A method and system for equivalent of rdl circuit composite based on convnext model
By using a neural network method based on the ConvNeXt model, the problem of low simulation accuracy of composite materials in large-scale RDL circuits was solved, achieving high-precision warpage prediction and improved packaging reliability.
Patent Information
- Application Number
- CN202510170209.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-17
- Publication Date
- 2025-11-21
- Estimated Expiration
- 2045-02-17
AI Technical Summary
Existing technologies struggle to perform high-fidelity simulations of composite materials for large-scale RDL circuits, especially regarding the impact of copper trace shape distribution on stress and warpage, which results in significant errors that affect wafer warpage and packaging reliability.
A neural network approach based on the ConvNeXt model is adopted. By constructing a database and a deep learning multimodal parameter analysis model, the neural network is trained to predict the thermodynamic parameters of the RDL circuit with high accuracy and to establish an equivalent material model.
High-fidelity simulation of large-scale RDL models was achieved, reducing warpage prediction errors and improving wafer production yield and packaging reliability.
Smart Images
Figure CN120164552B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of semiconductor manufacturing thermodynamic simulation, and in particular to an RDL circuit composite material equivalent method and system based on a ConvNeXt model. BACKGROUND
[0002] With the development of artificial intelligence, big data, cloud computing and other technologies, chips need to have higher operation speed, higher information transmission density, and relatively lower cost and power consumption. With the development of advanced chip processes tending to physical limits, Moore's Law is gradually failing, and the chip field has entered the post-Moore era.
[0003] To further improve the operation speed and data transmission rate of chips, advanced packaging technology has become a new solution. As an important component of 2.5D packaging, the silicon interposer plays a crucial role in improving the data transmission speed and interconnection density between chips. RDL (ReDistribution Layer) is a key technology in silicon interposers. Its main function is to redistribute the electrical connection points (I / O pad) on the chip, and by adding one or more layers of metal wiring on the chip surface, these connection points are rearranged to a more suitable position for packaging and interconnection. Due to problems such as mismatch of RDL material thermal expansion coefficient (Coefficient of Thermal Expansion, CTE) and uneven microstructure topology distribution, the wafer of the silicon interposer will experience stress and warping after heating and cooling during the manufacturing process. Excessive wafer warping can easily cause wafer position deviation, which has a negative impact on the alignment of photolithography during processing. In addition, excessive warping can also cause solder ball displacement, wafer breakage and other problems, which seriously affect the yield of the product. Therefore, by simulating and predicting wafer warping, in-depth exploration of the causes of warping in the redistribution process, and optimization of warping through design parameters, it is of great significance to improve the production yield of silicon interposer wafers, reduce production costs, and promote the development of China's semiconductor advanced packaging industry.
[0004] At the macro scale, one of the biggest technical challenges in large-scale RDL simulation research is the difficulty of completely distortionless modeling and simulation of its complex copper traces. General equivalent methods have not been able to capture the impact of copper trace shape distribution on stress and warping due to excessive simplification of the model, resulting in too large a result error. SUMMARY
[0005] One of the purposes of the present application is to provide a ConvNeXt model-based RDL circuit composite material equivalent method, which adopts a machine learning method, constructs a database according to simulation data, establishes a deep learning multi-modal parameter analysis model, trains a neural network model capable of high-precision prediction of equivalent model thermodynamic parameters according to RDL images, and thus realizes high-fidelity simulation of large-scale RDL models.
[0006] The second purpose of the present application is to provide a ConvNeXt model-based RDL circuit composite material equivalent system.
[0007] The scheme adopted by one of the purposes of the present application is: a ConvNeXt model-based RDL circuit composite material equivalent method, comprising the following steps:
[0008] S1: collecting image data, modeling multiple RDL circuit images and calculating related thermodynamic parameters, and establishing a data set after enhancing the data;
[0009] S2: establishing and improving the neural network model of ConvNeXt;
[0010] S3: inputting the data set established in step S1 into the ConvNeXt neural network model obtained in step S2 for training to optimize the neural network model of ConvNeXt, and generating a weight file corresponding to each training;
[0011] S4: using the trained ConvNeXt weight file to predict the thermodynamic parameters of the RDL circuit;
[0012] S5: importing the thermodynamic parameters predicted by ConvNeXt into the parameters of the equivalent material to form an equivalent material that can be used for RDL thermodynamic simulation.
[0013] Preferably, step S1 comprises the following steps:
[0014] S1a: obtaining images of RDL layouts with various circuit topologies;
[0015] S1b: extracting circuit contour data points of the images and numbering them;
[0016] S1c: importing the contour data point file into finite element software for automated modeling, assigning metal wires or dielectric material properties to the contours according to the circuit layout, and constructing an RDL composite material model;
[0017] S1d: performing thermodynamic simulation on the RDL composite material model to obtain thermodynamic parameters, and calculating the Young's modulus, thermal expansion coefficient in X, Y and Z directions, and shear modulus in XY, YZ and XZ planes;
[0018] S1e: Data augmentation is performed on the data obtained by simulation calculation according to geometric symmetry, and a data set is established.
[0019] Preferably, in step S1d, the specific calculation method of Young's modulus is to select two faces perpendicular to the direction, fix one face, stretch the other face, obtain the stress and strain values on the stretched face, and calculate the Young's modulus in the direction through the Young's modulus formula; the specific calculation method of the thermal expansion coefficient is to fix one of the faces perpendicular to the direction, set an initial temperature, calculate the thermal strain of the other face after heating, and calculate the thermal expansion coefficient in the direction through the thermal expansion coefficient calculation formula; the specific calculation method of the shear modulus is to fix one of the faces perpendicular to the face, apply a pulling force parallel to the face on the corresponding face, calculate the average stress and strain of the face, and calculate the shear modulus in the direction through the shear modulus calculation formula.
[0020] Preferably, in step S1e, the specific method of data augmentation on the data obtained by simulation calculation according to geometric symmetry is to rotate a picture 180° clockwise, flip it up and down, and then rotate it 180° clockwise again to correspond to the same thermodynamic parameter value.
[0021] Preferably, step S2 includes the following steps:
[0022] S2a: A module for feature extraction of a ConvNeXt neural network model is constructed, and the output is normalized by LayerNorm2d normalization operation after convolution kernel;
[0023] S2b: The first CNBlock module is used to increase the nonlinearity and feature extraction capability of the neural network;
[0024] S2c: After the first down-sampling layer convolution, higher-level feature information is extracted;
[0025] S2d: The second CNBlock module is used to increase the nonlinearity and feature extraction capability of the neural network;
[0026] S2e: After the second down-sampling layer convolution, higher-level feature information is extracted;
[0027] S2f: The third CNBlock module is used to increase the nonlinearity and feature extraction capability of the neural network;
[0028] S2g: After the third down-sampling layer convolution, higher-level feature information is extracted;
[0029] S2h: The fourth CNBlock module is used to increase the nonlinearity and feature extraction capability of the neural network;
[0030] S2i: regression prediction after normalization layer and full connection layer processing, completing the establishment and improvement of the neural network model of ConvNeXt.
[0031] Preferably, the structures of the first CNBlock module, the second CNBlock module, the third CNBlock module and the fourth CNBlock module each comprise, from top to bottom, a convolution layer, a LayerNorm layer, a linear layer, a GELU activation layer and a linear layer, wherein the GELU activation function can be represented as: x*sigmoid(1.703x).
[0032] Preferably, the input channels and output channels of the first CNBlock module, the second CNBlock module, the third CNBlock module and the fourth CNBlock module increase in turn.
[0033] Preferably, the input channels and output channels of the first down-sampling layer, the second down-sampling layer and the third down-sampling layer increase in turn.
[0034] Preferably, step S5 comprises the following steps:
[0035] S5a: inputting the RDL image into the constructed ConvNeXt neural network model to predict the Young's modulus, the thermal expansion coefficient in the X, Y and Z directions, and the shear modulus of the XY, YZ and XZ planes;
[0036] S5b: inputting the predicted values of the Young's modulus and the thermal expansion coefficient in the X, Y and Z directions, and the shear modulus of the XY, YZ and XZ planes into the finite element software to form an RDL equivalent material for subsequent thermodynamic simulation.
[0037] The scheme adopted by the second purpose of the application is: an RDL circuit composite material equivalent system based on a ConvNeXt model, comprising a processor and a memory, the memory is used for storing program instructions, and the processor is used for calling the storage instructions in the memory to execute the RDL circuit composite material equivalent method based on the ConvNeXt model.
[0038] The application has the following advantages and beneficial effects:
[0039] The method and system of the application are intended to adopt a machine learning method, construct a database according to simulation data, establish a deep learning multi-modal parameter analysis model, train a neural network model capable of predicting the thermodynamic parameters of the equivalent model with high precision according to the RDL image, and thus realize high-fidelity simulation of a large-scale RDL model. BRIEF DESCRIPTION OF DRAWINGS
[0040] Figure 1A flowchart of the RDL circuit composite material equivalent method based on the ConvNeXt of the present application is provided.
[0041] Figure 2 A flowchart of the RDL circuit composite material equivalent method based on the ConvNeXt of the present application is provided.
[0042] Figure 3 A method schematic diagram for obtaining RDL circuit thermodynamic parameters based on finite element simulation is provided for the embodiment of the present application.
[0043] Figure 4 A schematic diagram of the data enhancement method based on geometric transformation is provided for the embodiment of the present application.
[0044] Figure 5 A flowchart of the ConvNeXt neural network model for predicting RDL circuit thermodynamic parameters based on the deep learning method is provided for the embodiment of the present application.
[0045] Figure 6 A schematic diagram of the main network structure of the ConvNeXt neural network model is provided for the embodiment of the present application.
[0046] Figure 7 A performance comparison diagram of the ConvNeXt neural network model and the traditional Voigt method is provided for the embodiment of the present application. DETAILED DESCRIPTION
[0047] The following examples are further illustrations of the present application, but the scope of the present application is not limited to the following examples.
[0048] High-fidelity thermodynamic simulation of RDL circuit is of great significance to high-precision prediction of wafer warping and improvement of reliability of 2.5D packaging. The present application is aimed at the problem of low prediction accuracy of traditional methods for RDL circuit composite material thermodynamic properties. Through finite element simulation, a ConvNeXt neural network model is built to realize high-precision prediction of RDL circuit composite material thermodynamic properties.
[0049] As shown in Figure 1 A RDL circuit composite material equivalent method based on the ConvNeXt model, comprising the following steps:
[0050] S1: Collect image data, model multiple RDL circuit images and calculate related thermodynamic parameters, and establish a data set after data enhancement;
[0051] S2: Establish and improve the neural network model of ConvNeXt;
[0052] S3: input the data set established in step S1 into the ConvNeXt neural network model obtained in step S2 for training to optimize the neural network model of ConvNeXt, and a weight file is generated for each training;
[0053] S4: use the trained ConvNeXt weight file to predict the thermodynamic parameters of the RDL circuit;
[0054] S5: import the thermodynamic parameters predicted by ConvNeXt into the parameters of the equivalent material to form an equivalent material that can be used for RDL thermodynamic simulation.
[0055] Specifically, as shown in Figure 2 A ConvNeXt-based RDL circuit composite material equivalent method, comprising the following steps:
[0056] S1: Establish a data set:
[0057] Through image data acquisition, a large number of RDL circuit images are modeled and relevant thermodynamic parameters are calculated, and a data set is established after data enhancement;
[0058] S2: Establish and improve the neural network model of ConvNeXt:
[0059] Establish the neural network model of ConvNeXt: adjust and configure the corresponding parameters;
[0060] Improve the neural network model of ConvNeXt: improve the classifier of ConvNeXt to a regressor and optimize the performance of the regressor;
[0061] S3: Train to optimize the neural network model of ConvNeXt
[0062] Input the data set enhanced in step S2 into the ConvNeXt neural network model for training; the ConvNeXt neural network model is trained once for each training, and a weight file is generated;
[0063] S4: ConvNeXt neural network model performs regression task:
[0064] Save the trained ConvNeXt weight file and use the weight file to predict the thermodynamic parameters of the RDL circuit;
[0065] S5: Construct RDL equivalent material:
[0066] Import the thermodynamic parameters predicted by ConvNeXt into the parameters of the equivalent material to form an equivalent material that can be used for RDL thermodynamic simulation.
[0067] AsFigure 5 As shown, the core idea of the present application is to construct an RDL circuit diagram and a data set of corresponding thermodynamic parameters by a finite element simulation method, then train a ConvNeXt neural network model according to the data set, and finally predict the RDL parameters of the RDL circuit diagram through the trained model.
[0068] Further, the steps of step S1 are as follows:
[0069] (S1a) Import the RDL layout with various circuit topologies into the electronic design automation software and export a high-definition PDF image;
[0070] (S1b) Use a Python script to extract the circuit contour data points of the image and reasonably number them;
[0071] (S1c) Import the contour data point file into the finite element software and use the APDL script for automatic modeling. According to the circuit layout, give the copper or silicon dioxide material properties to the contour, and in other embodiments, other suitable metal wire materials and dielectric materials can be selected, and it is not limited to copper and silicon dioxide. Build an RDL composite material model, which is a cuboid in shape, and in other embodiments, other suitable shapes are selected;
[0072] (S1d) Perform thermodynamic simulation on the RDL composite material model to obtain thermodynamic parameters. Calculate the Young's modulus in X, Y, and Z directions respectively. The specific calculation method is to select two faces perpendicular to the direction, fix one face, and stretch the other face to obtain the stress and strain values on the stretched face. Calculate the Young's modulus in that direction through the Young's modulus formula. Calculate the thermal expansion coefficient in X, Y, and Z directions respectively. The specific calculation method is to fix one of the faces perpendicular to the direction, set the initial temperature to 20℃, and raise the temperature by 10℃. Calculate the thermal strain of the other face, and calculate the thermal expansion coefficient in that direction through the thermal expansion coefficient calculation formula. Calculate the shear modulus of XY, YZ, and XZ planes respectively. The specific calculation method is to fix one of the faces perpendicular to the plane, apply a tensile force parallel to the plane on the corresponding face, calculate the average stress and strain of the face, and calculate the shear modulus in that direction through the shear modulus calculation formula; as Figure 3 As shown, according to the finite element simulation method, the corresponding constraints and loads are applied to the model, and the thermodynamic parameters of the model are calculated through the related theory of material mechanics;
[0073] (S1e) According to the geometric symmetry, the data obtained by simulation calculation is data augmented. The specific method is to rotate a picture 180° clockwise, flip it up and down, flip it up and down again, and then rotate it 180° clockwise to correspond to the same thermodynamic parameter value; as Figure 4 As shown.
[0074] As Figure 6 Further, the step S2 is as follows:
[0075] (S2a) constructing a ConvNeXt neural network model composed of ConvNet modules,
[0076] The feature extraction module has an input channel of 1, passes through a 4x4 convolution kernel, and has an output channel of 128. The LayerNorm2d normalization operation is used to normalize the output.
[0077] (S2b) passing through 4 first CNBlock modules, each of which has the following parameters: an input channel and an output channel of 128, a 7x7 convolution kernel, a LayerNorm layer, a linear layer with an input of 128 and an output of 512, a GELU activation layer, wherein the GELU activation function can be represented as: x*sigmoid(1.703x), and a linear layer with an input of 512 and an output of 128;
[0078] (S2c) passing through a first down-sampling layer: an input channel of 128, an output channel of 256, and a 2x2 convolution kernel;
[0079] (S2b) passing through 3 second CNBlock modules, each of which has the following parameters: an input channel and an output channel of 256, a 7x7 convolution kernel, a LayerNorm layer, a linear layer with an input of 256 and an output of 1024, a GELU activation layer, wherein the GELU activation function can be represented as: x*sigmoid(1.703x), and a linear layer with an input of 1024 and an output of 256;
[0080] (S2d) passing through a second down-sampling layer: an input channel of 256, an output channel of 512, and a 2x2 convolution kernel;
[0081] (S2e) passing through 27 third CNBlock modules, each of which has the following parameters: an input channel and an output channel of 512, a 7x7 convolution kernel, a LayerNorm layer, a linear layer with an input of 512 and an output of 2048, a GELU activation layer, wherein the GELU activation function can be represented as: x*sigmoid(1.703x), and a linear layer with an input of 2048 and an output of 512;
[0082] (S2f) passing through a third down-sampling layer: an input channel of 512, an output channel of 1024, and a 2x2 convolution kernel;
[0083] (S2g) pass through 3 fourth CNBlock modules, and the parameters of each fourth CNBlock module are as follows: the input channel and the output channel are both 1024, a convolution layer with a convolution kernel of 7x7, a LayerNorm layer, a linear layer with an input of 1024 and an output of 4096, a GELU activation layer, wherein the GELU activation function can be expressed as: x*sigmoid(1.703x), a linear layer with an input of 4096 and an output of 1024;
[0084] (S2h) regression module: composed of a normalization layer and a fully connected layer; the normalization layer is used to standardize the input feature map, helping to speed up the training and reduce the gradient vanishing problem; the fully connected layer performs linear transformation, with an input channel of 1024 and an output channel of 1, performing regression prediction.
[0085] Further, the steps of step S3 are as follows:
[0086] (S3a) set the batch_size to 32 and the number of iterations epoch to 200;
[0087] (S3b) use the Adam optimizer.
[0088] Further, the steps of step S4 are as follows:
[0089] (S4a) define the model with the minimum test set loss in 200 iterations as the optimal model and save it;
[0090] (S4b) use this optimal model weight file to predict the thermodynamic parameters of the RDL circuit.
[0091] Further, the steps of step S5 are as follows:
[0092] (S5a) input the RDL image into the ConvNeXt optimal neural network model to predict the Young's modulus and thermal expansion coefficient in X, Y and Z directions, and the shear modulus of XY, YZ and XZ planes;
[0093] (S5b) import the predicted values of the Young's modulus and thermal expansion coefficient in X, Y and Z directions, and the shear modulus of XY, YZ and XZ planes into the finite element software to form an RDL equivalent material for subsequent thermodynamic simulation.
[0094] Figure 7A performance comparison chart of the ConvNeXt neural network model provided by the embodiment of the present application and the traditional Voigt method is shown in the figure, from which it can be seen that in the prediction error quantization analysis, the average absolute error (MAE) of the model is reduced by more than 71% compared with the traditional Voigt model, and is only 18.14% of the latter; the average relative error (MRE) of 0.65% is only 17.66% of 3.68% of the Voigt model, and the reduction is more than 82%. Through the comparison and verification of the two core indicators, the ConvNeXt deep learning model has more advantages than the traditional Voigt model in the field of thermodynamic parameter prediction.
[0095] The application further provides an RDL circuit composite material equivalent system based on a ConvNeXt model, comprising a processor and a memory, the memory is used for storing program instructions, and the processor is used for calling the storage instructions in the memory to execute the RDL circuit composite material equivalent method based on the ConvNeXt model.
[0096] The above is only a preferred embodiment of the present application, of course, cannot be used to limit the scope of the present application, should be noted that, for those skilled in the art, without departing from the principles of the present application, can make a number of improvements and changes, these improvements and changes are also considered to be within the scope of the present application.
Claims
1. A method for equivalent composite materials in RDL circuits based on the ConvNeXt model, characterized in that, Includes the following steps: S1: Acquire image data, model multiple RDL circuit images and calculate relevant thermodynamic parameters, and build a dataset after data enhancement. This includes the following steps: S1a: Obtain images of RDL layouts with various circuit topologies; S1b: Extract and number the circuit contour data points from the image; S1c: Import the contour data point file into the finite element software for automated modeling, assign the material properties of metal wires or dielectrics to the contour according to the circuit layout, and construct the RDL composite material model. S1d: Perform thermodynamic simulation on the RDL composite material model to obtain thermodynamic parameters, and calculate Young's modulus and coefficient of thermal expansion in the X, Y, and Z directions, as well as shear modulus on the XY, YZ, and XZ planes respectively. S1e: Perform data augmentation on the simulation data based on geometric symmetry and establish a dataset; S2: Establish and improve the neural network model of ConvNeXt, specifically including the following steps: S2a: The module for feature extraction in the ConvNeXt neural network model, which normalizes the output by LayerNorm2d normalization after passing through convolution kernels; S2b: The first CNBlock module enhances the nonlinearity and feature extraction capabilities of the neural network; S2c: Convolution after the first downsampling layer; S2d: The second CNBlock module enhances the nonlinearity and feature extraction capabilities of the neural network; S2e: Convolution after the second downsampling layer; S2f: The third CNBlock module enhances the nonlinearity and feature extraction capabilities of the neural network; S2g: Convolution after the third downsampling layer; S2h: The fourth CNBlock module enhances the nonlinearity and feature extraction capabilities of the neural network; S2i: After processing through normalization and fully connected layers, regression prediction is performed to complete the establishment and improvement of the ConvNeXt neural network model; S3: Input the dataset established in step S1 into the ConvNeXt neural network model obtained in step S2 for training to optimize the ConvNeXt neural network model. Each time the training is performed, a weight file is generated accordingly. S4: Predict the thermodynamic parameters of the RDL circuit using the trained ConvNeXt weight file; S5: Import the thermodynamic parameters predicted by ConvNeXt into the parameters of the equivalent material to form an equivalent material that can be used for RDL thermodynamic simulation.
2. The equivalent method for RDL circuit composite materials based on the ConvNeXt model according to claim 1, characterized in that, In step S1d, the specific method for calculating Young's modulus is to select two surfaces perpendicular to the direction, fix one surface, stretch the other surface, obtain the stress and strain values on the stretched surface, and calculate Young's modulus in that direction using the Young's modulus formula; the specific method for calculating thermal expansion coefficient is to fix one surface perpendicular to the direction, set an initial temperature, calculate the thermal strain of the other surface after heating, and calculate thermal expansion coefficient in that direction using the thermal expansion coefficient calculation formula; the specific method for calculating shear modulus is to fix one surface perpendicular to the direction, apply a tensile force parallel to the surface on the corresponding surface, calculate the average stress and strain of the surface, and calculate shear modulus in that direction using the shear modulus calculation formula.
3. The equivalent method for RDL circuit composite materials based on the ConvNeXt model according to claim 1, characterized in that, In step S1e, the specific method for data augmentation based on geometric symmetry is to rotate an image 180° clockwise, flip it vertically, and then rotate it 180° clockwise again to obtain the same thermodynamic parameter value.
4. The equivalent method for RDL circuit composite materials based on the ConvNeXt model according to claim 1, characterized in that, The first, second, third, and fourth CNBlock modules all consist of, from top to bottom, a convolutional layer, a LayerNorm layer, a linear layer, a GELU activation layer, and another linear layer, where the GELU activation function is expressed as: x sigmoid (1.703x).
5. The equivalent method for RDL circuit composite materials based on the ConvNeXt model according to claim 1, characterized in that, The input and output channels of the first, second, third, and fourth CNBlock modules increase sequentially.
6. The equivalent method for RDL circuit composite materials based on the ConvNeXt model according to claim 1, characterized in that, The input and output channels of the first, second, and third downsampling layers increase sequentially.
7. The equivalent method for RDL circuit composite materials based on the ConvNeXt model according to claim 1, characterized in that, Step S4 specifically includes the following steps: The RDL image is input into the constructed ConvNeXt neural network model to predict the Young's modulus and thermal expansion coefficient in the X, Y, and Z directions, as well as the shear modulus on the XY, YZ, and XZ planes, respectively. Step S5 specifically includes the following steps: The predicted values of Young's modulus and coefficient of thermal expansion in the X, Y, and Z directions, as well as the predicted values of shear modulus on the XY, YZ, and XZ planes, are imported into the finite element software to form an RDL equivalent material for subsequent thermodynamic simulation.
8. An equivalent system of composite materials for RDL circuits based on the ConvNeXt model, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute the equivalent method of RDL circuit composite material based on the ConvNeXt model as described in any one of claims 1-7.