A method for fast simulation of drop impact response of an integrated circuit package

Abstract

The application provides a kind of integrated circuit package drop impact response fast simulation method, comprising: establishing the integrated circuit package drop impact simulation model considering plastic deformation and contact friction effect;Obtain drop impact samples under different spatial positions and working conditions by parameterized scanning, construct data set;Pretreatment is carried out to sample data, key area node data is extracted, and consistent expression of sample node and network input node is realized through coordinate conversion;The prediction variable is normalized, and deep features are extracted based on spatial parameters and node coordinates to construct input feature set;Multi-task deep learning network based on cross-scale feature splicing and multi-gate hybrid expert mechanism is constructed, and data set is trained;The network model trained is used to quickly predict the integrated circuit package drop impact response, and the field distribution of equivalent plastic deformation, von mises stress and deformation energy density is reconstructed.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit packaging drop simulation technology, and more specifically, to a rapid simulation method for the drop impact response of integrated circuit packaging. Background Technology

[0002] With the rapid development of integrated circuit packaging technology, the dynamic response analysis of packaging structures under different operating environments has become a key issue in electronic system design. In particular, drop impact, as an occasional extreme load, has a significant impact on the mechanical behavior and reliability of integrated circuit packaging structures. Current technologies for drop impact response analysis of integrated circuit packaging structures mainly employ experimental testing methods and numerical simulation methods to obtain the response characteristics of the packaging structure under drop impact loads.

[0003] However, existing technologies still have the following problems: First, traditional experimental testing is costly and time-consuming, and numerical simulation is computationally inefficient when dealing with short-term, high-energy, and highly nonlinear problems such as drop impacts, making it difficult to quickly analyze various collision scenarios; Second, the data samples used by existing deep learning methods cover a limited range of impact postures, making it difficult to comprehensively represent actual drop conditions; Third, existing network models have limited ability to extract and fuse input features, making it difficult to fully represent node spatial information and response patterns. In multi-output prediction scenarios, the modeling of the correlation between different output tasks is insufficient, thus limiting the predictive performance and applicability of the model. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a fast simulation method for the drop impact response of integrated circuit packages. This method combines deep learning technology and constructs a multi-task deep learning network to quickly predict the mechanical response of integrated circuit packages under drop impact conditions, which significantly improves simulation efficiency and ensures calculation accuracy.

[0005] The technical solution of the present invention is as follows:

[0006] A rapid simulation method for drop shock response of integrated circuit packages includes the following steps:

[0007] Establish a drop impact simulation model for integrated circuit packaging that considers plastic deformation and contact friction effects;

[0008] A dataset was constructed by acquiring drop impact samples under different spatial locations and working conditions through parametric scanning.

[0009] The sample data is preprocessed to extract key region node data, and coordinate transformation is used to achieve a consistent representation between the sample nodes and the network input nodes.

[0010] The predicted variables are normalized, and deep features are extracted based on spatial parameters and node coordinates to construct an input feature set;

[0011] A multi-task deep learning network based on cross-scale feature concatenation and a multi-gated hybrid expert mechanism is constructed and trained on the dataset.

[0012] The trained network model is used to quickly predict the drop impact response of integrated circuit packages and reconstruct the field distribution of equivalent plastic deformation, von Mises stress and deformation energy density.

[0013] Preferably, the integrated circuit package drop impact simulation model is a three-dimensional finite element model including the package structure, solder joint area, chip area and substrate area, wherein the part used for training and output only includes the solder joint area. The simulation model considers the material plastic deformation effect and contact friction effect, ignores air resistance, and the package structure can change the drop height and impact posture during the simulation process.

[0014] Preferably, the step of acquiring drop impact samples under different spatial locations and working conditions through parametric scanning and constructing a dataset specifically includes:

[0015] Step 1: Establish a parametric simulation model of drop impact on integrated circuit packaging;

[0016] Step 2: Set the range of parameter variations for drop height, spatial location, and impact attitude;

[0017] Step 3: Perform batch simulation calculations for each parameter combination to obtain multiple drop impact samples;

[0018] Step 4: Extract the equivalent plastic deformation, von Mises stress and deformation energy density distribution results for each sample at the corresponding time point to construct a drop impact dataset.

[0019] Preferably, the steps of preprocessing the sample data, extracting key region node data, and achieving a consistent representation of the sample nodes and network input nodes through coordinate transformation specifically include:

[0020] Step 1: Extract the node coordinates and corresponding physical quantity data of the weld area from the drop impact simulation results;

[0021] Step 2: Anchor a set of reference node coordinates, establish coordinate transformation relationships with other samples, and achieve alignment of node data from different samples;

[0022] Step 3: Construct network training samples based on the aligned node data.

[0023] Preferably, in the step of normalizing the predicted variables and extracting deep features based on spatial parameters and node coordinates to construct an input feature set, the predicted variables include equivalent plastic deformation, von Mises stress, and deformation energy density; the input features include node coordinate features, node position features relative to the whole, periodic features, Fourier features, and deep features formed by the combination of spatial parameters and derived values.

[0024] Preferably, in the step of constructing a multi-task deep learning network based on cross-scale feature concatenation and a multi-gated hybrid expert mechanism, and training the dataset, the multi-task deep learning network includes an input layer, a feature extraction layer, a cross-scale feature fusion layer, an expert module, a gating module, a task fusion layer, and a multi-task output head; the cross-scale fusion layer is used to enhance the network's ability to extract multi-scale spatial features and alleviate the gradient vanishing problem during deep network training; the expert module is used to learn shared and differential features of different prediction tasks; the gating module is used to dynamically allocate the output weights of each expert module according to different tasks; and the multi-task output head is used to output the prediction results of equivalent plastic deformation, von Mises stress, and deformation energy density respectively.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. By preprocessing samples and transforming coordinates, the input node data was anchored, which solved the problem of inconsistent distribution between sample nodes and network input nodes, and constructed a drop impact dataset covering all scenarios.

[0027] 2. Based on the data characteristics, spatial features, periodic features, absolute and relative node position features, and Fourier features were extracted, and a cross-scale feature fusion layer was designed to effectively improve the accuracy of feature extraction and the learning ability of the network.

[0028] 3. In the network design, a multi-gated hybrid expert mechanism and a sharing mechanism are combined. By using a task fusion layer and gating to dynamically allocate the output weights of expert modules, the adaptability of the model to multi-physical fields and multi-task collaborative learning is enhanced. Attached Figure Description

[0029] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0030] Figure 1 Flowchart of a rapid simulation method for drop impact response of integrated circuit packages;

[0031] Figure 2 A simulation flowchart for drop impact samples of integrated circuit packages;

[0032] Figure 3This is a schematic diagram of a multi-task deep learning network structure;

[0033] Figure 4 This is a schematic diagram of an integrated circuit packaging structure according to a specific embodiment;

[0034] Figure 5 A schematic diagram of the package drop impact posture in a specific embodiment;

[0035] Figure 6 This is a schematic diagram of the von Mises stress reconstruction results;

[0036] Figure 7 This is a schematic diagram of the equivalent plastic deformation reconstruction result;

[0037] Figure 8 This is a schematic diagram of the deformation energy density reconstruction results. Detailed Implementation

[0038] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0039] Specifically, this invention provides a fast simulation method for the drop impact response of integrated circuit packages, such as... Figure 1 As shown, the method includes the following steps:

[0040] S1: Establish a drop impact simulation model for integrated circuit packaging that considers plastic deformation and contact friction effects;

[0041] The integrated circuit package drop impact simulation model is a three-dimensional finite element model that includes the package structure, solder joint area, chip area and substrate area. The part used for training and output only includes the solder joint area. The simulation model considers the plastic deformation effect and contact friction effect of the material, and ignores air resistance. The package structure can change the drop height and impact posture during the simulation.

[0042] Specifically, in step S1, the electronic packaging structure falls from a height h. Assuming negligible air resistance, its velocity and time upon contact with the ground are obtained from the law of conservation of energy.

[0043]

[0044] Where g is the acceleration due to gravity, t f The initial impact velocity v0 is determined by the free fall phase, which in turn determines the subsequent impact intensity.

[0045] After the impact contact occurs, the system enters a short-time strongly nonlinear response stage, and its dynamic equilibrium equation is given by the conservation of momentum (or the principle of virtual work):

[0046]

[0047] Where ρ is density, u is displacement vector, σ is Cauchy stress tensor, and b is body force (including gravity). At the contact surface Γ c Above, contact constraints and contact forces need to be applied:

[0048]

[0049] Where g n For the normal gap, p n The normal contact pressure is t; the tangential force t can be described by Coulomb friction.

[0050]

[0051] Where μ is the coefficient of friction. Von Mises stress considering plastic deformation. Compliance with J2 yield criterion:

[0052]

[0053] Where S is the deviatoric stress tensor, k is the stress direction label, and σ y (k) Takes the isotropic hardening mode. The other two output responses are equivalent to plastic deformation. and deformation energy density w d (t) are respectively:

[0054]

[0055]

[0056] Where σ is the stress tensor, It is the instantaneous value of plastic deformation.

[0057] S2: Obtain drop impact samples under different spatial locations and working conditions through parametric scanning to construct a dataset;

[0058] Step S2 specifically includes:

[0059] Step 1: Establish a parametric simulation model of drop impact on integrated circuit packaging;

[0060] Step 2: Set the range of parameter variations for drop height, spatial location, and impact attitude;

[0061] Step 3: Perform batch simulation calculations for each parameter combination to obtain multiple drop impact samples;

[0062] Step 4: Extract the equivalent plastic deformation, von Mises stress and deformation energy density distribution results for each sample at the corresponding time point to construct a drop impact dataset.

[0063] Specifically, such as Figure 2 As shown, a geometric model of the encapsulation structure is first established in COMSOL, and material parameters such as elastoplastic parameters, density, and damping are input. Then, the physical field and its governing equations are set, and elastoplastic constitutive relations and contact constraints are applied. Afterward, the model is meshed, and the weld area and interface are locally refined to improve the simulation accuracy of key areas. After the model is built, the angle α between the encapsulation structure and the xOy plane, the angle β between the encapsulation structure and the xOz plane, and the height h of the lowest point of the encapsulation from the rigid ground are used as parameterized variables for scanning, thereby generating a batch of multi-condition drop impact samples. Time-integrated transient analysis is used in the solution process, and the corresponding time step and convergence strategy are set. The result output time is selected as 20 μs after the collision. Finally, the von Mises stress, strain energy density, and equivalent plastic deformation at each node of the encapsulation structure are exported to form drop impact simulation sample data.

[0064] S3: Preprocess the sample data, extract key region node data, and achieve a consistent representation of the sample nodes and network input nodes through coordinate transformation;

[0065] Specifically, the steps for coordinate transformation and alignment of sample nodes in step S3 are as follows:

[0066] Since the area outside the solder ball corresponds to linear material, COMSOL marks the equivalent plastic deformation of this area as NaN when outputting the results. Based on this, it identifies and removes node data in the non-solder ball area, and only retains the prediction variable data corresponding to the solder ball area.

[0067] To ensure that the number of selected nodes and their relative positions remain consistent under different attitude and spatial position parameters, the coordinates (x0, y0, z0) of any node in the parameter combination (0,0,0) are selected as the reference coordinates. For example... Figure 5 As shown, let the horizontal axis be a plane passing through the origin and parallel to the xOy plane, and let the central axis be a plane passing through the origin and parallel to the xOz plane. When the parameter combination changes to (α, β, h), the corresponding node coordinates change accordingly. To achieve a consistent expression of node coordinates for different samples, this invention employs the following coordinate transformation method:

[0068] The first step is to rotate the original node coordinates x = (x0, y0, z0) around the horizontal axis by an angle α, thereby transforming them into transition coordinates x. h = (x h , y h , z h The calculation formula is as follows:

[0069]

[0070] in λ1 = arcsin(z0 / L0), and sign(·) denotes the sign function.

[0071] The second step is to rotate the plane around the central axis by an angle β, and raise it by a height h in the vertical direction, thus adjusting the transition coordinate x. h Convert to final coordinates x f The calculation formula is as follows:

[0072]

[0073] in λ2 = arccos(z) h / L h ).

[0074] Through the above two-step coordinate transformation, nodes with different parameter combinations can be mapped to a unified coordinate system, thereby ensuring the consistency of node position representation in subsequent network input samples.

[0075] S4: Normalize the predicted variables and extract deep features based on spatial parameters and node coordinates to construct the input feature set;

[0076] Specifically, in step S4, the predicted variables include equivalent plastic deformation, von Mises stress, and deformation energy density. Since the three predicted variables have significantly different orders of magnitude, they are normalized to improve the numerical stability during network training. The normalization formula is as follows:

[0077]

[0078] Where Y represents the value of a certain predictor, Y max and Y min These are its maximum and minimum values, respectively.

[0079] In step S4, the steps for constructing the input feature set are as follows:

[0080] The first step is to extract the three-dimensional coordinates (x, y, z) of each node as the basic spatial features, and extract the attitude parameters α, β and the height parameter h corresponding to the current fall condition as the condition features.

[0081] The second step is to calculate the distance feature l from node i to the origin, which is expressed as:

[0082]

[0083] Where (x) i , y i , z i ) represents the position coordinates of node i.

[0084] The third step is to normalize the node index to construct the relative position feature pos1 of node i. i Furthermore, a periodic feature pos2 is constructed through sine and cosine transformations. i and pos3 i Its expression is:

[0085]

[0086]

[0087] The fourth step is to construct the interaction features and derived features between attitude parameters, including αβ, α 2 β 2 , inv α and inv β, where:

[0088]

[0089] Where ε is the disturbance constant to prevent the denominator from being zero;

[0090] The fifth step involves randomly projecting the node coordinates to generate Fourier features, thereby enhancing the network's ability to express nonlinear spatial relationships. The expression for this feature is:

[0091]

[0092] Where coords represents the node coordinate matrix and W represents the random projection matrix;

[0093] Step 6: Expand the Fourier features according to the sample conditions to obtain the Fourier input features corresponding to the task:

[0094]

[0095] Where r represents the Fourier feature order.

[0096] Step 7: Concatenate the basic spatial features, working condition features, distance features, periodic features, parameter interaction features, and Fourier features to form the final network input feature matrix X. i Its expression is:

[0097]

[0098] Through the above steps, absolute position features, relative position features, periodic features, parameter interaction features, and Fourier features can be extracted from the node spatial distribution and drop condition parameters, thereby constructing input features for subsequent network training and response prediction.

[0099] S5: Construct a multi-task deep learning network based on cross-scale feature concatenation and a multi-gated hybrid expert mechanism, and train it on the dataset;

[0100] The multi-task deep learning network includes an input layer, a feature extraction layer, a cross-scale feature fusion layer, an expert module, a gating module, a task fusion layer, and a multi-task output head. The cross-scale fusion layer is used to enhance the network's ability to extract multi-scale spatial features and alleviate the gradient vanishing problem during deep network training. The expert module is used to learn shared and differential features for different prediction tasks. The gating module is used to dynamically allocate the output weights of each expert module according to different tasks. The multi-task output head is used to output the prediction results of equivalent plastic deformation, von Mises stress, and deformation energy density, respectively.

[0101] Specifically, such as Figure 3 As shown, the construction steps of the multi-task deep learning network based on cross-scale feature concatenation and multi-gated hybrid expert mechanism are as follows:

[0102] The first step is to input the input features obtained in step S4 into the expert module to establish the input layer of the deep learning network;

[0103] The second step is to introduce the Inception-Res module into the expert module, which extracts local response features at different scales by combining multi-scale parallel convolution with residual connections, so as to enhance the network's ability to express complex spatial features.

[0104] The third step is to set up multiple expert sub-networks and realize the sharing of underlying features between different prediction tasks through an expert sharing mechanism, so that the network can retain the commonalities of tasks while taking into account the differences of tasks.

[0105] The fourth step is to set up a feature fusion layer after the expert shared output, and to perform cross-scale feature splicing and fusion on features extracted by different scales and different experts, so as to improve the network's ability to represent multi-level spatial information.

[0106] The fifth step is to set up a gating module to weight the contribution of each expert subnetwork output according to different tasks, thereby achieving dynamic selection of expert outputs. The expression of the weighting function is:

[0107]

[0108] in, It is the weight value of the task. It is the loss value for each task.

[0109] The sixth step is to set a task-specific header after the gating module, process the fusion features of different tasks separately, and output the prediction results of equivalent plastic deformation, von Mises stress and deformation energy density.

[0110] Step 7: Train the multi-task deep learning network based on the training samples. A relative error loss function is used for equivalent plastic deformation, and a Huber loss function is used for von Mises stress and deformation energy density to improve the prediction accuracy and training stability of each task. The definitions of the relative error loss function and the Huber loss function are as follows:

[0111]

[0112]

[0113] in Y is the actual value, Y is the predicted value, and m is a small constant used to prevent the denominator from being zero. It is a threshold, usually set to 1.0 or other hyperparameters.

[0114] The eighth step is to perform a weighted summation of the losses from each task to form a total loss function, and then use an optimization algorithm to update the network parameters until the preset training termination conditions are met.

[0115] S6: Utilize the trained network model to quickly predict the drop impact response of integrated circuit packages and reconstruct the field distribution of equivalent plastic deformation, von Mises stress, and deformation energy density.

[0116] Specifically, the trained network model is used to quickly predict the drop impact response of integrated circuit packages and reconstruct the equivalent plastic deformation. von Mises stress and deformation energy density The field distribution.

[0117] Based on the simulation method of the present invention described above, a specific embodiment is calculated.

[0118] The following is in conjunction with the accompanying drawings, Figure 4 The implementation of the technical solution will be further described in detail using the integrated circuit packaging structure shown as an example. Figure 4 This embodiment illustrates a flip-chip ball grid array (BGA) package structure, including an epoxy molding compound, a die, a substrate, a package substrate, a solder ball array, and a printed circuit board. The solder balls in the package have a radius of 0.5 mm and are cut along a plane located 3 / 5 of the radius above and below the center of each solder ball. The spacing between the solder balls is 2 mm. To more clearly describe the spatial position and attitude parameters of the package in response to a drop impact, combined with... Figure 5 As shown, different samples are characterized using a combination of parameters (α, β, h), where α is the tilt angle of the horizontal plane (xOy plane), β is the tilt angle of the axial plane (xOz plane), and h is the height of the lowest point of the package from the rigid ground.

[0119] In this embodiment, by adjusting the tilt angles α and β and the height parameter h, 1200 sets of drop impact simulation samples were generated using the commercial software COMSOL and used as the training dataset. The multi-task deep learning network based on cross-scale feature stitching and a multi-gated hybrid expert mechanism proposed in this invention achieves the following accuracies in predicting equivalent plastic deformation, von Mises stress, and deformation energy density: the maximum absolute error is 3.15 × 10⁻⁶. -3 2.26 MPa and 1.74 × 10 4 J / m 3 The average relative errors were 1.7%, 0.388%, and 4.48%, respectively, indicating that the overall prediction accuracy was better than that of traditional neural networks. For example... Figure 6 , Figure 7 , Figure 8 As shown, the prediction results of the method of the present invention are highly consistent with the simulation results of COMSOL.

[0120] In terms of computational efficiency, the multi-task deep learning network based on cross-scale feature splicing and multi-gated hybrid expert mechanism requires approximately 360 seconds to complete training of 1200 sets of samples, with an average data processing and inference time of approximately 0.05 seconds per case. In contrast, the traditional finite element method requires approximately 126,000 seconds to complete simulation of the same scale of samples, and the method of this invention can achieve an acceleration effect of approximately 350 times.

[0121] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A fast simulation method for drop impact response of integrated circuit packages, characterized in that, The method includes the following steps: Establish a drop impact simulation model for integrated circuit packaging that considers plastic deformation and contact friction effects; A dataset was constructed by acquiring drop impact samples under different spatial locations and working conditions through parametric scanning. The sample data is preprocessed to extract key region node data, and coordinate transformation is used to achieve a consistent representation between the sample nodes and the network input nodes. The predicted variables are normalized, and deep features are extracted based on spatial parameters and node coordinates to construct an input feature set; A multi-task deep learning network based on cross-scale feature concatenation and a multi-gated hybrid expert mechanism is constructed and trained on the dataset. The trained network model is used to quickly predict the drop impact response of integrated circuit packages and reconstruct the field distribution of equivalent plastic deformation, von Mises stress and deformation energy density.

2. The rapid simulation method for drop impact response of integrated circuit packaging according to claim 1, characterized in that, The integrated circuit package drop impact simulation model is a three-dimensional finite element model that includes the package structure, solder joint area, chip area and substrate area. The part used for training and output only includes the solder joint area. The simulation model considers the plastic deformation effect and contact friction effect of the material, and ignores air resistance. The package structure can change the drop height and impact posture during the simulation.

3. The rapid simulation method for drop impact response of integrated circuit packaging according to claim 1, characterized in that, The step of acquiring drop impact samples under different spatial locations and working conditions through parametric scanning and constructing a dataset specifically includes: Step 1: Establish a parametric simulation model of drop impact on integrated circuit packaging; Step 2: Set the range of parameter variations for drop height, spatial location, and impact attitude; Step 3: Perform batch simulation calculations for each parameter combination to obtain multiple drop impact samples; Step 4: Extract the equivalent plastic deformation, von Mises stress and deformation energy density distribution results for each sample at the corresponding time point to construct a drop impact dataset.

4. The rapid simulation method for drop impact response of integrated circuit packaging according to claim 1, characterized in that, The steps of preprocessing the sample data, extracting key region node data, and achieving a consistent representation of the sample nodes and network input nodes through coordinate transformation specifically include: Step 1: Extract the node coordinates and corresponding physical quantity data of the weld area from the drop impact simulation results; Step 2: Anchor a set of reference node coordinates, establish coordinate transformation relationships with other samples, and achieve alignment of node data from different samples; Step 3: Construct network training samples based on the aligned node data.

5. The rapid simulation method for drop impact response of integrated circuit packaging according to claim 1, characterized in that, In the step of normalizing the predicted variables and extracting deep features based on spatial parameters and node coordinates to construct an input feature set, the predicted variables include equivalent plastic deformation, von Mises stress, and deformation energy density; the input features include node coordinate features, node position features relative to the whole, periodic features, Fourier features, and deep features formed by the combination of spatial parameters and derived values.

6. The rapid simulation method for drop impact response of integrated circuit packaging according to claim 1, characterized in that, The step of constructing a multi-task deep learning network based on cross-scale feature concatenation and a multi-gated hybrid expert mechanism, and training it on the dataset, includes an input layer, a feature extraction layer, a cross-scale feature fusion layer, an expert module, a gating module, a task fusion layer, and a multi-task output head. The cross-scale fusion layer is used to enhance the network's ability to extract multi-scale spatial features and alleviate the gradient vanishing problem during deep network training. The expert module is used to learn shared and differential features for different prediction tasks. The gating module is used to dynamically allocate the output weights of each expert module according to different tasks; the multi-task output head is used to output the prediction results of equivalent plastic deformation, von Mises stress and deformation energy density respectively.

Images