A machine learning-based substrate warpage prediction method and system

Abstract

This invention relates to the field of chip packaging technology and proposes a substrate warpage prediction method and system based on machine learning. The method includes: acquiring the temperature of the substrate at several moments during temperature loading; obtaining several sampling moments and their corresponding DIC views for each temperature change stage; obtaining several nodes for various materials; obtaining the similar response time difference for each node; quantifying the interference response time difference for each node based on the change in pixel values ​​at corresponding positions in the DIC view as the sampling moments change; obtaining the similar influence degree of each node for various materials; obtaining the comprehensive response time of each node and constructing several response isochrones; acquiring the heat conduction distribution value of each pixel; training a substrate warpage prediction model; and thus predicting substrate warpage. This invention aims to solve the problem that the neural network training process for substrate warpage prediction models based on DIC measurement images learns a lot of irrelevant information.

Description

Technical Field

[0001] This invention relates to the field of chip packaging technology, and more specifically to a substrate warpage prediction method and system based on machine learning. Background Technology

[0002] Due to the difference in thermal expansion coefficients between components and the substrate, the shrinkage ratios differ during the reflow soldering cooling process, resulting in stress differences that cause the substrate to warp. The deformation is greatest at the four corners of the substrate. Traditional methods for measuring substrate warpage are divided into contact and non-contact methods. Contact methods include the pad method and contact probe methods. Contact methods are simple to operate but cannot dynamically measure the deformation process. Non-contact methods include the Moiré fringe method, optical interferometry, and DIC method. However, the Moiré fringe method requires close proximity to the target surface and is limited in high-temperature environments, while the optical interferometry method requires expensive equipment and is also unsuitable for high-temperature environments. DIC is a high-precision full-field measurement method based on the principle of binocular stereo vision. It reproduces the three-dimensional shape of the target object according to the camera's calibration parameters, enabling accurate warpage measurement.

[0003] Existing technologies extract features by preprocessing various processing parameters, input them into a neural network, and use the warp values ​​measured by DIC (Digital Curve Analyzer) as training labels for warp prediction. However, even though the features and the warp values ​​measured by DIC are combined for training, the content learned by the neural network model during training is not limited, which may cause the model to learn a lot of irrelevant information. Therefore, it is necessary to strengthen the connection between input and output data, that is, the connection between processing parameters and DIC measurement results, during training. Summary of the Invention

[0004] This invention provides a machine learning-based method and system for predicting substrate warpage, addressing the problem that existing neural network training models for substrate warpage prediction based on DIC measurement images learn a significant amount of irrelevant information. The specific technical solution adopted is as follows:

[0005] This invention proposes a machine learning-based method for predicting substrate warpage, which includes the following steps:

[0006] Obtain the temperature of the substrate at several moments during the temperature loading process;

[0007] The temperature change trend at each temperature change stage during the temperature loading process is analyzed to obtain several sampling times at each temperature change stage, and the DIC view at each sampling time is obtained through the DIC measurement unit; based on the distribution of different types of materials in the substrate wiring diagram, several nodes of each type of material are obtained through corner detection; the similarity relationship between the pixel values ​​at corresponding positions in the DIC view of different nodes of the same type of material with changes in sampling time is analyzed to obtain the response time difference of each node.

[0008] Based on the distribution of nodes in different types of materials, and combined with the changes in pixel values ​​at corresponding positions in the DIC view of nodes changing with sampling time, the interference response time difference of each node is quantified; based on the distribution of nodes in the same type of materials and their similar response time differences, the average response time of each type of material is obtained, and then the similar influence degree of each node in each type of material is obtained; combined with the similar response time difference and the interference response time difference, the comprehensive response time of each node is obtained, and several response isochrones are constructed.

[0009] Based on the distribution of pixels and response isochrones in the DIC view, the thermal conductivity distribution value of each pixel is obtained, and then several analysis image blocks are determined and their feature vectors are obtained through KSVD decomposition. The substrate warpage prediction model is trained to predict substrate warpage.

[0010] Optionally, the specific method for obtaining several sampling times for each temperature change stage includes:

[0011] For any temperature change phase, the slope at each moment is calculated based on the temperature difference between adjacent moments; the absolute value of the average slope at all moments in any temperature change phase is taken as the change coefficient of that temperature change phase; the sum of 1 plus the change coefficient is obtained; a preset unit interval reference value is set; the product of the sum and the unit interval reference value and rounded down is taken as the unit interval of that temperature change phase.

[0012] The first moment in the temperature change phase is taken as a sampling moment, and a sampling moment is obtained at every unit interval until the end of the temperature change phase, thus obtaining a number of sampling moments in the temperature change phase.

[0013] Optionally, the specific method for obtaining several nodes of various materials includes:

[0014] Obtain the substrate wiring diagram. The same materials in the substrate wiring diagram are marked with the same color. Perform corner detection on the area corresponding to any type of material to obtain several corner points. Use the obtained corner points as several nodes of that type of material.

[0015] Optionally, the specific method for obtaining the similar response time difference of each node includes:

[0016] For any node of any material, obtain the pixel value of the corresponding position of the node in the DIC view at each sampling time, construct a coordinate system with the sampling time as the horizontal axis and the pixel value in the DIC view as the vertical axis, map the pixel value of the corresponding position of the node in the DIC view at each sampling time to the coordinate system, obtain several data points, and connect them to form the change curve of the node.

[0017] Obtain the Euclidean distances between a given node and all other nodes in a set of nodes for this type of material. The node with the minimum Euclidean distance is designated as the matching node. Perform DTW matching on the change curve of this node and its matching nodes to obtain several matching time pairs. Subtract the sampling time of the matching node from the sampling time of any matching time pair, and use this difference as the response time difference for that matching time pair. The average of the response time differences for all matching time pairs is used as the response time difference for that node.

[0018] Optionally, the interference response time difference of each node is obtained by the following method:

[0019] For any node of any material class, and all nodes of any other material class, obtain the Euclidean distance between each node and the node, and take the node with the minimum Euclidean distance as the reference node of the node in the other material class.

[0020] Obtain several matching time pairs between the change curve of the node and the change curve of the reference node, and obtain the response time difference of each matching time pair. Use the average of the response time differences of all matching time pairs between the change curve of the node and the change curve of the reference node as the response time difference of the node under the influence of other types of materials.

[0021] The average of the response time differences of the node to all other types of materials other than the one in question is taken as the disturbance response time difference of the node.

[0022] Optionally, the specific methods for obtaining the average response time of various materials include:

[0023] For any node of any material class, if the node and its matching node are in the same region of the same material class, obtain the number of nodes in that region, and use the ratio of the number of nodes in that region to the total number of nodes in all material classes as the response factor of that node;

[0024] If the node and its matching node are in different regions of the same material, obtain the connection between the node and its matching node, and obtain the regions of other material types that the connection passes through. If the connection passes through regions of other material types but does not pass through nodes of other material types, the corresponding region count is 1. If the connection passes through regions of other material types and passes through nodes of other material types in the region, the number of nodes of other material types in the region is used as the count of the corresponding region. The ratio of the sum of the counts of all other material types that the connection passes through to the total number of nodes of all material types is used as the interference factor of the node. The difference obtained by subtracting the interference factor from 1 is used as the response factor of the node.

[0025] The response factors of all nodes in this type of material are weighted and normalized to obtain the response weight of each node. The response time differences of the same type of nodes are weighted and summed based on the response weights, and the result is used as the average response time of this type of material.

[0026] Optionally, the specific methods for obtaining the degree of similar influence at each node of various materials include:

[0027] Obtain the absolute value of the difference between the response time difference of each node in any type of material and the average response time of that type of material. Take the ratio of the absolute value of the difference corresponding to any node to the maximum value among all absolute values ​​of the difference as the degree of influence of that node in that type of material.

[0028] Optionally, the specific methods for obtaining the comprehensive response time of each node and constructing several response isochrones include:

[0029] For any node of any type of material, the degree of similar influence of the node is used as the weight of the similar response time difference, and the difference obtained by subtracting the degree of similar influence of the node from 1 is used as the weight of the disturbance response time difference. The similar response time difference and the disturbance response time difference of the node are weighted and summed, and the result is used as the comprehensive response time of the node.

[0030] The maximum and minimum values ​​of the comprehensive response time of all nodes of all materials are obtained. The difference between the maximum and minimum values ​​is divided by 10 and used as the classification threshold.

[0031] Based on the contour line construction process, isochrones are constructed for the comprehensive response time of each node, and these isochrones are used as response isochrones. If the absolute value of the difference between the comprehensive response times of each node on the same response isochrone does not exceed the classification threshold, then several response isochrones are constructed.

[0032] Optionally, the specific method for obtaining the heat conduction distribution value of each pixel and then determining several image blocks for analysis includes:

[0033] For any pixel in the DIC view at any sampling time, obtain the Euclidean distance between the pixel and each response isochrone. Take the response isochrone corresponding to the minimum Euclidean distance as the reference isochrone for the pixel. Obtain the average of the comprehensive response times of all nodes on the reference isochrone as the response time of the reference isochrone. Multiply the inversely proportional normalized result of the minimum Euclidean distance by the response time of the reference isochrone as the heat conduction distribution value of the pixel.

[0034] Obtain the heat conduction distribution value of each pixel except the pixel on the response isochron. Use the response time of each response isochron as the heat conduction distribution value of the pixel on it to form a heat conduction distribution map.

[0035] For any DIC view at any sampling time, the DIC view is divided into several image blocks, and the mean value of the heat conduction distribution of all pixels in any image block is obtained as the heat conduction comprehensive factor of the image block;

[0036] All image blocks are arranged in descending order of thermal conductivity comprehensive factor. A preset number of images are selected, and the first number of images selected are used as a number of images for analysis.

[0037] The present invention also proposes a substrate warpage prediction system based on machine learning, the system comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0038] The beneficial effects of this invention are as follows: This invention performs DIC measurement on a chip substrate, and based on the measured warpage information and the distribution information of various materials on the substrate, comprehensively analyzes the relationship between the time change of thermal conduction response at various locations on the substrate and the substrate warpage performance. Using this, KSVD decomposition is used to construct feature vectors, and a substrate warpage prediction model is trained to predict substrate warpage. Specifically, the invention obtains the response time difference of the same type and the interference response time difference, and performs matching analysis on the pixel values ​​in the DIC view between nodes of the same type of material as the sampling time changes, thereby quantifying the response time difference. Furthermore, it performs matching analysis on the pixel value changes in the DIC view between nodes of different types of materials to obtain the interference response time difference. Simultaneously, it considers whether the response time difference of the same type is affected by other types of materials, thereby quantifying the average response time of each type of material. Weights are then constructed to obtain the comprehensive response time of the nodes, and a response isochron isochron is constructed for subsequent acquisition of the thermal conduction distribution values ​​of pixels in the DIC view. This allows for the selection and analysis of image blocks to reference more thermal conduction information for neural network model learning, enabling the final substrate warpage prediction model to learn more about the relationship between thermal conduction information and warpage information, thus improving the accuracy of substrate warpage prediction. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of a machine learning-based substrate warpage prediction method provided in one embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figure 1 The diagram illustrates a flowchart of a machine learning-based substrate warpage prediction method according to an embodiment of the present invention. The method includes the following steps:

[0043] Step S001: Obtain the temperature of the substrate at several moments during the temperature loading process.

[0044] The purpose of this embodiment is to predict the thermal variables of the chip substrate under temperature changes, so as to achieve substrate warpage prediction. This requires measuring the thermal deformation of the chip substrate in the laboratory using a heating device, and then constructing a substrate warpage prediction model to achieve actual substrate warpage prediction.

[0045] Specifically, a heating device and a DIC measurement device are set up in the laboratory, including a DIC measurement unit and a temperature loading unit. The room temperature is 30°C, the chip size is 5mm, and the DIC measurement unit includes a camera, light source, calibration board and calibration device, etc. The temperature loading unit is a system that supports programmable heating and cooling.

[0046] Furthermore, the temperature loading program is set up as follows: the initial preheating stage is " ", meaning the temperature rises from 30℃ to 150℃ within 160 seconds; the preheating and heat preservation stage is " ", meaning the temperature is maintained at 150℃ for 160s to 520s; the welding heating stage is " ", meaning the temperature rises from 150℃ to 240℃ within 520s to 700s; the welding heat preservation stage is " ", meaning the temperature is maintained at 240℃ for 700 to 860 seconds; the cooling phase is " "This means that the temperature will drop back to 30℃ within 860s to 1100s; the temperature at each moment will be recorded during the temperature loading process."

[0047] It should be noted that since the substrate is composed of multiple materials with different coefficients of thermal expansion, the temperature is uneven and the expansion and contraction amounts vary in different areas of the substrate. Therefore, the relationship between reflow soldering process parameters and warpage is obtained by observing the different responses of the substrate to temperature changes, and the neural network learns this relationship. At the same time, feature parameters are extracted from the results measured by DIC and used as input data to better represent the relationship between reflow soldering process parameters and warpage, thereby improving the warpage prediction results.

[0048] It should be further explained that, after obtaining the changes in warpage after the reflow soldering process parameters change, the feature extraction process is adjusted based on the results of DIC measurement, and the extraction process can fully include response information. However, when extracting features in the existing technology, spatial response information is lacking, resulting in single feature parameters that do not express the characteristics of changes with process parameters. KSVD decomposition has good adaptive ability and can adaptively adjust during the feature extraction process, while also having a good ability to preserve local features. Therefore, the adaptive KSVD method is adopted to use feature vectors that better represent the relationship between warpage information and changes in process parameters for subsequent training.

[0049] Step S002: Analyze the temperature change trend at each temperature change stage during the temperature loading process, obtain several sampling times at each temperature change stage, and obtain the DIC view at each sampling time through the DIC measurement unit; based on the distribution of different types of materials in the substrate wiring diagram, obtain several nodes of each type of material through corner detection; analyze the similarity relationship between the pixel values ​​at corresponding positions in the DIC view of different nodes of the same type of material as the sampling time changes, and obtain the response time difference of the same type of node.

[0050] It should be noted that the complete reflow soldering process involves temperature changes, including a preheating stage (further divided into preheating and preheating holding stages), a soldering stage (soldering heating and soldering holding stages), and a cooling stage. During the soldering stage, a solid-liquid phase transition of the solder occurs. Therefore, the process changes at different stages are not uniform, resulting in non-uniform changes in the DIC displacement field. Referring to the concept of load steps in finite element analysis, different unit variation intervals need to be set for different stages. Smaller unit variation intervals are set for the nonlinear region where the solder undergoes a phase transition, while larger unit variation intervals are set for the holding stage. Then, DIC views are acquired based on these unit variation intervals, and the changes in the displacement field information in the DIC views are analyzed to represent the spatial response information. The heat conduction path represents the path of heat propagation in the reflow soldering process, which leads to uneven temperature changes in different areas of the substrate during heating / cooling, i.e., uneven temperature distribution on the substrate. Furthermore, because each material has a different coefficient of thermal expansion, the expansion or contraction of different materials in different areas will vary.

[0051] Preferably, in one embodiment of the present invention, the temperature change trend at each temperature change stage during the temperature loading process is analyzed to obtain several sampling times at each temperature change stage, and a DIC view at each sampling time is obtained through a DIC measurement unit. The specific method includes:

[0052] For each moment and its temperature, the temperature loading process has been divided into several temperature change stages. For any temperature change stage, the slope of each moment is calculated based on the temperature difference between adjacent moments. Specifically, the difference is obtained by subtracting the temperature of the previous moment from the temperature of the next moment. The ratio of the difference to the sampling time interval is used as the slope of the next moment. The slope of the first moment in each temperature change stage is calculated in combination with the temperature of the last moment of the previous temperature change stage. The slope of the first moment in the entire temperature loading process is set to be the same as the slope of the second moment.

[0053] Furthermore, the absolute value of the average slope of all moments in any temperature change phase is used as the change coefficient of that temperature change phase. The sum of 1 and the change coefficient is obtained. A preset unit interval reference value is set. In this embodiment, the unit interval reference value is described as 10 seconds. The product of the sum and the unit interval reference value and the result is rounded down as the unit interval of that temperature change phase. The first moment in that temperature change phase is taken as a sampling moment, and a sampling moment is obtained at every unit interval until the end of the temperature change phase, thereby obtaining a number of sampling moments of that temperature change phase. Based on each sampling moment of that temperature change phase, the DIC view is acquired through the DIC measurement unit.

[0054] It should be further noted that since the substrate is composed of multiple materials (silicon chip, FR4 / ceramic substrate, solder, copper foil, heat sink, etc.), each material has a different coefficient of thermal expansion. Therefore, the temperature change is not uniform and the thermal conduction response varies. Thus, when analyzing the displacement field change with temperature, it is necessary to determine the nodes in multiple regions based on the wiring structure of the substrate.

[0055] Preferably, in one embodiment of the present invention, based on the distribution of different types of materials in the substrate wiring diagram, several nodes of each type of material are obtained through corner detection, including the following specific method:

[0056] Obtain the substrate wiring diagram. In the substrate wiring diagram, the same material is marked with the same color. That is, the area marked with the same color corresponds to the same type of material. Perform corner detection on the area corresponding to any type of material to obtain several corner points. Use the obtained corner points as several nodes of that type of material.

[0057] It should be noted that the nodes obtained through corner detection are the intersection points between regions corresponding to different types of materials, that is, the interaction points between different types of materials.

[0058] Preferably, in one embodiment of the present invention, the method for analyzing the similarity relationship between the pixel values ​​at corresponding positions in the DIC view of different nodes of the same material changing with the sampling time to obtain the similar response time difference of each node includes:

[0059] For any node of any material, obtain the pixel value of the corresponding position of the node in the DIC view at each sampling time. The pixel value of each pixel in the DIC view represents the displacement of the substrate warping. Construct a coordinate system with the sampling time as the abscissa and the pixel value in the DIC view as the ordinate. Map the pixel value of the corresponding position of the node in the DIC view at each sampling time to the coordinate system to obtain several data points, and connect them to form the change curve of the node.

[0060] Furthermore, among several nodes of this type of material, the Euclidean distance between the node and other nodes is obtained. The node corresponding to the minimum Euclidean distance is taken as the matching node of the node. DTW matching is performed on the change curve of the node and the change curve of its matching node to obtain several matching time pairs. The difference between the sampling time of the node and the sampling time of its matching node in any matching time pair is taken as the response time difference of the matching time pair. The mean of the response time differences of all matching time pairs is taken as the same type of response time difference of the node.

[0061] It should be noted that if the thermal conductivity of similar materials is the same, the thermal conductivity response time difference between nodes of similar materials can be reflected by analyzing the response time difference between the nearest matching nodes.

[0062] Thus, the response time differences of various materials at different points are obtained.

[0063] Step S003: Based on the distribution of nodes in different types of materials, and combined with the changes in pixel values ​​at corresponding positions in the DIC view of the nodes as the sampling time changes, quantify the interference response time difference of each node; based on the distribution of nodes in the same type of materials and their similar response time differences, obtain the average response time of each type of material, and then obtain the similar influence degree of each node in each type of material; combined with the similar response time difference and the interference response time difference, obtain the comprehensive response time of each node, and construct several response isochrones.

[0064] It should be noted that since the thermal conductivity of similar materials is the same, the change in the response time difference of similar materials is used to determine the thermal conductivity of the node. If the change in the first response time difference is relatively regular with the change of the node, it can be said that the thermal conductivity of the node is basically the same as that of the node of similar materials, that is, it is greatly affected by the same type of materials. Conversely, it is greatly affected by other types of materials.

[0065] Preferably, in one embodiment of the present invention, based on the distribution of nodes of different types of materials and combined with the changes in pixel values ​​at corresponding positions in the DIC view of the nodes changing with sampling time, the interference response time difference of each node is quantified, including the following specific method:

[0066] For any node of any material class, and all nodes of any other material class, obtain the Euclidean distance between each node and the node. The node with the minimum Euclidean distance is taken as the reference node for that other material class. Following the same method, obtain several matching time pairs between the node's change curve and the reference node's change curve, also obtained through DTW matching. Calculate the response time difference for each matching time pair. The average of the response time differences for all matching time pairs between the node's change curve and the reference node's change curve is taken as the node's response time difference under the influence of that other material class. The average of the response time differences for the node under the influence of all other materials is taken as the node's disturbance response time difference.

[0067] It should be noted that after obtaining the similar response time difference and the interference response time difference, they respectively reflect the thermal response differences between similar materials and between different materials. For similar materials, it is necessary to further analyze their average response time. Based on the distribution relationship between the similar response time difference obtained at each node and its matching node, the more nodes of other types of materials there are, the lower the confidence of the similar response time difference. The average response time is obtained by weighting in this way, and the similar response time difference of each node is compared with the average response time to quantify the degree of similar influence of each type of material at each node.

[0068] Preferably, in one embodiment of the present invention, the average response time of various materials is obtained based on the distribution of nodes in similar materials and the time difference of similar responses, thereby obtaining the degree of similar influence of each node in various materials. The specific method includes:

[0069] For any node of any material class, if the node and its matching node are in the same region of the material class, the number of nodes in that region is obtained, and the ratio of the number of nodes in that region to the total number of nodes in all material classes is used as the response factor of the node. If the node and its matching node are in different regions of the material class, the connection between the node and its matching node is obtained, and the regions of other material classes that the connection passes through are obtained. If the connection passes through regions of other material classes but does not pass through nodes of other material classes, the corresponding region is counted as 1. If the connection passes through regions of other material classes and passes through nodes of other material classes in those regions, the number of nodes of other material classes passed through in those regions is used as the count of the corresponding region. The sum of the counts of all other material classes that the connection passes through is used as the ratio of the total number of nodes of all material classes to the total number of nodes in all material classes, and the difference between 1 and the interference factor is used as the response factor of the node. The response factors of all nodes in the material class are weighted and normalized to obtain the response weight of each node. The response time differences of the same type of nodes are weighted and summed based on the response weights, and the result is used as the average response time of the material class.

[0070] Furthermore, the absolute value of the difference between the response time difference of each node in this type of material and the average response time of this type of material is obtained. The ratio of the absolute value of the difference corresponding to any node to the maximum value among all absolute values ​​of the difference is taken as the degree of similar influence of this node in this type of material.

[0071] It should be noted that the average response time is compared and normalized using the maximum value to obtain the degree of influence of the same type. This is used as the weight of the response time difference of the same type for subsequent comprehensive response time acquisition.

[0072] Preferably, in one embodiment of the present invention, the comprehensive response time of each node is obtained by combining the similar response time difference and the interference response time difference, and several response isochrones are constructed. The specific method includes:

[0073] For any node of any type of material, the degree of similar influence of the node is used as the weight of the similar response time difference, and the difference obtained by subtracting the degree of similar influence of the node from 1 is used as the weight of the disturbance response time difference. The similar response time difference and the disturbance response time difference of the node are weighted and summed, and the result is used as the comprehensive response time of the node.

[0074] Furthermore, the maximum and minimum values ​​of the comprehensive response time of all nodes for all types of materials are obtained. The difference between the maximum and minimum values ​​is divided by 10, and the result is used as the classification threshold. Referring to the contour line construction method, isochrones are constructed for the comprehensive response time of each node, and these isochrones are used as response isochrones. If the absolute value of the difference between the comprehensive response times of each node on the same response isochrone does not exceed the classification threshold, then several response isochrones are constructed, and adjacent nodes on the same response isochrone are connected by straight lines.

[0075] Thus, isochronous response curves were constructed for the substrate wiring diagram, reflecting the distribution of nodes with the same degree of thermal conductivity response in various materials of the substrate.

[0076] Step S004: Based on the distribution of pixels and response isochrones in the DIC view, obtain the heat conduction distribution value of each pixel, then determine several analysis image blocks and obtain their feature vectors through KSVD decomposition, train the substrate warping prediction model, and thus predict substrate warping.

[0077] It should be noted that the location of the response isochrones represents a stable region of thermal conduction response characteristics that can be referenced. Therefore, it is expected that during the process of obtaining the feature vector through KSVD decomposition, the distribution of isochrones in the reference image can be used to extract feature vectors that express thermal conduction information, so that the feature vectors used in the subsequent prediction model training process are more representative of thermal conduction.

[0078] Specifically, for any pixel in the DIC view at any sampling time, the Euclidean distance between the pixel and each response isochronous line is obtained (the response isochronous lines are distributed in the substrate wiring diagram and the DIC view, and the corresponding positions of the pixels are the same). The response isochronous line corresponding to the minimum Euclidean distance is taken as the reference isochronous line for the pixel. The average of the comprehensive response times of all nodes on the reference isochronous line is obtained as the response time of the reference isochronous line. The product of the inversely proportional normalized result of the minimum Euclidean distance and the response time of the reference isochronous line is taken as the heat conduction distribution value of the pixel. The heat conduction distribution values ​​of each pixel other than those on the response isochronous line are obtained according to the above method. The response time of each response isochronous line is taken as the heat conduction distribution value of the pixels on it, thus forming a heat conduction distribution map.

[0079] Furthermore, for any DIC view at any sampling time, the DIC view is divided into several image blocks consisting of 16*16 pixels, i.e., each image block is 16*16 in size and does not overlap. The mean value of the heat conduction distribution of all pixels in any image block is obtained as the heat conduction comprehensive factor of the image block. All image blocks are arranged in descending order of heat conduction comprehensive factor, and a preset number of screenings is set. In this embodiment, the number of screenings is described as 64. The image blocks of the first number of screenings are used as several analysis image blocks and as the initial dictionary atoms in the convergence process of KSVD decomposition. The image blocks are sparsely encoded, and the distance is updated according to KSVD. This process is repeated iteratively until convergence. Then, the image blocks are encoded using the dictionary trained by KSVD, thereby obtaining several feature vectors of the DIC view at that sampling time. KSVD decomposition is a prior art and will not be described in detail in this embodiment.

[0080] Furthermore, in this embodiment, models such as RNN, CNN, and random forest are used for training. The training dataset consists of several feature vectors of the DIC view at each sampling time, which are input into the model to output the degree of warpage at each position of the substrate under each DIC view, thereby obtaining several trained warpage prediction models. The warpage prediction models are evaluated using model evaluation metrics, such as root mean square error and absolute error, to obtain the prediction performance of each warpage prediction model. The output degree of warpage is compared with the degree of warpage at each position marked manually for evaluation. The warpage prediction model with the highest prediction performance is selected as the substrate warpage prediction model. Then, during the temperature loading process, the DIC view at each time is input into the substrate warpage prediction model to output the predicted degree of warpage at each position of the substrate, thus realizing substrate warpage prediction.

[0081] It should be noted that this embodiment adopts... The model is used to represent the inverse proportional relationship and for normalization processing. As input to the model, It is an exponential function with the natural constant as the base. Implementers can set an inverse proportional function and a normalization function according to the actual situation.

[0082] This concludes the embodiment.

[0083] Another embodiment of the present invention provides a substrate warpage prediction system based on machine learning. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the above-described method steps S001 to S004.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning based substrate warpage prediction method, characterized by, The method includes the following steps: Obtain the temperature of the substrate at several moments during the temperature loading process; The temperature change trend at each temperature change stage during the temperature loading process is analyzed to obtain several sampling times at each temperature change stage, and the DIC view at each sampling time is obtained through the DIC measurement unit; based on the distribution of different types of materials in the substrate wiring diagram, several nodes of each type of material are obtained through corner detection; the similarity relationship between the pixel values ​​at corresponding positions in the DIC view of different nodes of the same type of material with changes in sampling time is analyzed to obtain the response time difference of each node. Based on the distribution of nodes in different types of materials, and combined with the changes in pixel values ​​at corresponding positions in the DIC view of nodes changing with sampling time, the interference response time difference of each node is quantified; based on the distribution of nodes in the same type of materials and their similar response time differences, the average response time of each type of material is obtained, and then the similar influence degree of each node in each type of material is obtained; combined with the similar response time difference and the interference response time difference, the comprehensive response time of each node is obtained, and several response isochrones are constructed. Based on the distribution of pixels and response isochrones in the DIC view, the thermal conductivity distribution value of each pixel is obtained, and then several analysis image blocks are determined and their feature vectors are obtained through KSVD decomposition. The substrate warpage prediction model is trained to predict substrate warpage. The specific method for obtaining the time difference of similar responses of each node is as follows: For any node of any material, obtain the pixel value of the corresponding position of the node in the DIC view at each sampling time, construct a coordinate system with the sampling time as the horizontal axis and the pixel value in the DIC view as the vertical axis, map the pixel value of the corresponding position of the node in the DIC view at each sampling time to the coordinate system, obtain several data points, and connect them to form the change curve of the node. Among several nodes of this type of material, the Euclidean distance between the node and all other nodes is obtained. The node corresponding to the minimum Euclidean distance is taken as the matching node of the node. The change curve of the node and the change curve of its matching node are matched by DTW to obtain several matching time pairs. The difference between the sampling time of the node and the sampling time of its matching node in any matching time pair is taken as the response time difference of the matching time pair. The average of the response time differences of all matching time pairs is taken as the same type of response time difference of the node. The interference response time difference of each node is obtained by the following method: For any node of any material class, and all nodes of any other material class, obtain the Euclidean distance between each node and the node, and take the node with the minimum Euclidean distance as the reference node of the node in the other material class. Obtain several matching time pairs between the change curve of the node and the change curve of the reference node, and obtain the response time difference of each matching time pair. Use the average of the response time differences of all matching time pairs between the change curve of the node and the change curve of the reference node as the response time difference of the node under the influence of other types of materials. The average of the response time differences of the node to all other types of materials except for this type is taken as the disturbance response time difference of the node. The specific methods for obtaining the degree of similar influence at each node of various materials include: Obtain the absolute value of the difference between the response time difference of each node in any type of material and the average response time of that type of material. Take the ratio of the absolute value of the difference corresponding to any node to the maximum value among all absolute values ​​of the difference as the degree of influence of that node in that type of material. The specific methods for obtaining the comprehensive response time of each node and constructing several response isochrones are as follows: For any node of any type of material, the degree of similar influence of the node is used as the weight of the similar response time difference, and the difference obtained by subtracting the degree of similar influence of the node from 1 is used as the weight of the disturbance response time difference. The similar response time difference and the disturbance response time difference of the node are weighted and summed, and the result is used as the comprehensive response time of the node. The maximum and minimum values ​​of the comprehensive response time of all nodes of all materials are obtained. The difference between the maximum and minimum values ​​is divided by 10 and used as the classification threshold. Based on the contour line construction process, isochrones are constructed for the comprehensive response time of each node, and these isochrones are used as response isochrones. If the absolute value of the difference between the comprehensive response times of each node on the same response isochrone does not exceed the classification threshold, then several response isochrones are constructed.

2. The method of claim 1, wherein: The specific method for obtaining several sampling times for each temperature change stage is as follows: For any temperature change phase, the slope at each moment is calculated based on the temperature difference between adjacent moments; the absolute value of the average slope of all moments in any temperature change phase is taken as the change coefficient of that temperature change phase; the sum of 1 and the change coefficient is obtained; a unit interval reference value is preset; the product of the sum and the unit interval reference value and the result of rounding down is taken as the unit interval of that temperature change phase. The first moment in the temperature change phase is taken as a sampling moment, and a sampling moment is obtained at every unit interval until the end of the temperature change phase, thus obtaining a number of sampling moments in the temperature change phase.

3. The substrate warpage prediction method based on machine learning according to claim 1, characterized in that, The specific methods for obtaining several nodes of various materials are as follows: Obtain the substrate wiring diagram. The same materials in the substrate wiring diagram are marked with the same color. Perform corner detection on the area corresponding to any type of material to obtain several corner points. Use the obtained corner points as several nodes of that type of material.

4. The substrate warpage prediction method based on machine learning according to claim 3, characterized in that, The specific methods for obtaining the average response time of various materials are as follows: For any node of any material class, if the node and its matching node are in the same region of the same material class, obtain the number of nodes in that region, and use the ratio of the number of nodes in that region to the total number of nodes in all material classes as the response factor of that node; If the node and its matching node are in different regions of the same material, obtain the connection between the node and its matching node, and obtain the regions of other material types that the connection passes through. If the connection passes through regions of other material types but does not pass through nodes of other material types, the corresponding region count is 1. If the connection passes through regions of other material types and passes through nodes of other material types in the region, the number of nodes of other material types in the region is used as the count of the corresponding region. The ratio of the sum of the counts of all other material types that the connection passes through to the total number of nodes of all material types is used as the interference factor of the node. The difference obtained by subtracting the interference factor from 1 is used as the response factor of the node. The response factors of all nodes in this type of material are weighted and normalized to obtain the response weight of each node. The response time differences of the same type of nodes are weighted and summed based on the response weights, and the result is used as the average response time of this type of material.

5. The substrate warpage prediction method based on machine learning according to claim 1, characterized in that, The specific method for obtaining the heat conduction distribution value of each pixel and then determining several image blocks for analysis includes: For any pixel in the DIC view at any sampling time, obtain the Euclidean distance between the pixel and each response isochrone. Take the response isochrone corresponding to the minimum Euclidean distance as the reference isochrone for the pixel. Obtain the average of the comprehensive response times of all nodes on the reference isochrone as the response time of the reference isochrone. Multiply the inversely proportional normalized result of the minimum Euclidean distance by the response time of the reference isochrone as the heat conduction distribution value of the pixel. Obtain the heat conduction distribution value of each pixel except the pixel on the response isochron. Use the response time of each response isochron as the heat conduction distribution value of the pixel on it to form a heat conduction distribution map. For any sampling time, the DIC view is divided into several image blocks. The mean value of the heat conduction distribution of all pixels in any image block is obtained as the heat conduction comprehensive factor of that image block. All image blocks are arranged in descending order of thermal conductivity comprehensive factor. A preset number of images are selected, and the first number of images selected are used as a number of images for analysis.

6. A machine learning-based substrate warpage prediction system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the substrate warpage prediction method based on machine learning as described in any one of claims 1-5.

Images