present application with reference to the accompanying drawings. It should be understood that these detailed descriptions are only descriptions of exemplary embodiments of the present application, and are not intended to limit the scope of the present application in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0042] In the accompanying drawings, the size, dimensions and shapes of the elements have been slightly adjusted for ease of illustration. The drawings are for illustration only and are not strictly drawn to scale. As used herein, the terms "substantially", "approximately" and similar terms are used as terms of approximation, not as terms of degree, and are intended to illustrate the inherent deviations in measurements or calculations that will be recognized by those of ordinary skill in the art. In addition, in this application, the order in which the steps are described does not necessarily represent the order in which these processes occur in actual operation, unless otherwise specified or can be derived from the context.

[0043] It should also be understood that expressions such as "include", "including", "have", "contain" and / or "comprising" are open rather than closed expressions in this specification, which indicate the presence of the stated features, elements and / or components, but do not exclude the presence of one or more other features, elements, components and / or combinations thereof. In addition, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire list of features rather than just the individual elements in the list. In addition, when describing embodiments of the present application, "may" is used to mean "one or more embodiments of the present application". And, the term "exemplary" is intended to refer to an example or illustration.

[0044] Unless otherwise defined, all words (including engineering terms and scientific and technological terms) used in this article have the same meaning as those commonly understood by ordinary technicians in the field to which this application belongs. It should also be understood that unless there is a clear explanation in this application, words defined in commonly used dictionaries should be interpreted as having the same meaning as their meaning in the context of the relevant technology, and should not be interpreted in an idealized or overly formal sense.

[0045] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0046] Example 1

[0047] like Figure 1 As shown, the process parameter adaptive optimization method for HDI printed circuit board manufacturing includes: step S100: obtaining the morphology and hole wall state characteristic image data of historical micro blind holes and the corresponding historical micro blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate;

[0048] Step S200: using a convolutional neural network to analyze the historical micro blind hole morphology and hole wall state characteristic image data, and extracting micro blind hole and hole wall characteristic vectors of the micro blind hole morphology and hole wall state characteristic image data;

[0049] Step S300: training a first regression model based on the micro-blind hole and hole wall feature vectors, historical micro-blind hole electroplating numerical parameters, and corresponding historical electroplating thickness and hole filling rate, wherein the output of the first regression model is electroplating thickness and hole filling rate;

[0050] Step S400: Taking the electroplating thickness and hole filling rate as targets, a genetic algorithm is applied to optimize the process parameters to obtain an optimal process parameter combination.

[0051] The following will be combined Figure 1 to Figure 2 The specific steps of each step of the above optimization method are described in detail.

[0052] S100 acquires the historical micro blind hole morphology and hole wall state characteristic image data and the corresponding historical micro blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate;

[0053] Use high-precision microscopes (such as scanning electron microscopes SEM) and optical scanning equipment (such as confocal microscopes) to obtain images of the morphology and wall state of micro blind holes. For example, use SEM at different magnifications (such as 500 times, 1000 times, 5000 times) to take top views, side views, and cross-sectional views of micro blind holes to obtain morphological features such as diameter, depth, and wall roughness of micro blind holes; use confocal microscopes to obtain three-dimensional morphology and wall state images of micro blind holes.

[0054] The quality of the historical micro-blind hole morphology and hole wall state characteristic image data directly affects the model quality. In a more preferred embodiment, the screening of the historical micro-blind hole morphology and hole wall state characteristic image data adopts indicators based on pixel micro-blind hole aperture size, pixel micro-blind hole wall roughness, pixel micro-blind hole bottom flatness, and pixel micro-blind hole electroplating uniformity to comprehensively screen and obtain the historical micro-blind hole morphology and hole wall state characteristic image data: specifically as follows:

[0055] Pixel-based blind hole aperture size estimation function :

[0056] ;

[0057] Where H is the number of pixels in the hole depth direction, is the hole radius at the ith pixel (in pixels). The smaller it is, the closer the aperture size is to the target range. and Respectively represent the minimum and maximum values ​​set by professional technicians. For example, if the target aperture should be between 50 μm and 70 μm, then That is 50 microns. That is 70 microns. The formula evaluates the accuracy of the aperture size by calculating the relative deviation between the aperture at each depth position and the midpoint of the target interval. The formula calculates the relative deviation of the aperture at each depth position and takes the average value as the evaluation value of the overall aperture size. The smaller the deviation, the closer the aperture is to the target interval.

[0058] When screening images, set the aperture size deviation threshold ,when When the aperture size of the micro blind hole sample exceeds the target range, it is considered that the aperture size of the micro blind hole sample exceeds the target range and is removed. The characteristic value distribution of aperture size deviation is obtained by analyzing the experimental data, and the aperture size deviation of 95% of the samples is within 0.05, then the threshold can be set. .

[0059] Pixel-based evaluation function for the roughness of micro blind hole wall :

[0060]

[0061] in, is the local average of the hole radius at the pixel at the i-th center point, which can be obtained by averaging the hole radius of nearby pixels. The smaller the value, the smoother the hole wall. This formula evaluates the hole wall roughness by calculating the relative deviation between the hole diameter and the local average hole diameter. The formula calculates the relative deviation of the hole diameter at each depth position and takes the average value as the evaluation value of the overall hole wall roughness. The smaller the deviation, the smoother the hole wall.

[0062] When filtering images, set the hole wall roughness threshold ,when When the roughness of the micro blind hole sample wall exceeds the standard, it is considered that the hole wall roughness of the micro blind hole sample exceeds the standard and is removed. Hole wall roughness threshold Assuming that the process requires that the hole wall roughness should be less than 0.01, and through statistical analysis it is found that the roughness of most samples is within 0.008, the threshold can be set ;

[0063] Pixel-based evaluation function for the flatness of the bottom of micro blind vias :

[0064]

[0065] in, is the number of pixels in the hole bottom area B, is the pixel point on the bottom surface of the hole The gray value of is the average gray value of pixels in the bottom area of ​​the well. The smaller the value, the flatter the bottom surface of the hole. This formula evaluates the flatness by calculating the deviation between the gray value of each pixel in the bottom area of ​​the hole and the average gray value.

[0066] When screening images, set the threshold for hole bottom flatness ,when When the bottom of the micro blind hole sample is not flat, it is considered that it is not flat and is removed. The formula calculates the absolute deviation between the gray value of each pixel in the bottom area and the average value, and takes the average value as the evaluation value of the bottom flatness. The smaller the deviation, the flatter the bottom surface of the hole. Bottom flatness threshold Assuming that the experience determines that the bottom flatness of the hole should be less than 0.02, and through optimization and verification tests, the optimal threshold is determined to be around 0.015, then the threshold can be set .

[0067] Pixel-based evaluation function for micro blind hole electroplating uniformity :

[0068]

[0069] in, is the number of pixels in the hole wall area W, is the pixel point on the hole wall Gray value is the average grayscale value of pixels in the hole wall area. Assuming that the grayscale value of the electroplating material is positively correlated with its thickness, then The smaller it is, the more uniform the plating thickness distribution is.

[0070] When screening images, set the plating uniformity threshold ,when When the micro blind hole sample is considered to have uneven electroplating thickness distribution, it is removed. This formula evaluates electroplating uniformity by calculating the relative deviation between the grayscale value of each pixel in the hole wall area and the average grayscale value. The formula calculates the relative deviation between the grayscale value of each pixel in the hole wall area and the average value, and takes the average value as the evaluation value of electroplating uniformity. The smaller the deviation, the more uniform the electroplating thickness distribution. Electroplating uniformity threshold Assuming that through process requirements and experimental data analysis, the characteristic value of electroplating uniformity should be less than 0.03, and through verification testing, it is found that this threshold screening effect is the best, then the threshold can be set .

[0071] The above pixel-based evaluation methods make full use of image processing and analysis technology to achieve efficient and accurate image screening by extracting key features of micro blind holes (such as hole size, hole wall roughness, hole bottom flatness, electroplating uniformity, etc.) and comparing them with corresponding thresholds. These formulas avoid the difficulty of directly measuring physical parameters, but use the pixel information of the image to improve the quality of the data.

[0072] When screening, the images can be screened in the order of aperture size, hole wall roughness, hole bottom flatness, and electroplating uniformity. Only images that pass the previous index screening can enter the evaluation of the next index. You can also perform a weighted average of the evaluation values ​​of the four indicators of each sample to obtain a comprehensive quality score for selection to obtain the morphology and hole wall state characteristic image data of the historical micro blind hole.

[0073] Record the micro-blind hole electroplating numerical parameters during the micro-blind hole electroplating process, including:

[0074] Pretreatment conditions: degumming time and temperature, micro-etching time and temperature, water washing time, etc. For example, degumming time is 10 minutes, temperature is 30°C; micro-etching time is 5 minutes, temperature is 25°C; water washing time is 2 minutes.

[0075] Plating solution composition and temperature: such as CuSO4 concentration, H2SO4 concentration, additive type and concentration, plating solution temperature, etc. For example, CuSO4 concentration is 250g / L, H2SO4 concentration is 60g / L, and temperature is 45°C.

[0076] Electroplating parameters: current density, electroplating time, etc. For example, current density is 2A / dm² and electroplating time is 15 minutes.

[0077] Record the pH value, conductivity and other physical and chemical parameters of the plating solution during the electroplating process, and monitor the stability of the plating solution components;

[0078] Record the voltage change curve during the electroplating process and analyze the stability and controllability of the electroplating process. During the electroplating process, the voltage applied between the cathode (workpiece) and the anode will change over time. The voltage change curve reflects the stability and controllability of the electroplating process. Ideally, the voltage change curve should be smooth without large fluctuations. If the voltage change curve shows abnormal fluctuations, such as sudden increases or decreases, it may mean that there is a problem in the electroplating process, such as cathode surface passivation, anode passivation, changes in the composition of the plating solution, etc. By recording and analyzing the voltage change curve, abnormal conditions in the electroplating process can be discovered in time, and corresponding measures can be taken, such as adjusting the current density, replacing the anode, replenishing the composition of the plating solution, etc., to ensure the stability and controllability of the electroplating process.

[0079] Conductivity: The conductivity of the plating solution reflects the concentration of conductive ions in the plating solution. If the conductivity is too low, the current efficiency will be reduced and the plating speed will be slow; if the conductivity is too high, the crystal grain size of the coating will become coarser and the quality of the coating will be reduced. Therefore, it is necessary to use a conductivity meter to monitor the conductivity of the plating solution in real time to ensure that it is always within the optimal range. If the conductivity deviates from the optimal range, it is necessary to adjust the composition of the plating solution in time, such as adding conductive salt. pH value: The pH value of the plating solution is an important factor affecting the quality of electroplating. Different plating solutions have different optimal pH ranges. For example, the optimal pH value of acid copper plating solution is 3.0-4.0. During the electroplating process, a pH meter is needed to monitor the pH value of the plating solution in real time to ensure that it is always within the optimal range. If the pH value deviates from the optimal range, it needs to be adjusted in time to avoid affecting the quality of electroplating. Conductivity: The conductivity of the plating solution reflects the concentration of conductive ions in the plating solution. If the conductivity is too low, the current efficiency will be reduced and the electroplating speed will be slow; if the conductivity is too high, the crystal grain size of the coating will become coarser and the quality of the coating will be reduced. Therefore, it is necessary to use a conductivity meter to monitor the conductivity of the plating solution in real time to ensure that it is always within the optimal range. If the conductivity deviates from the optimal range, it is necessary to adjust the composition of the plating solution in time, such as adding conductive salt. Other physical and chemical parameters: Depending on the type of plating solution and the requirements of the electroplating process, other physical and chemical parameters may also need to be monitored, such as the density, viscosity, and metal ion concentration of the plating solution. Changes in these parameters will also affect the quality of electroplating and need to be monitored and controlled in real time.

[0080] Quality inspection data: the corresponding micro blind hole plating thickness and hole filling rate after electroplating. For example, the plating thickness is 20 microns and the hole filling rate is 95%.

[0081] The real-time monitoring data is associated with the electroplating parameters and quality inspection data to establish a data set for the electroplating process, providing support for subsequent modeling and optimization.

[0082] By acquiring a large amount of real-time monitoring data (such as pretreatment conditions, pH value, conductivity, voltage change curve, etc.) and also recording the electroplating parameters (such as current density, electroplating time, etc.), the numerical parameters of micro blind hole electroplating are constructed, and the electroplating thickness and hole filling rate are obtained through quality inspection. Such data are easy to obtain to obtain the data set for training the first regression model.

[0083] S200 uses a convolutional neural network (CNN) to process the morphology and hole wall state characteristic image data of the first historical micro blind hole to obtain the morphology and hole wall state characteristic image data of the micro blind hole;

[0084] The acquired first historical micro-blind hole morphology and hole wall state characteristic image data are preprocessed, such as image size normalization, pixel value normalization, data enhancement (rotation, flipping, noise addition), etc., to improve the training effect of CNN.

[0085] Design the CNN network structure, including:

[0086] Input layer: receiving the pre-processed micro-blind hole morphology and hole wall state feature image data.

[0087] Hidden layer: It consists of multiple convolutional layers and pooling layers, which are used to extract local and global features of the image. For example, 3-5 convolutional layers are used, each layer uses a 3x3 or 5x5 convolution kernel, and the activation function is ReLU; 2-3 maximum pooling layers are used, and the pooling kernel size is 2x2 or 3x3.

[0088] Output layer: Flatten the extracted image features into a one-dimensional feature vector as the input of the subsequent regression model. For example, use a fully connected layer to flatten the feature map and use the Dropout regularization method to prevent overfitting.

[0089] Use the labeled micro-blind hole morphology and hole wall state feature image data to train the CNN and optimize the network parameters. For example, use the Adam optimizer, set the learning rate to 0.001, set the batch size to 32, and train for 100 epochs.

[0090] The trained CNN model is used to extract features from the new micro blind hole morphology and hole wall state feature image data to obtain the micro blind hole and hole wall feature vectors of the micro blind hole morphology and hole wall state feature image data.

[0091] In order to improve the performance and adaptability in the feature extraction of micro blind hole morphology and hole wall state, in a more preferred embodiment, a spatial attention enhanced convolutional neural network (Spatial Attention-Enhanced CNN, SAE-CNN) is constructed, and the basic architecture of the convolutional neural network is designed, including a convolutional layer, a pooling layer, a fully connected layer, etc., which are used to extract the hierarchical features of the micro blind hole morphology and hole wall state. On the basis of the convolutional neural network, a spatial attention mechanism is introduced to adaptively adjust the importance of different areas in the feature map.

[0092] like Figure 2 As shown, the steps for training the AE-CNN model and extracting the features of micro blind hole morphology and hole wall state are as follows:

[0093] S201. The filtered historical micro blind hole morphology and hole wall state feature image data are used as training data, represented, and input into the AE-CNN model.

[0094] The filtered historical micro-blind hole morphology and hole wall state feature image data are expressed as ,in Indicates Input micro blind hole SEM image, size , representing height and width respectively The one-hot encoded label vector represents the quality level of the corresponding micro-blind hole, C is the total number of quality level categories, and N is the total number of training samples. The image is input into the SAE-CNN model, and the high-level features of the micro-blind hole morphology and hole wall state are extracted through the convolution layer, pooling layer and spatial attention mechanism.

[0095] S202: Construct a spatial attention enhanced convolutional neural network model, input the training data of step S201 into the SAE-CNN model, and extract high-level features of micro blind hole morphology and hole wall state through convolution layer, pooling layer and spatial attention mechanism.

[0096] Construct a SAE-CNN model, which includes convolutional layers, pooling layers, and a spatial attention mechanism. Input the micro-blind hole SEM image prepared in step 1 into the SAE-CNN model, extract the local features of the image layer by layer through the convolutional layer and the pooling layer, and adaptively enhance the feature representation of important areas through the spatial attention mechanism, and finally obtain the high-level features of the micro-blind hole morphology and hole wall state. The specific steps are as follows:

[0097] S2021 Introducing channel-weighted spatial attention mechanism

[0098] The channel-weighted spatial attention mechanism is introduced, and the global information description of the convolutional feature map is obtained through global average pooling and global maximum pooling operations. and Then, and After concatenation, input two fully connected layers to learn the channel weights and spatial attention weights Finally, the original convolution feature map F is combined with and Multiply element by element and add to F to get the feature map after spatial attention enhancement The mathematical expression is as follows:

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] ;

[0104] in: Represents a convolutional feature map. , , Represent the height, width and number of channels of the feature map respectively. and Represent the results of global average pooling and global maximum pooling respectively. Represents the channel weight. , represents the weight matrix of the first fully connected layer, Represents the weight matrix of the second fully connected layer. represents the Sigmoid function, Represents element-wise multiplication. Represents the feature map after spatial attention enhancement. Represents the feature map after spatial attention enhancement.

[0105] A SAE-CNN model consisting of convolutional layers, pooling layers, and a spatial attention mechanism was constructed. By extracting local features of the image layer by layer and enhancing the feature representation of important areas, high-level features of the micro-blind hole morphology and hole wall state were obtained. Among them, the channel-weighted spatial attention mechanism (S2021) further improved the effect of feature extraction by adaptively adjusting the weights of different areas in the convolution feature map;

[0106] S203 defines a weighted cross entropy loss function and designs a weighted cross entropy loss function to assign different weights to samples of different quality levels;

[0107] In order to alleviate the problem of uneven distribution of micro-blind hole quality grade samples, a weighted cross entropy loss function is designed. , different weights are assigned to samples of different quality levels. The mathematical expression is as follows:

[0108] ;

[0109] in: Indicates the quality grade of micro blind holes. The weight is set according to the inverse of the number of samples, that is, . Micro blind hole quality grade The number of samples. Indicates a micro blind hole sample Belong to the quality level The true label (0 or 1). Represents the SAE-CNN model predicting the micro blind hole sample Belong to the quality level probability.

[0110] S204. Model training: using the Adam optimization algorithm to train the SAE-CNN model of step S202, minimizing the loss function of step S203, and iteratively updating the model parameters;

[0111] The SAE-CNN model is trained using the Adam optimization algorithm. By minimizing the weighted cross entropy loss and iteratively updating the model parameters, the model can learn the mapping relationship between the micro blind hole morphology and hole wall state characteristics and the quality grade.

[0112] Use the Adam optimization algorithm to minimize the weighted cross entropy loss , iteratively update the parameters of the SAE-CNN model:

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] in: and denote the first-order and second-order moment estimates of the gradient, respectively. and is the decay rate hyperparameter. and is the correction value of the moment estimate. represents the initial learning rate, is the smoothing term. Represents the parameters of the SAE-CNN model, and := represents the assignment operation. Indicates the current iteration step number.

[0119] S205 model evaluation and selection, using the K-fold cross-validation strategy to evaluate and select the SAE-CNN model trained in step S204, to obtain the final micro blind hole morphology and hole wall state feature extraction model.

[0120] The K-fold cross-validation strategy is adopted to divide the micro-blind hole SEM image dataset into K parts (such as K=5), and K-1 parts are selected as training sets each time, and the remaining 1 part is used as validation set. The SAE-CNN model is trained on each fold, and the model performance is evaluated on the corresponding validation set, such as accuracy, precision, recall, F1 score and other indicators. Finally, the model hyperparameter combination with the best average validation performance is selected, and the SAE-CNN model is retrained on the entire dataset to obtain the final micro-blind hole morphology and hole wall state feature extraction model.

[0121] By constructing a spatial attention enhanced convolutional neural network (SAE-CNN) model, the high-level features of the micro blind hole morphology and hole wall state are effectively extracted, laying the foundation for the subsequent electroplating thickness and filling rate prediction.

[0122] S300 training a first regression model based on the micro-blind hole and hole wall feature vectors, historical micro-blind hole electroplating numerical parameters, and corresponding historical electroplating thickness and hole filling rate, wherein the output of the first regression model is electroplating thickness and hole filling rate;

[0123] The micro-blind hole and hole wall feature vectors extracted by CNN are combined with the micro-blind hole electroplating numerical parameters recorded in step S100 (such as pre-treatment conditions, plating solution composition and temperature, current density, electroplating time, etc.) to form the input features of the regression model.

[0124] Select a suitable regression model, such as support vector regression (SVR), random forest regression (Random Forest Regression), multi-layer perceptron (MLP), etc. For example, use the SVR model, select the Gaussian kernel (RBF) as the kernel function, set the regularization parameter C to 1.0, and set the kernel function parameter gamma to "auto".

[0125] The input features and the corresponding plating thickness and hole filling rate labels are divided into training sets and test sets to train and evaluate the regression model. For example, a 5-fold cross-validation method is used to evaluate the mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R²) of the model on the test set.

[0126] Use grid search or random search to optimize the hyperparameters of the regression model, such as the C and gamma parameters of SVR, to improve the predictive performance of the model.

[0127] The micro-blind hole and hole wall feature vectors extracted by CNN are combined with the micro-blind hole electroplating numerical parameters to form the input features of the regression model, and the appropriate regression model is selected for training and evaluation. This step establishes the mapping relationship between the micro-blind hole morphology, hole wall state characteristics, electroplating parameters, electroplating thickness, and hole filling rate, providing a prediction model for process parameter optimization.

[0128] The optimized regression model is used to predict the new input features to obtain the predicted electroplating thickness and filling rate of the micro blind vias.

[0129] S500 takes electroplating thickness and hole filling rate as the target, applies genetic algorithm to optimize process parameters to obtain the optimal process parameter combination, and defines the optimization objective function to maximize the electroplating thickness and hole filling rate;

[0130] Exemplarily, a weighted combination of electroplating thickness and hole filling rate is defined as the optimization objective function, and the weight can be adjusted according to actual needs. For example, objective function = 0.6 × electroplating thickness + 0.4 × hole filling rate, indicating that the weight of electroplating thickness is 0.6 and the weight of hole filling rate is 0.4.

[0131] Set the constraints of the process parameter optimization objective function, such as the value range of plating thickness and hole filling rate, the value range of process parameters, etc. For example, the plating thickness range is 10-30 microns, the hole filling rate range is 90%-100%, the current density range is 1-5A / dm², and the plating time range is 10-30 minutes.

[0132] The genetic algorithm (GA) is used to optimize the process parameters to obtain the optimal process parameter combination, including the following steps:

[0133] S501 Initialize population

[0134] According to the constraint conditions of the process parameter optimization objective function set in step S400, a number of first micro blind hole electroplating numerical parameter combinations are randomly generated as the initial population of the genetic algorithm.

[0135] The population size is set to 50, and each individual contains 9 process parameters: degumming time, degumming temperature, micro-etching time, micro-etching temperature, water washing time, plating solution composition, plating solution temperature, current density and plating time.

[0136] Each process parameter can be encoded using real number encoding or binary encoding. Real number encoding is selected here, and the value range of each parameter corresponds to a real number interval.

[0137] S502 Fitness Evaluation

[0138] Several groups of first micro blind hole electroplating numerical combination parameters are respectively input into the trained first regression model to predict the corresponding electroplating thickness and hole filling rate, and the corresponding fitness value is obtained according to the corresponding electroplating thickness and hole filling rate.

[0139] According to the optimization objective function defined in step S400, the fitness value of each individual is calculated. For example, the fitness value = 0.6 × predicted electroplating thickness + 0.4 × predicted hole filling rate.

[0140] The higher the fitness value, the larger the optimization objective function value of the process parameter combination, that is, the closer the electroplating thickness and hole filling rate are to the target value.

[0141] S503 Select operation

[0142] A genetic algorithm is used to perform selection, crossover and mutation operations on each group of first micro blind hole electroplating numerical parameters to generate optimal micro blind hole electroplating numerical parameters.

[0143] The Roulette Wheel Selection method is used to assign selection probabilities based on the fitness values ​​of individuals. Individuals with higher fitness values ​​have greater probabilities of being selected.

[0144] Generate the same number of random numbers as the population size, and select the corresponding individuals according to the interval within which the random numbers fall in the roulette wheel.

[0145] The selected individuals are used as parents and prepared for crossover operation.

[0146] S504 Cross Operation

[0147] The selected parent individuals are randomly paired in pairs, and each pair of individuals decides whether to perform a crossover operation based on the crossover probability (such as 0.8).

[0148] If crossover is performed, one or more crossover points are randomly selected, the two parent individuals are disconnected at the crossover points, and some of their gene fragments are exchanged to generate two new offspring individuals.

[0149] Using the arithmetic crossover method, each parameter of the two parent individuals is linearly combined to generate new parameter values. For example, the offspring parameter = α × parent 1 parameter + (1-α) × parent 2 parameter, where α is a randomly generated real number between (0, 1).

[0150] S505 Mutation Operation

[0151] The offspring individuals generated by the crossover operation are mutated, and a certain mutation probability (such as 0.1) is used to determine whether each parameter mutates.

[0152] If a parameter mutates, a new value is randomly generated within its value range to replace the original parameter value.

[0153] The mutation operation introduces new parameter values, which helps maintain the diversity of the population and avoids premature convergence to the local optimal solution.

[0154] S506 New population generation

[0155] The new generation of individuals generated through selection, crossover and mutation operations are merged with the parent individuals to form a new population.

[0156] The size of the new population is the same as the initial population, keeping the population size constant.

[0157] S507 Termination condition judgment

[0158] Set the termination conditions of the genetic algorithm, such as the maximum number of iterations (such as 100) or the improvement of the optimal fitness value after several consecutive iterations is less than a certain threshold (such as 0.1%).

[0159] If the current number of iterations reaches the maximum number of iterations, or the optimal fitness value is improved very little after several consecutive iterations, it is considered that the algorithm has converged, and the process jumps to step S508.

[0160] Otherwise, return to step S502 to continue the next round of iterative optimization.

[0161] S508 Output optimization results

[0162] In the last generation of the population, the individual with the highest fitness value is selected as the optimal solution, and the corresponding process parameter combination is output.

[0163] The optimal process parameter combination is input into the regression model to predict the corresponding plating thickness and hole filling rate as the optimization result.

[0164] The above is the detailed process of using genetic algorithm to optimize micro blind via electroplating process parameters. This process covers the key steps of genetic algorithm, including population initialization, fitness evaluation, selection, crossover, mutation, new population generation and termination condition judgment. Through continuous iterative optimization, the optimal process parameter combination is finally obtained when the predicted electroplating thickness and hole filling rate are maximized.

[0165] In practical applications, it is also necessary to reasonably set the algorithm parameters (such as population size, crossover probability, mutation probability, etc.) according to specific problems, and verify the effectiveness and convergence of the algorithm through experiments. In addition, it is also possible to consider introducing some improvement strategies, such as elite retention strategy, adaptive parameter adjustment strategy, etc., to further improve the optimization performance of the algorithm.

[0166] Taking the electroplating thickness and hole filling rate as the optimization targets, the genetic algorithm is used to optimize the process parameters. By defining the weighted combination objective function and setting constraints, the optimal combination of process parameters is searched and obtained while meeting the process requirements, thereby improving the electroplating quality.

[0167] S600 adjusts the pre-treatment conditions and electroplating parameters of micro blind holes in real time during production according to the optimal process parameter combination

[0168] The optimal process parameter combination obtained by optimization is applied to the actual production process, and the pre-treatment conditions of the micro blind holes (such as degumming time, temperature, micro-etching time, temperature, washing time) and electroplating parameters (such as electroplating solution composition, temperature, current density, electroplating time) are adjusted in real time.

[0169] The plating thickness and hole filling rate in the production process are continuously monitored and compared with the predicted values. If there is a large deviation between the actual value and the predicted value, new data can be collected, the regression model can be retrained, and the process parameters can be optimized to achieve closed-loop optimization and control.

[0170] Example 2

[0171] like Figure 3 As shown, this embodiment provides a process parameter adaptive optimization system for HDI printed circuit board manufacturing based on embodiment 1, and the same points as the embodiment are not described again.

[0172] The process parameter adaptive optimization system 400 for HDI printed circuit board manufacturing includes:

[0173] A data acquisition module, which acquires the morphology and hole wall state characteristic image data of the historical micro blind holes and the corresponding historical micro blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate;

[0174] A convolutional neural network model processing module uses a convolutional neural network to analyze the morphology and hole wall state feature image data of historical micro blind holes, and extracts the micro blind hole and hole wall feature vectors of the morphology and hole wall state feature image data of the micro blind holes;

[0175] A first regression model training module is used to train a first regression model based on the micro-blind hole and hole wall feature vectors, historical micro-blind hole electroplating numerical parameters, and corresponding historical electroplating thickness and hole filling rate, wherein the output of the first regression model is electroplating thickness and hole filling rate;

[0176] The optimization module uses genetic algorithm to optimize the process parameters with electroplating thickness and hole filling rate as the target to obtain the optimal process parameter combination.

[0177] Example 3

[0178] Figure 4 1 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Figure 4 As shown, according to another aspect of the present application, an electronic device 700 is also provided. The electronic device 700 may include one or more processors and one or more memories. The memories store computer readable codes, and when the computer readable codes are run by one or more processors, the process parameter adaptive optimization method described above may be executed.

[0179] The method or system according to the embodiment of the present application can also be used by Figure 4 The electronic device architecture shown in FIG. Figure 4 As shown, the electronic device 700 may include a bus 701, one or more CPUs 702, a read-only memory (ROM) 703, a random access memory (RAM) 704, a communication port 705 connected to a network, an input / output component 706, a hard disk 707, etc. The storage device in the electronic device 700, such as the ROM 703 or the hard disk 707, may store the process parameter adaptive optimization method provided in the present application. The process parameter adaptive optimization method may, for example, include: obtaining the historical micro-blind hole morphology and hole wall state characteristic image data and the corresponding historical micro-blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate; using a convolutional neural network to analyze the historical micro-blind hole morphology and hole wall state characteristic image data, and extracting the micro-blind hole and hole wall characteristic vectors of the micro-blind hole morphology and hole wall state characteristic image data; training a first regression model based on the micro-blind hole and hole wall characteristic vectors, historical micro-blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate, and the first regression model outputs the electroplating thickness and hole filling rate; using a genetic algorithm to optimize the process parameters with the electroplating thickness and hole filling rate as the target to obtain the optimal process parameter combination. Further, the electronic device 700 may also include a user interface 708. Of course, Figure 4 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 4 One or more components of an electronic device are shown.

[0180] Example 4

[0181] Figure 5 Schematic diagram of a computer-readable storage medium structure provided by an embodiment of the present application. Figure 5As shown, a computer-readable storage medium 800 according to an embodiment of the present application is shown. Computer-readable instructions are stored on the computer-readable storage medium 800. When the computer-readable instructions are executed by the processor, the process parameter adaptive optimization method according to the embodiment of the present application described with reference to the above figures can be executed. The storage medium 800 includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0182] In addition, according to the embodiments of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, which stores machine-readable instructions, and the machine-readable instructions can be executed by a processor to execute instructions corresponding to the method steps provided in the present application, such as: obtaining the morphology and hole wall state feature image data of the historical micro blind holes and the corresponding historical micro blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate; using a convolutional neural network to analyze the morphology and hole wall state feature image data of the historical micro blind holes, and extracting the micro blind hole and hole wall feature vectors of the morphology and hole wall state feature image data of the micro blind holes; based on the micro blind hole and hole wall feature vectors, the historical micro blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate, a first regression model is trained, and the output of the first regression model is the electroplating thickness and hole filling rate; with the electroplating thickness and hole filling rate as the target, a genetic algorithm is used to optimize the process parameters to obtain the optimal process parameter combination. When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0183] The methods, apparatuses, and devices of the present application may be implemented in many ways. For example, the methods, apparatuses, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0184] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0185] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A process parameter adaptive optimization method for HDI printed circuit board manufacturing, characterized in that: include: Obtaining the morphology and hole wall state characteristic image data of the historical micro blind hole and the corresponding historical micro blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate; The convolutional neural network is used to analyze the morphology and hole wall state characteristic image data of the historical micro blind holes, and the micro blind hole and hole wall characteristic vectors of the morphology and hole wall state characteristic image data of the micro blind holes are extracted; A first regression model is trained based on the micro-blind hole and hole wall feature vectors, historical micro-blind hole electroplating numerical parameters, and corresponding historical electroplating thickness and hole filling rate, wherein the output of the first regression model is electroplating thickness and hole filling rate; With the electroplating thickness and hole filling rate as the target, the genetic algorithm is used to optimize the process parameters to obtain the optimal process parameter combination; The convolutional neural network is an attention-enhanced convolutional neural network model; The attention-enhanced convolutional neural network model training method is: Step S201: using the filtered historical micro blind hole morphology and hole wall state feature image data as training data and representing them; Step S202: constructing a spatial attention enhanced convolutional neural network model, inputting the training data of step S1 into the spatial attention enhanced convolutional neural network model, and extracting high-level features of micro blind hole morphology and hole wall state through convolution layers, pooling layers and spatial attention mechanisms; Step S203: designing a weighted cross entropy loss function to assign different weights to samples of different quality levels; Step S204: Use the Adam optimization algorithm to train the spatial attention enhanced convolutional neural network model of step S202, minimize the loss function of step S203, and iteratively update the model parameters; Step S205: using a K-fold cross-validation strategy to evaluate and select the spatial attention enhanced convolutional neural network model trained in step S204, to obtain a final micro blind hole morphology and hole wall state feature extraction model; The weighted cross entropy loss function is: ； in: Indicates the quality level of micro blind holes The weight is set according to the inverse of the number of samples, that is, ; Quality grade of micro blind holes The number of samples, Indicates a micro blind hole sample Belong to the quality level The real label, Represents the SAE-CNN model predicting the micro blind hole sample Belong to the quality level probability.

2. The process parameter adaptive optimization method for HDI printed circuit board manufacturing according to claim 1, characterized in that: The numerical parameters of micro blind hole electroplating include degumming time and temperature, micro-etching time and temperature, water washing time, electroplating solution composition, electroplating solution temperature and electroplating solution pH value, current density, electroplating time and conductivity.

3. The process parameter adaptive optimization method for HDI printed circuit board manufacturing according to claim 1, characterized in that: Taking the electroplating thickness and hole filling rate as the target, the steps of optimizing the process parameters by using genetic algorithm to obtain the optimal process parameter combination include: generating a plurality of first micro blind hole electroplating numerical parameter combinations; Inputting several groups of first micro blind hole electroplating numerical combination parameters into the first regression model respectively, predicting the corresponding electroplating thickness and hole filling rate, and obtaining the corresponding fitness value according to the corresponding electroplating thickness and hole filling rate; A genetic algorithm is used to perform selection, crossover and mutation operations on each group of first micro blind hole electroplating numerical parameters to generate optimal micro blind hole electroplating numerical parameters.

4. The process parameter adaptive optimization method for HDI printed circuit board manufacturing according to claim 1, characterized in that: include, Based on the indexes of pixel micro blind hole aperture size, pixel micro blind hole wall roughness, pixel micro blind hole bottom flatness and pixel micro blind hole electroplating uniformity, the morphology and hole wall state characteristic image data of historical micro blind holes are comprehensively screened.

5. A process parameter adaptive optimization system for HDI printed circuit board manufacturing, which is used to implement any process parameter adaptive optimization method in claims 1-4, characterized in that: include: A data acquisition module, which acquires the morphology and hole wall state characteristic image data of the historical micro blind holes and the corresponding historical micro blind hole electroplating numerical parameters and the corresponding historical electroplating thickness and hole filling rate; A convolutional neural network model processing module uses a convolutional neural network to analyze the morphology and hole wall state feature image data of historical micro blind holes, and extracts the micro blind hole and hole wall feature vectors of the morphology and hole wall state feature image data of the micro blind holes; A first regression model training module is used to train a first regression model based on the micro-blind hole and hole wall feature vectors, historical micro-blind hole electroplating numerical parameters, and corresponding historical electroplating thickness and hole filling rate, wherein the output of the first regression model is electroplating thickness and hole filling rate; The optimization module uses genetic algorithm to optimize the process parameters with electroplating thickness and hole filling rate as the target to obtain the optimal process parameter combination.

6. An electronic device, characterized in that The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the process parameter adaptive optimization method as claimed in any one of claims 1 to 4 are implemented.

7. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the steps in the process parameter adaptive optimization method according to any one of claims 1 to 4.

Images