Bonding pressure dynamic adjustment method and device and storage medium

Abstract

The invention discloses a bonding pressure dynamic adjustment method and device and a storage medium, and belongs to the technical field of semiconductor manufacturing and packaging, and the method comprises the steps: obtaining bonding surface image data and technological parameters in a bonding process; inputting the bonding surface image data into a first model for feature extraction to obtain bonding surface features; inputting the process parameters into a second model for feature extraction to obtain process parameter features; performing feature fusion processing on the bonding surface features and the process parameter features, and outputting bonding pressure adjustment parameters; and dynamically adjusting the pressure parameter of the bonding equipment according to the bonding pressure adjusting parameter. According to the method, the multi-source data fusion and machine learning technology is adopted, more accurate and adaptive dynamic adjustment of the bonding pressure is realized, the limitation of a traditional method is effectively solved, and the bonding yield and the stability and consistency of the bonding quality are improved.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing and packaging technology, specifically to a method, apparatus and storage medium for dynamic adjustment of bonding pressure. Background Technology

[0002] In semiconductor manufacturing, bonding is a core process that uses physical or chemical methods to electrically connect and mechanically fix the substrate to other components (such as substrates, lead frames, etc.), and its quality directly affects device reliability. Bonding process parameters include bonding pressure, temperature, time (holding pressure / heating / cooling duration), and bonding head displacement signals. Traditional bonding process parameters rely on fixed parameter settings and manual experience, making it difficult to adapt to changes in different materials and operating conditions, and the lack of real-time monitoring leads to low bonding yields. Summary of the Invention

[0003] This application provides a method, apparatus, and storage medium for dynamic adjustment of bonding pressure, which can dynamically adjust the pressure parameters of the bonding equipment by real-time acquisition of bonding surface image data and process parameters, followed by feature extraction and fusion processing.

[0004] Firstly, a method for dynamically adjusting bonding pressure is provided, comprising:

[0005] Obtain bonding surface image data and process parameters during the bonding process;

[0006] The bonding surface image data is input into the first model for feature extraction to obtain the bonding surface features;

[0007] The process parameters are input into the second model for feature extraction to obtain process parameter features;

[0008] The bonding surface features and the process parameter features are fused together to output bonding pressure adjustment parameters.

[0009] The pressure parameters of the bonding equipment are dynamically adjusted according to the bonding pressure adjustment parameters.

[0010] Secondly, a bonding pressure dynamic adjustment device is also provided, comprising:

[0011] The data acquisition module is configured to acquire bonding surface image data and process parameters during the bonding process;

[0012] The first feature extraction module is configured to input the bonding surface image data into the first model for feature extraction to obtain bonding surface features;

[0013] The second feature extraction module is configured to input the process parameters into the second model for feature extraction to obtain process parameter features.

[0014] The feature fusion module is configured to perform feature fusion processing on the bonding surface features and the process parameter features, and output bonding pressure adjustment parameters.

[0015] The pressure regulation module is configured to dynamically adjust the pressure parameters of the bonding equipment according to the bonding pressure regulation parameters.

[0016] Thirdly, a computer-readable storage medium is also provided, on which a computer program is stored, which is loaded by a processor to perform the steps in the above-described dynamic adjustment method for bonding pressure.

[0017] Beneficial effects:

[0018] This application acquires real-time image data of the bonding surface and process parameters from the bonding equipment. The image data is input into a first model to extract bonding surface features (such as solder joint contours, abnormal areas, and porosity), while the process parameters (such as pressure, temperature, and displacement) are input into a second model to extract temporal features. These two types of features are then fused to output precise bonding pressure adjustment parameters, which are fed back to the equipment's actuator in real time, thereby achieving dynamic optimization and adjustment of pressure parameters. This application has the following technical advantages: First, multi-source data fusion overcomes the limitations of traditional single-parameter adjustment, making the pressure setting more closely match the actual process conditions. Second, the feature extraction capability based on the machine learning model can effectively identify the complex correlation between microscopic defects on the bonding surface and process parameters. Finally, the entire adjustment process forms a closed-loop control, improving bonding yield and the stability and consistency of bonding quality. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a method for dynamically adjusting bonding pressure provided in some embodiments of this application;

[0021] Figure 2 This is a schematic diagram of the process for training the first model provided in some embodiments of this application;

[0022] Figure 3 This is a schematic diagram of the process for training a second model provided in some embodiments of this application;

[0023] Figure 4 This is a structural block diagram of a bonding pressure dynamic adjustment device provided in some embodiments of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0026] "A and / or B" includes the following three combinations: A only, B only, and a combination of A and B.

[0027] The use of "applies to" or "configured to" in this application implies open and inclusive language, which does not exclude the applicability to or configuration to devices performing additional tasks or steps. Additionally, the use of "based on" implies openness and inclusivity, because processes, steps, calculations, or other actions "based on" one or more of the stated conditions or values ​​may in practice be based on additional conditions or values ​​beyond those stated.

[0028] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0029] In the bonding process, specific process parameters, such as pressure and temperature, are applied to form a reliable electrical and mechanical connection between the substrate (such as a chip or wafer) and the substrate. The magnitude of the pressure directly affects the contact quality of the bonding interface: too little pressure leads to incomplete bonding, resulting in poor solder joints or excessive contact resistance; too much pressure causes substrate deformation or damage, producing microcracks and porosity. Furthermore, due to differences in material properties, ambient temperature fluctuations, and the surface morphology of the substrate during bonding, static pressure parameters are difficult to adapt to dynamic process requirements. Therefore, real-time monitoring of the bonding status and dynamic adjustment of pressure parameters are necessary to achieve a highly reliable, low-defect bonding interface, which plays a crucial role in improving semiconductor packaging yield and product lifespan.

[0030] Existing bonding technologies primarily rely on fixed parameter settings or manual adjustments based on experience, making it difficult to adapt in real-time to dynamic changes in process variables such as material properties and environmental conditions. Furthermore, the lack of effective real-time monitoring methods prevents the timely detection of quality defects during the bonding process (such as poor solder joints and porosity), resulting in persistently high defect rates. In addition, existing automated equipment generally lacks intelligent feedback adjustment mechanisms, exhibiting significant limitations when handling complex conditions such as substrate warping and material differences, severely restricting the stability and adaptability of the bonding process. These shortcomings collectively lead to technical bottlenecks such as low production efficiency and large fluctuations in product quality. Based on this, this application provides a method, apparatus, and storage medium for dynamic adjustment of bonding pressure, which can dynamically adjust the pressure parameters of the bonding equipment by real-time acquisition of bonding surface image data and process parameters, followed by feature extraction and fusion processing.

[0031] On the one hand, this embodiment provides a method for dynamically adjusting bonding pressure, such as... Figure 1 As shown, it includes:

[0032] S110: Obtain bonding surface image data and process parameters for the bonding process.

[0033] Understandably, industrial cameras (such as infrared cameras and infrared interferometers) can capture real-time microscopic image data of the bonding surface (e.g., weld joint morphology, porosity distribution, etc.), while simultaneously using pressure sensors, temperature sensors, and displacement sensors to collect time-series data of process parameters (including real-time pressure values, temperature profiles, and bonding head displacement). These two types of data together form the input foundation for the deep learning model: image data is used to analyze bonding quality defect characteristics, while process parameter data is used to monitor equipment operating status. The combination of these two provides data support for subsequent feature extraction and dynamic adjustment. The purpose of this step is to achieve simultaneous acquisition of multimodal data from the bonding process, overcoming the limitations of traditional methods that rely on only a single parameter source.

[0034] For example, the dynamic adjustment method for bonding pressure further includes data preprocessing: firstly, noise filtering and outlier cleaning are performed on the collected sensor signals such as bonding pressure, temperature, and displacement to ensure data accuracy; simultaneously, the bonding surface images acquired by the industrial camera are standardized (e.g., normalized, size-unified). For time-series process parameter data, a sliding window method is used to extract time-series features and perform normalization; for image data, data augmentation is performed by adding noise, rotation, and flipping to improve the model's generalization ability. Finally, the processed multi-source data (such as sensor signals, image features, and process parameters) are unified in dimensions and a feature vector is constructed.

[0035] S120: Input the bonding surface image data into the first model for feature extraction to obtain the bonding surface features.

[0036] Understandably, a pre-trained first model, such as a convolutional neural network (CNN) model, can intelligently analyze and process real-time acquired bonding surface images (e.g., solder joint morphology, porosity distribution, etc.). The first model automatically extracts multi-level visual features from the images (including micro-texture, edge contours, abnormal regions, etc.). These features can quantitatively characterize the current quality state of the bonding surface (e.g., solder joint shape, porosity, and other defect features). This step aims to overcome the limitations of traditional manual visual inspection, achieving intelligent diagnosis of bonding quality and providing an objective basis for subsequent dynamic pressure adjustment. In particular, for complex conditions such as substrate warping, the first model can also perform partitioned feature extraction, supporting differentiated local pressure adjustment. Meanwhile, the substrate includes, but is not limited to, chips, wafers, etc.

[0037] S130: Input the process parameters into the second model for feature extraction to obtain process parameter features.

[0038] Understandably, real-time data on process parameters such as pressure, temperature, and displacement collected from sensors in the bonding equipment are input into a pre-trained second model, such as a recurrent neural network (e.g., Long Short-Term Memory Neural Network (LSTM) or Gated Recurrent Unit Network (GRU), for deep analysis. This second model can automatically extract key dynamic features from the process parameters, including: 1) pressure change trends (e.g., pressure fluctuation amplitude and frequency); 2) temperature-pressure coupling characteristics (reflecting the correlation of material thermodynamic properties); and 3) displacement-pressure synergistic characteristics (characterizing the bonding head's motion state). These features accurately depict the deviation between the current process state and standard operating conditions, providing a quantitative basis for subsequent intelligent decision-making. Compared to traditional fixed-parameter control, this time-series feature extraction-based method enables the system to identify subtle anomalies in process parameters (e.g., differences in material properties), forming the technological foundation for dynamic adjustment.

[0039] S140: Perform feature fusion processing on the bonding surface features and the process parameter features to output bonding pressure adjustment parameters.

[0040] Understandably, the visual features of the bonding surface extracted in step S120 (such as image features like solder joint morphology and porosity distribution) and the temporal features of process parameters extracted in step S130 (such as dynamic features like pressure fluctuations, temperature changes, and displacement changes) are integrated into multimodal data through a feature fusion layer. Specifically, this includes the following processing: First, spatial-temporal feature alignment is achieved through feature stitching or an attention mechanism; then, a fully connected network or cross-attention model is used to mine the deep correlation between the two types of features (e.g., the correspondence between solder joint anomalies and pressure fluctuations in the image); finally, a quantified bonding pressure adjustment value (e.g., ±5% pressure correction) is output through a regression network. The purpose of this step is: 1) to overcome the limitations of traditional methods' single-dimensional adjustment, achieving a dual decision-making basis of "visual quality + process status" (capable of dynamically optimizing bonding pressure); 2) to address the substrate warping problem, enabling differentiated pressure output across zones through feature weight allocation; and 3) to automatically compensate for optimal pressure parameters through feature interaction analysis to address material differences.

[0041] S150: Dynamically adjust the pressure parameters of the bonding equipment according to the bonding pressure adjustment parameters.

[0042] Understandably, the bonding pressure adjustment parameters (such as pressure correction values ​​and zoned pressure distribution) output by the deep learning model are fed back to the pressure control system of the bonding equipment in real time, achieving closed-loop adjustment through the following methods: First, the digital signal is converted into physical control commands (such as voltage / current signals) to drive the pressure actuator (such as a servo motor or piezoelectric ceramic) to precisely adjust the applied pressure of the bonding head; for substrate warping, the system will perform zoned pressure differential adjustment (such as increasing the pressure by 10% in the central area and decreasing the pressure by 5% in the edge area); for material differences, global pressure compensation is implemented (such as an overall increase of 8% to adapt to high-hardness materials). The purpose of this step is to: 1) achieve closed-loop control from "data analysis" to "physical execution" (achieving online prediction and real-time adjustment); 2) solve the defect of traditional processes that "cannot respond to changes in operating conditions in real time after parameter setting" through a dynamic feedback mechanism; 3) support the continuous optimization capability of the reinforcement learning model, so that the adjustment accuracy continues to improve with the accumulation of production data.

[0043] In some embodiments, step S140 involves performing feature fusion processing on the bonding surface features and the process parameter features to output bonding pressure adjustment parameters, including:

[0044] S141: Concatenate and splice the bonding surface features and process parameter features into a joint feature vector.

[0045] S142: Input the joint feature vector into the reinforcement learning model, wherein: the joint feature vector is used as the state input; the action value output by the reinforcement learning model is the change value of the bonding pressure adjustment amount; and the reward function is calculated based on the real-time detected bonding strength value and / or pore distribution density.

[0046] S143: Output the bonding pressure adjustment parameter according to the action value.

[0047] Understandably, the feature fusion processing in step S140 can be implemented through the following steps to achieve intelligent pressure regulation decision-making: First, the visual features of the bonding surface extracted by the first model (such as weld joint morphology, porosity distribution, etc.) and the temporal features of the process parameters extracted by the second model (such as recurrent neural network) (such as pressure fluctuation curve, temperature gradient, etc.) are spliced ​​together to form a joint feature vector containing spatial-temporal information; then, this joint feature is used as the state input of the reinforcement learning model (such as the deep deterministic policy gradient (DDPG) algorithm or the proximal policy optimization (PPO) algorithm). The reinforcement learning model outputs the action value of the pressure regulation amount (such as ±5% pressure change) through the policy network, and at the same time, it uses the real-time detected bonding strength (shear force test value) and / or porosity (image analysis result) to construct a reward function (such as +0.1 reward for every 1% reduction in porosity) to drive the reinforcement learning model to learn the optimal regulation strategy; finally, the action value is converted into executable bonding pressure regulation parameters (such as servo motor control voltage). The purpose of steps S141-S143 is to: 1) achieve multimodal fusion of "visual-process parameters" through feature cascading; 2) solve the pressure adaptation problem caused by different material properties / processes by using a reinforcement learning reward mechanism; and 3) enable the system to continuously optimize the adjustment strategy through dynamic reward feedback (achieving the technical effect of continuous optimization and iteration).

[0048] For example, when the system detects that a batch of substrates has insufficient solder joint diffusion in the bonding image due to high material hardness (visual feature extracted by the first model) and the real-time pressure curve shows abnormal pressure feedback (temporal feature extracted by the second model), the two types of features are concatenated into a joint vector; the PPO algorithm is selected to output an action value such as "pressure +8%", and a +0.2 reward is given based on the improvement effect of reducing the current porosity from, for example, 0.5% to 0.3%; finally, the action value is converted into a 12V voltage signal to drive the servo motor to increase pressure. This process demonstrates: 1) how visual-process feature fusion can identify differences in material properties; 2) how to optimize adjustment strategies through real-time feedback of porosity reduction; and 3) how to automatically adapt to different materials / processes without human intervention.

[0049] In some other embodiments, the bonding pressure dynamic adjustment method further includes:

[0050] When substrate warping is detected, the bonding surface image is divided into M×N independent adjustment regions; where M and N are positive integers.

[0051] Extract the partitioned bond surface features for each independent regulation region;

[0052] The bonding surface features of each partition are fused with the process parameter features to output the bonding pressure adjustment parameters for each partition.

[0053] Understandably, intelligent partitioned pressure adjustment methods for substrate warping involve the following steps: When the system detects substrate warping through image analysis, it first divides the bonding surface image into an M×N grid region (e.g., a 3×3 partition). Then, it extracts the microscopic features of each partition (such as local solder joint morphology and gap size) using a first model. Next, it fuses the visual features of each partition extracted by the first model with the process parameter features extracted by the second model (such as overall temperature and reference pressure), ultimately outputting differentiated partitioned pressure adjustment parameters. For example, the pressure in the central warped region can be automatically increased by 15%, while the pressure in the flat edge region remains unchanged, thus achieving a dynamic partitioned pressurization strategy. This method overcomes the limitations of the traditional bonding process's globally uniform pressure, compensating for substrate deformation through spatially adaptive pressure distribution, making it suitable for scenarios requiring high-precision packaging.

[0054] In other embodiments, the bonding pressure dynamic adjustment method further includes: calculating defect occurrence probability features based on bonding surface image data using a first model; performing feature fusion processing on the defect occurrence probability features and process parameter features to output a pressure compensation coefficient; wherein the pressure compensation coefficient is used to correct the bonding pressure adjustment parameters.

[0055] Understandably, a defect prediction-based intelligent pressure compensation mechanism can be implemented using deep learning: First, a first model (such as CNN) analyzes real-time bonding images to calculate the probability of occurrence of key defects (such as porosity, solder joint misalignment, etc.), generating a quantified probability feature vector. Then, this probability feature is fused with the time-series features of process parameters (such as pressure fluctuations and temperature changes) extracted by a second model (such as LSTM / GRU) through multimodal fusion. The correlation between defects and process parameters is analyzed using a fully connected network or attention mechanism (e.g., if an increase in porosity probability is detected, and a simultaneous decrease in pressure is observed, then pressure compensation is required). Finally, a pressure compensation coefficient (e.g., +0.08 represents an 8% increase in baseline pressure) is output to dynamically correct the original adjustment parameters. The purpose of this method is: 1) to break through the traditional method's reliance on passive adjustment of explicit defects and achieve predictive compensation for defect risks; 2) to identify implicit process mismatch problems through probability features in material / process difference scenarios; and 3) to combine reinforcement learning mechanisms to continuously optimize the compensation strategy based on production data.

[0056] In some other specific examples, in step S150, dynamically adjusting the pressure parameters of the bonding device according to the bonding pressure adjustment parameters includes: superimposing the adjustment amount corresponding to the pressure compensation coefficient on the bonding pressure adjustment parameters to generate the final pressure parameters of the bonding device.

[0057] Understandably, the basic bonding pressure adjustment parameters output by the deep learning model (such as the partitioned pressure adjustment value) are intelligently superimposed with the calculated pressure compensation coefficient (such as +8% correction) to generate the final execution command. Specifically, when the system detects that the probability of porosity on the bonding surface in a certain area rises to a threshold (such as 30%), a compensation amount (such as +5%) is generated based on historical data; at the same time, a partitioned pressure suggestion value (such as +10%) is output for the substrate warping situation in that area; finally, the two are superimposed to generate the pressure parameters of the bonding equipment (such as 10% + 5% = 15%), which is then executed by the servo system. The purpose of this method is to: 1) achieve dual optimization of "defect prevention compensation + working condition adaptive adjustment" (such as the joint processing of warping partitioning and porosity early warning); 2) enable the compensation coefficient to be dynamically optimized with the production process through the reward mechanism of the reinforcement learning model; 3) effectively solve the overpressure / underpressure problem caused by the lack of compensation mechanism in traditional processes in material / process difference scenarios.

[0058] In some other embodiments, the bonding pressure dynamic adjustment method further includes: when a bonding surface image cannot be obtained, inputting the substrate surface data of the previous process into a first model to extract the bonding surface features of the previous process; performing feature fusion processing on the bonding surface features of the previous process and the process parameter features, and outputting bonding pressure adjustment parameters.

[0059] Understandably, this example presents an alternative solution for application scenarios where real-time image acquisition equipment cannot be installed: when image acquisition equipment such as infrared cameras / infrared interferometers is unavailable, the system automatically switches to a backup mode, using substrate surface data (such as warpage and thickness distribution) obtained from previous processes (such as grinding and cutting processes) as input, and extracting equivalent bonding surface features using a first model. Specifically: 1) Features (such as regional warpage gradient and surface roughness) are extracted from the 3D scanning data of previous processes using a pre-trained first model to replace real-time image features; 2) These features are fused with current process parameters (such as temperature and reference pressure); 3) Pressure adjustment parameters are output (such as increasing pressure by 12% for high warpage areas). The purpose of this method is to: 1) expand the applicability of the invention; 2) achieve near real-time monitoring by utilizing the data correlation between processes (such as predicting bonding pressure requirements through previous surface data); 3) maintain dynamic optimization capabilities even when fusing previous features with process parameters within a reinforcement learning framework.

[0060] Before performing the aforementioned data acquisition, preprocessing, and feature fusion, the first and second models require specialized training and optimization. The first model is primarily trained on bond surface image data, learning to extract effective visual features from a large number of labeled bond defect samples (such as porosity and solder joint misalignment). The second model is trained based on time-series data of (historical) process parameters, capturing the time dependence of parameters such as pressure and temperature. Both models are optimized using cross-entropy loss and mean squared error loss, respectively, and the network weights are adjusted through backpropagation to ensure accurate extraction of feature information for their respective modalities, providing reliable input for subsequent feature fusion and dynamic pressure adjustment.

[0061] In some embodiments, such as Figure 2 As shown, the bonding pressure dynamic adjustment method further includes:

[0062] Training the first model includes:

[0063] S201: Obtain bonding surface image samples and corresponding bonding quality label data during the bonding process. The bonding quality label data includes the solder joint morphology defect level and / or porosity distribution density.

[0064] S202: Perform normalization and sample augmentation operations on the bonding surface image samples.

[0065] S203: Extract contour features and abnormal region features of bonding surface image samples using the first model.

[0066] S204: Calculate the deviation between the predicted classification result and the bonded quality label data using the cross-entropy loss function, and optimize the network parameters of the first model.

[0067] Understandably, the training process of the first model includes: firstly, collecting labeled bonding surface image samples and corresponding quality labels (e.g., solder joint defects are classified into 1-5 levels, and quantified porosity values); then, performing standardization preprocessing on the images (e.g., normalizing to the 0-1 range) and data augmentation (e.g., rotating / flipping to expand samples); automatically learning and extracting key visual features of the bonding surface through a CNN network (e.g., identifying solder joint integrity based on contour texture and detecting porosity locations in abnormal areas); finally, using the cross-entropy loss function and a backpropagation algorithm to continuously reduce the error between the model's predicted defect classification results (e.g., predicting "level 3 solder joint defect") and the actual labels (e.g., labeling "level 4 defect"), thereby optimizing the network weight parameters. This training method enables the first model to accurately identify substrate warping areas and material property differences, providing reliable visual feature input for subsequent dynamic pressure adjustment.

[0068] For example, for regression problems that predict optimal process parameters (such as target pressure and time), mean squared error (MSE) or Huber loss can be used as loss functions.

[0069] Mean Squared Error (MSE): The formula is as follows

[0070] Here, m and n represent the number of rows and columns of the image, respectively (in a bond surface image, this is used to calculate the error between the predicted value I(i,j) and the true value K(i,j) for each pixel (i,j) in the image). The performance of the model in regression tasks is measured by averaging the squared errors between the predicted and true values. For example, when predicting the curve of bond pressure changing over time, MSE can be used to evaluate the deviation between the model's predicted values ​​and the actual measured values.

[0071] Huber loss: The formula is as follows

[0072] It is a loss function that is less sensitive to outliers. When the error (difference between the predicted and true values) is less than or equal to δ, the Huber loss is equivalent to the mean squared error loss. When the error is greater than δ, the Huber loss becomes a linear function, which reduces the impact of outliers on the loss calculation. For example, in the prediction of bonding process parameters, if occasional sensor measurement outliers occur, the Huber loss can make the model training more robust.

[0073] For classification problems that determine bonding quality (such as whether the bonding surface has defects, the defect level, etc.), cross-entropy loss can be used as the loss function. Cross-entropy loss measures the difference between the probability distribution predicted by the model and the probability distribution of the true labels. By minimizing cross-entropy loss, the model can learn a more accurate classification decision boundary. For example, in determining whether solder joints in a bonding surface image are acceptable (acceptable / unacceptable), cross-entropy loss can help the model optimize classification performance.

[0074] For reinforcement learning tasks (such as the PPO algorithm for dynamic bond pressure regulation), training can be performed using a reward maximization strategy. In the bond pressure regulation scenario, the model takes action (outputs the pressure regulation amount) based on the current state (such as bond surface features and process parameter features), and then adjusts its strategy based on rewards from environmental feedback (such as rewards for increased bond strength and reduced porosity) to maximize long-term cumulative rewards, thereby learning the optimal pressure regulation strategy.

[0075] During training, model hyperparameters can be optimized using methods such as grid search or Bayesian optimization to achieve better training results. Grid search is an exhaustive search method. For example, for the hyperparameters of the first model, such as kernel size, number of layers, and learning rate, a range of values ​​for each hyperparameter is predefined. Then, each combination of hyperparameters is systematically traversed, the model performance is evaluated on the validation set, and the optimal hyperparameter combination is selected. Although this method is simple and direct, it has a high computational cost, especially when there are many hyperparameters. Bayesian optimization is a probabilistic optimization method. It constructs a probabilistic model of the objective function (the model's performance on the validation set) and, based on existing hyperparameter performance data, predicts the next hyperparameter combination most likely to improve performance for evaluation. Compared to grid search, Bayesian optimization is more efficient in handling high-dimensional hyperparameter spaces and can find better hyperparameter combinations faster.

[0076] In some embodiments, such as Figure 3 As shown, the bonding pressure dynamic adjustment method further includes:

[0077] Training the second model includes:

[0078] S301: Obtain process parameter samples for the bonding process and corresponding pressure adjustment tag data, wherein the process parameter samples include time-series data from pressure, temperature, and displacement sensors.

[0079] S302: Perform sliding window sampling on the time series data to construct time series features.

[0080] S303: The second model is used to train the time series features and pressure regulation amount label data.

[0081] S304: Calculate the deviation between the predicted pressure regulation amount and the pressure regulation amount label data using the mean square error loss function, and optimize the network parameters of the second model.

[0082] Understandably, the training process for the second model (such as LSTM / GRU) includes: first, collecting multi-dimensional process parameter samples of the bonding equipment (such as time-series data of pressure curves, temperature fluctuations, and displacement changes) and their corresponding manually optimized pressure adjustment label data; then, using a sliding window sampling method (e.g., a time window of 5 seconds), converting the continuous time-series data into feature segments with temporal correlation; inputting these features into the second model for training, the network learns the temporal dependencies between process parameters (e.g., the hysteretic effect of temperature rise on pressure); finally, using the mean squared error loss function (MSE), the deviation between the pressure adjustment predicted by the model (e.g., +8% pressure) and the labeled optimized adjustment is quantified, and the network parameters are continuously optimized through the backpropagation algorithm. This training method enables the second model to predict the pressure compensation requirements corresponding to substrate warping, as well as the differentiated pressure parameters required by different material / process characteristics, providing a reliable process feature basis for dynamic adjustment.

[0083] The bonding pressure dynamic adjustment device provided in this application is described below. The bonding pressure dynamic adjustment device described below can be referred to in correspondence with the bonding pressure dynamic adjustment method described above.

[0084] On the other hand, such as Figure 4 As shown, this embodiment provides a bonding pressure dynamic adjustment device, including a data acquisition module 401, a first feature extraction module 402, a second feature extraction module 403, a feature fusion module 404, and a pressure adjustment module 405.

[0085] For example, the data acquisition module 401 is configured to acquire bonding surface image data and process parameters of the bonding process.

[0086] For example, the first feature extraction module 402 is configured to input the bonding surface image data into the first model for feature extraction to obtain bonding surface features.

[0087] For example, the second feature extraction module 403 is configured to input the process parameters into the second model for feature extraction to obtain process parameter features.

[0088] For example, the feature fusion module 404 is configured to perform feature fusion processing on the bonding surface features and the process parameter features, and output bonding pressure adjustment parameters.

[0089] For example, the pressure regulation module 405 is configured to dynamically adjust the pressure parameters of the bonding device according to the bonding pressure regulation parameters.

[0090] Understandably, the bonding pressure dynamic adjustment device achieves intelligent pressure regulation through multimodal data collaborative processing: the data acquisition module acquires infrared images of the bonding surface in real time (e.g., capturing solder joint morphology and porosity distribution) and process parameters (e.g., pressure / temperature / displacement time-series data); the first model (e.g., CNN) extracts spatial features from the images (e.g., warp region partitioning features), and the second model (e.g., LSTM / GRU) analyzes the temporal correlation of process parameters (e.g., the delayed effect of temperature fluctuations on pressure); the feature fusion module intelligently associates the two types of features through a cross-attention mechanism (e.g., identifying the causal relationship between increased porosity and insufficient pressure), and finally outputs an accurate pressure compensation value; the pressure adjustment module adjusts the servo system in real time based on the fused multidimensional features (e.g., increasing the pressure in the warp region by 15%). This application breaks through the traditional fixed parameter mode and achieves dynamic optimization through closed-loop feedback of image and parameters: it can adjust the pressure for substrate warp partitioning and can adapt to different material properties, achieving the technical effects of dynamically optimizing bonding pressure and accurately monitoring bonding quality.

[0091] It should be noted that the bonding pressure dynamic adjustment device provided in this application embodiment can realize all the method steps implemented in the above method embodiment and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0092] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to execute the arrangement in any of the methods described above.

[0093] In the embodiments of this application, the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0094] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0095] The above provides a detailed description of a bonding pressure dynamic adjustment method, apparatus, and storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method of dynamically adjusting a bonding pressure, the method comprising: The method comprises the following steps: Obtaining bonding surface image data and process parameters of a bonding process; Inputting the bonding surface image data into a first model for feature extraction to obtain bonding surface features; Inputting the process parameters into a second model for feature extraction to obtain process parameter features; Performing feature fusion processing on the bonding surface features and the process parameter features to output a bonding pressure adjustment parameter; According to the bonding pressure adjustment parameter, dynamically adjusting the pressure parameter of the bonding equipment.

2. The method of bonded pressure dynamic adjustment of claim 1, wherein, Further comprising: Training the first model, including: Obtaining bonding surface image samples and corresponding bonding quality label data of a bonding process, the bonding quality label data including solder joint morphology defect level and / or pore distribution density; Performing normalization processing and sample expansion operation on the bonding surface image samples; Extracting contour features and abnormal area features of the bonding surface image samples through the first model; Calculating the deviation of the predicted classification result and the bonding quality label data with a cross-entropy loss function to optimize the network parameters of the first model.

3. The method of bonded pressure dynamic adjustment of claim 1, wherein, Further comprising: Training the second model, including: Obtaining process parameter samples and corresponding pressure adjustment amount label data of a bonding process, the process parameter samples including time series data of pressure, temperature, and displacement sensor; Performing sliding window sampling on the time series data to construct time series features; Training the time series features and the pressure adjustment amount label data through the second model; Calculating the deviation of the predicted pressure adjustment amount and the pressure adjustment amount label data with a mean square error loss function to optimize the network parameters of the second model.

4. The method of bonded pressure dynamic adjustment of claim 1, wherein, Performing feature fusion processing on the bonding surface features and the process parameter features to output a bonding pressure adjustment parameter includes: Concatenating the bonding surface features and the process parameter features into a joint feature vector; Inputting the joint feature vector into a reinforcement learning model, wherein the joint feature vector is input as a state, the action value output by the reinforcement learning model is the change value of the bonding pressure adjustment amount, and the reward function is calculated based on the real-time detected bonding strength value and / or pore distribution density; Outputting the bonding pressure adjustment parameter according to the action value.

5. The method of bonded pressure dynamic adjustment of claim 1, wherein, Further comprising: When detecting that the bonding substrate is warped, dividing the bonding surface image into M×N independent adjustment regions; Wherein M and N are positive integers; Extracting partition bonding surface features for each independent adjustment region; Performing feature fusion processing on the partition bonding surface features and the process parameter features to output bonding pressure adjustment parameters for each partition.

6. The method of bonded pressure dynamic adjustment of claim 1, wherein, Further comprising: Based on the bonding surface image data, calculating defect occurrence probability features through the first model; Performing feature fusion processing on the defect occurrence probability features and the process parameter features to output a pressure compensation coefficient; Wherein, the pressure compensation coefficient is used to correct the bonding pressure adjustment parameter.

7. The method of bonded pressure dynamic adjustment of claim 6, wherein, The dynamically adjusting the pressure parameter of the bonding equipment according to the bonding pressure adjustment parameter includes: Superimposing the bonding pressure adjustment parameter and the adjustment amount corresponding to the pressure compensation coefficient to generate the final pressure parameter of the bonding equipment.

8. The method of bonded pressure dynamic adjustment of claim 1, wherein, Further comprising: When the bonding surface image cannot be acquired, the substrate surface profile data of the previous process is input into the first model to extract the bonding surface feature of the previous process; The bonding surface feature of the previous process and the process parameter feature are subjected to feature fusion processing, and a bonding pressure adjustment parameter is output.

9. A bonding pressure dynamic adjustment apparatus, characterized by, Comprise: A data acquisition module configured to acquire bonding surface image data and process parameters of a bonding process; A first feature extraction module configured to input the bonding surface image data into a first model for feature extraction to obtain a bonding surface feature; A second feature extraction module configured to input the process parameters into a second model for feature extraction to obtain a process parameter feature; A feature fusion module configured to perform feature fusion processing on the bonding surface feature and the process parameter feature to output a bonding pressure adjustment parameter; A pressure adjustment module configured to dynamically adjust the pressure parameters of a bonding device according to the bonding pressure adjustment parameter.

10. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and the computer program is loaded by a processor to execute the steps in the bonding pressure dynamic adjustment method of any one of claims 1-8. A computer program is stored thereon, and the computer program is loaded by a processor to execute the steps in the bonding pressure dynamic adjustment method of any one of claims 1-8.

Images