RDL circuit composite material equivalence method and system based on ConvNeXt model

Abstract

The invention relates to the technical field of semiconductor manufacturing thermodynamic simulation, in particular to an RDL circuit composite material equivalence method and system based on a ConvNeXt model, and the method comprises the following steps: collecting image data, carrying out the modeling of a plurality of RDL circuit images, calculating the related thermodynamic parameters, carrying out the enhancement of the data, and building a data set; establishing and improving a neural network model of ConvNeXt; inputting the data set into the neural network model for training so as to optimize the model, and correspondingly generating a weight file each time training is performed; performing thermodynamic parameter prediction on the RDL circuit by using the trained weight file; and the predicted thermodynamic parameters are imported into parameters of the equivalent material to form the equivalent material capable of being used for RDL thermodynamic simulation. According to the method, the database is constructed according to the simulation data, the neural network model capable of predicting the thermodynamic parameters of the equivalent model according to the RDL image is established and trained, and the simulation of the large-scale RDL model is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor manufacturing thermodynamics simulation, and particularly relates to a method and system for equivalent of RDL circuit composite materials based on the ConvNeXt model. Background Art

[0002] With the development of technologies such as artificial intelligence, big data, and cloud computing, chips are required to have higher computing speeds, higher information transmission densities, as well as relatively lower costs and power consumptions. As the development of advanced chip manufacturing processes approaches the physical limit, Moore's Law gradually fails, and the chip field has entered the post-Moore era.

[0003] To further improve the computing speed and data transmission speed of chips, advanced packaging technologies have become new solutions. 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 the silicon interposer. Its main function is to redistribute the electrical connection points (I / O pads) on the chip. By adding one or more metal wiring layers on the chip surface, these connection points are rearranged to more suitable positions for packaging and interconnection. Due to problems such as the mismatch of the coefficient of thermal expansion (CTE) of RDL materials and the uneven topological distribution of the microstructure, after the silicon interposer wafer experiences heating and cooling during the manufacturing process, internal stress will be generated, resulting in warping. Excessive wafer warping is likely to cause the wafer position to shift, which has a negative impact on the alignment of lithography during the processing. In addition, excessive warping may also cause solder ball displacement, wafer fracture and other problems, seriously affecting the product yield. Therefore, predicting wafer warping through simulation, deeply exploring the causes of warping in the redistribution process, and optimizing warping through design parameters are of great significance for improving the production yield of silicon interposer wafers, reducing production costs, and promoting the development of China's semiconductor advanced packaging industry.

[0004] At the macroscopic scale, one of the biggest technical problems in large-scale RDL simulation research is the difficulty of completely distortion-free modeling and simulation of its complex copper traces. General equivalent methods simplify the model too much and fail to capture the influence of the shape distribution of copper traces on stress and warping, resulting in excessive errors in the results. Summary of the Invention

[0005] One of the objectives of the present invention is to provide an equivalent method for RDL circuit composite materials based on the ConvNeXt model. It is intended to use the method of machine learning to construct a database according to simulation data, establish a deep learning multi-modal parameter analysis model, and train a neural network model that can accurately predict the thermodynamic parameters of the equivalent model based on the RDL image, so as to achieve high-fidelity simulation of large-scale RDL models.

[0006] Another objective of the present invention is to provide an equivalent system for RDL circuit composite materials based on the ConvNeXt model.

[0007] The solution adopted by the present invention to achieve the first objective is as follows: An equivalent method for RDL circuit composite materials based on the ConvNeXt model, including the following steps:

[0008] S1: Collect image data, model multiple RDL circuit images and calculate relevant thermodynamic parameters, and establish a data set after data enhancement;

[0009] S2: Establish and improve the neural network model of ConvNeXt;

[0010] S3: Input the data set 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 of training generates a weight file correspondingly;

[0011] S4: Use the trained ConvNeXt weight file to predict the thermodynamic parameters of the RDL circuit;

[0012] 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.

[0013] Preferably, step S1 includes the following steps:

[0014] S1a: Obtain images of RDL layouts with various circuit topologies;

[0015] S1b: Extract the circuit contour data points of the image and number them;

[0016] S1c: Import the contour data point file into finite element software for automated modeling, and assign the material properties of metal wires or dielectrics to the contour according to the circuit layout to construct an RDL composite material model;

[0017] S1d: Perform thermodynamic simulation on the RDL composite material model to obtain thermodynamic parameters, and calculate the Young's modulus, coefficient of thermal expansion in the X, Y, and Z directions respectively, and the shear modulus on the XY, YZ, and XZ planes;

[0018] S1e: Enhance the data obtained from the simulation calculation according to geometric symmetry and establish a data set.

[0019] Preferably, in step S1d, the specific calculation method of Young's modulus is to select two faces perpendicular to this direction, fix one of the faces, stretch the other face, obtain the stress and strain values on the stretched face, and calculate the Young's modulus in this direction through the Young's modulus formula; the specific calculation method of the coefficient of thermal expansion is to fix one of the faces perpendicular to this direction, set an initial temperature, calculate the thermal strain of the other face after heating, and calculate the coefficient of thermal expansion in this direction through the coefficient of thermal expansion calculation formula; the specific calculation method of the shear modulus is to fix one of the faces perpendicular to this face, apply a tensile force parallel to this face on the corresponding face, calculate the average stress and strain of this face, and calculate the shear modulus in this direction through the shear modulus calculation formula.

[0020] Preferably, in step S1e, the specific method of enhancing the data obtained from the simulation calculation according to geometric symmetry is to perform a 180° clockwise rotation, an up-down flip, and an up-down flip followed by a 180° clockwise rotation on a picture, corresponding to the same thermodynamic parameter value.

[0021] Preferably, step S2 includes the following steps:

[0022] S2a: Construct a feature extraction module of the ConvNeXt neural network model, and normalize the output through the LayerNorm2d normalization operation after the convolutional kernel.

[0023] S2b: Increase the nonlinearity and feature extraction ability of the neural network through the first CNBlock module.

[0024] S2c: Extract higher-level feature information through the first downsampling layer convolution.

[0025] S2d: Increase the nonlinearity and feature extraction ability of the neural network through the second CNBlock module.

[0026] S2e: Extract higher-level feature information through the second downsampling layer convolution.

[0027] S2f: Increase the nonlinearity and feature extraction ability of the neural network through the third CNBlock module.

[0028] S2g: Extract higher-level feature information through the third downsampling layer convolution.

[0029] S2h: Increase the nonlinearity and feature extraction ability of the neural network through the fourth CNBlock module.

[0030] S2i: After being processed by the normalization layer and the fully connected layer, regression prediction is performed to complete 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 all sequentially include a convolutional layer, a LayerNorm layer, a linear layer, a GELU activation layer, and a linear layer from top to bottom, where the GELU activation function can be expressed 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 sequentially.

[0033] Preferably, the input channels and output channels of the first downsampling layer, the second downsampling layer, and the third downsampling layer increase sequentially.

[0034] Preferably, step S5 includes the following steps:

[0035] S5a: Input the RDL image into the constructed ConvNeXt neural network model to respectively predict the Young's modulus, coefficient of thermal expansion in the X, Y, and Z directions, and the shear modulus on the XY, YZ, and XZ planes;

[0036] S5b: Import the predicted values of the Young's modulus, coefficient of thermal expansion in the X, Y, and Z directions, and the shear modulus on the XY, YZ, and XZ planes into the finite element software to form an RDL equivalent material for subsequent thermodynamic simulation.

[0037] The solution adopted by the present invention to achieve the second object is: An RDL circuit composite material equivalent system based on the ConvNeXt model, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the RDL circuit composite material equivalent method based on the ConvNeXt model.

[0038] The present invention has the following advantages and beneficial effects:

[0039] The method and system of the present invention are intended to adopt the method of machine learning, construct a database according to simulation data, establish a deep learning multi-modal parameter analysis model, and train a neural network model that can accurately predict the thermodynamic parameters of the equivalent model based on the RDL image, so as to achieve high-fidelity simulation of large-scale RDL models. Description of the Drawings

[0040] Figure 1Flow chart of the RDL circuit composite material equivalent method based on ConvNeXt of the present invention;

[0041] Figure 2 Flow chart of the RDL circuit composite material equivalent method based on ConvNeXt provided by an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the method for obtaining the thermodynamic parameters of the RDL circuit based on finite element simulation provided by an embodiment of the present invention;

[0043] Figure 4 Schematic diagram of the method for realizing data augmentation based on geometric transformation provided by an embodiment of the present invention;

[0044] Figure 5 Flow chart of the method for constructing a ConvNeXt neural network model to predict the thermodynamic parameters of the RDL circuit based on the deep learning method provided by an embodiment of the present invention;

[0045] Figure 6 Schematic diagram of the backbone network structure of the ConvNeXt neural network model provided by an embodiment of the present invention;

[0046] Figure 7 Performance comparison chart of the ConvNeXt neural network model and the traditional Voigt method provided by an embodiment of the present invention. Detailed implementation manners

[0047] For a better understanding of the present invention, the following embodiments are further descriptions of the present invention, but the content of the present invention is not limited to the following embodiments only.

[0048] High-fidelity thermodynamic simulation of the RDL circuit is of great significance for high-precision prediction of wafer warping and improvement of the reliability of 2.5D packaging. The present invention aims at the problem of low prediction accuracy of the thermodynamic properties of RDL circuit composite materials by traditional methods at present, and realizes high-precision prediction of the thermodynamic properties of RDL circuit composite materials by building a ConvNeXt neural network model through finite element simulation.

[0049] As Figure 1 shown, a method for equivalent of RDL circuit composite materials based on the ConvNeXt model includes the following steps:

[0050] S1: Collect image data, model multiple RDL circuit images and calculate relevant thermodynamic parameters, and establish a data set after data augmentation;

[0051] S2: Establish and improve the neural network model of ConvNeXt;

[0052] 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 training is performed, a weight file is generated correspondingly.

[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 Figure 2 shown, a method for equivalent of RDL circuit composite materials based on ConvNeXt includes the following steps:

[0056] S1: Establish a dataset:

[0057] Through the acquisition of image data, model a large number of RDL circuit images and calculate relevant thermodynamic parameters. After data augmentation, establish a dataset.

[0058] S2: Establish and improve the ConvNeXt neural network model:

[0059] Establish the ConvNeXt neural network model: Adjust and configure corresponding parameters;

[0060] Improve the ConvNeXt neural network model: Improve the classifier of ConvNeXt into a regressor and optimize the performance of the regressor;

[0061] S3: Train to optimize the ConvNeXt neural network model

[0062] Input the dataset enhanced in step S2 into the ConvNeXt neural network model for training; Each time the ConvNeXt neural network model is trained, a weight file is generated correspondingly.

[0063] S4: The ConvNeXt neural network model performs a regression task:

[0064] Save the trained ConvNeXt weight file and use this weight file to predict the thermodynamic parameters of the RDL circuit;

[0065] S5: Construct an 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 invention is to construct a dataset of RDL circuit diagrams and corresponding thermodynamic parameters through the finite element simulation method, then train a ConvNeXt neural network model based on the dataset, 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) Obtain RDL layouts with various circuit topologies, import them into electronic design automation software, and export high-definition PDF images;

[0070] (S1b) Use a Python script to extract the circuit contour data points of the image and number them reasonably;

[0071] (S1c) Import the contour data point file into finite element software and perform automated modeling with APDL scripts. Assign material properties of copper or silicon dioxide to the contour according to the circuit layout. In other embodiments, other suitable metal wire materials and dielectric materials can be selected, not limited to copper and silicon dioxide. Build an RDL composite material model. The overall shape of this model is a cuboid, and other suitable shapes can be selected in other embodiments;

[0072] (S1d) Perform thermodynamic simulation on the RDL composite material model to obtain thermodynamic parameters. Calculate the Young's modulus in the X, Y, and Z directions respectively. The specific calculation method is to select two planes perpendicular to this direction, fix one of the planes, stretch the other plane, and obtain the stress and strain values on the stretched plane, and calculate the Young's modulus in this direction through the Young's modulus formula; calculate the coefficient of thermal expansion in the X, Y, and Z directions respectively. The specific calculation method is to fix one of the planes perpendicular to this direction, set the initial temperature to 20°C, raise the temperature by 10°C, calculate the thermal strain of the other plane, and calculate the coefficient of thermal expansion in this direction through the coefficient of thermal expansion calculation formula; calculate the shear modulus on the XY, YZ, and XZ planes respectively. The specific calculation method is to fix one of the planes perpendicular to this plane, apply a tensile force parallel to this plane on the corresponding plane, calculate the average stress and strain of this plane, and calculate the shear modulus in this direction through the shear modulus calculation formula; As Figure 3 shown, according to the finite element simulation method, apply corresponding constraints and loads to the model, and then calculate the thermodynamic parameters of the model through relevant theories of material mechanics;

[0073] (S1e) Perform data augmentation on the data obtained from the simulation calculation according to geometric symmetry. The specific method is to rotate a picture 180° clockwise, flip it up and down, and then rotate it 180° clockwise after flipping it up and down, corresponding to the same thermodynamic parameter value; As Figure 4 shown.

[0074] As Figure 6 shown, further, the steps of step S2 are as follows:

[0075] (S2a) Construct a ConvNeXt neural network model composed of ConvNet modules,

[0076] Feature extraction module: The input channel is 1, passing through a 4x4 convolutional kernel, and the output channel is normalized by a LayerNorm2d normalization operation;

[0077] (S2b) Pass through 4 first CNBlock modules, and the parameters of each first CNBlock module are as follows: both the input channel and the output channel are 128, a convolutional layer with a 7×7 convolutional kernel, a LayerNorm layer, a linear layer with an input of 128 and an output of 512, a GELU activation layer, where the GELU activation function can be expressed as: x*sigmoid(1.703x), and a linear layer with an input of 512 and an output of 128;

[0078] (S2c) Pass through the first downsampling layer: the input channel is 128, the output channel is 256, and a convolutional layer with a 2×2 convolutional kernel;

[0079] (S2b) Pass through 3 second CNBlock modules, and the parameters of each second CNBlock module are as follows: both the input channel and the output channel are 256, a convolutional layer with a 7×7 convolutional kernel, a LayerNorm layer, a linear layer with an input of 256 and an output of 1024, a GELU activation layer, where the GELU activation function can be expressed as: x*sigmoid(1.703x), and a linear layer with an input of 1024 and an output of 256;

[0080] (S2d) Pass through the second downsampling layer: the input channel is 256, the output channel is 512, and a convolutional layer with a 2×2 convolutional kernel;

[0081] (S2e) Pass through 27 third CNBlock modules, and the parameters of each third CNBlock module are as follows: both the input channel and the output channel are 512, a convolutional layer with a 7×7 convolutional kernel, a LayerNorm layer, a linear layer with an input of 512 and an output of 2048, a GELU activation layer, where the GELU activation function can be expressed as: x*sigmoid(1.703x), and a linear layer with an input of 2048 and an output of 512;

[0082] (S2f) Pass through the third downsampling layer: the input channel is 512, the output channel is 1024, and a convolutional layer with a 2×2 convolutional kernel;

[0083] (S2g) Pass through three fourth CNBlock modules. The parameters of each fourth CNBlock module are as follows: both the input channels and output channels are 1024, a convolutional layer with a 7×7 convolutional kernel, a LayerNorm layer, a linear layer with an input of 1024 and an output of 4096, a GELU activation layer, where the GELU activation function can be expressed as: x * sigmoid(1.703x), and 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 to help accelerate training and reduce the problem of gradient disappearance; the fully connected layer performs a linear transformation, with an input channel of 1024 and an output channel of 1, and performs regression prediction.

[0085] Further, the steps of step S3 are as follows:

[0086] (S3a) Set batch_size to 32 and the number of epochs 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 optimal ConvNeXt neural network model to predict its Young's modulus and coefficient of thermal expansion in the X, Y, and Z directions, as well as the shear modulus on the XY, YZ, and XZ planes;

[0093] (S5b) Import the predicted values of its Young's modulus and coefficient of thermal expansion in the X, Y, and Z directions, as well as the shear modulus on the XY, YZ, and XZ planes into the finite element software to construct an RDL equivalent material for subsequent thermodynamic simulations.

[0094] Figure 7This is a performance comparison chart of the ConvNeXt neural network model provided by the embodiments of the present invention and the traditional Voigt method. It can be seen from the chart that in the prediction error quantization analysis, the mean absolute error (MAE) of this model is reduced by more than 71% compared with the traditional Voigt model, only being 18.14% of the latter; the error value of its mean relative error (MRE) of 0.65% is only equivalent to 17.66% of 3.68% of the Voigt model, with a decrease of more than 82%. Through the comparison and verification of these 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 present invention also provides an equivalent system of an RDL circuit composite material based on the ConvNeXt model, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the equivalent method of the RDL circuit composite material based on the ConvNeXt model.

[0096] The above is the preferred implementation manner of the present invention. Of course, the scope of the rights of the present invention cannot be limited by this. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and changes can still be made, and these improvements and changes are also regarded as the protection scope of the present invention.

Claims

1. A method for equivalent RDL circuit composite materials based on ConvNeXt model, characterized in that: The following steps are involved: S1: Collect image data, model multiple RDL circuit images and calculate relevant thermodynamic parameters, and establish a data set after enhancing the data; S2: Build and improve the ConvNeXt neural network model; S3: input the data set established in step S1 into the ConvNeXt neural network model obtained in step S2 for training to optimize the ConvNeXt neural network model, and generate a weight file for each training; S4: Use the trained ConvNeXt weight file to predict the thermodynamic parameters of the RDL circuit; 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 RDL circuit composite material equivalent method based on the ConvNeXt model according to claim 1, characterized in that: Step S1 includes the following steps: S1a: Acquire images of RDL layouts with various circuit topologies; S1b: extracting the circuit contour data points of the image and numbering them; S1c: Import the contour data point file into the finite element software for automatic modeling, assign the material properties of the metal wire or dielectric to the contour according to the circuit layout, and build 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 thermal expansion coefficient in the three directions of X, Y, and Z, and shear modulus on the three planes of XY, YZ, and XZ; S1e: Perform data enhancement on the data obtained from simulation calculations based on geometric symmetry and establish a data set.

3. The RDL circuit composite material equivalent method based on the ConvNeXt model according to claim 2 is characterized in that: In step S1d, the specific calculation method of Young's modulus is to select two surfaces perpendicular to the direction, fix one of the surfaces, stretch the other surface, obtain the stress and strain values ​​on the stretched surface, 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 surfaces perpendicular to the direction, set an initial temperature, calculate the thermal strain of the other surface 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 surfaces perpendicular to the surface, apply a tensile force parallel to the surface on the corresponding surface, calculate the average stress and strain of the surface, and calculate the shear modulus in the direction through the shear modulus calculation formula.

4. The RDL circuit composite material equivalent method based on the ConvNeXt model according to claim 2, characterized in that: In step S1e, a specific method of performing data enhancement on the data obtained by simulation calculation according to geometric symmetry is to rotate an image 180° clockwise, flip it upside down, flip it upside down and then rotate it 180° clockwise to obtain the same thermodynamic parameter value.

5. The RDL circuit composite material equivalent method based on the ConvNeXt model according to claim 1, characterized in that: Step S2 includes the following steps: S2a: Constructs the feature extraction module of the ConvNeXt neural network model, and normalizes the output by the LayerNorm2d normalization operation after the convolution kernel; S2b: Increase the nonlinearity and feature extraction capabilities of the neural network through the first CNBlock module; S2c: convolution after the first downsampling layer; S2d: The nonlinearity and feature extraction capabilities of the neural network are increased through the second CNBlock module; S2e: convolution after the second downsampling layer; S2f: The third CNBlock module increases the nonlinearity and feature extraction capabilities of the neural network; S2g: convolution after the third downsampling layer; S2h: The nonlinearity and feature extraction capabilities of the neural network are increased through the fourth CNBlock module; S2i: After being processed by the normalization layer and the fully connected layer, regression prediction is performed to complete the establishment and improvement of the ConvNeXt neural network model.

6. The RDL circuit composite material equivalent method based on the ConvNeXt model according to claim 5, characterized in that: The structures of the first CNBlock module, the second CNBlock module, the third CNBlock module, and the fourth CNBlock module all include, from top to bottom, a convolutional layer, a LayerNorm layer, a linear layer, a GELU activation layer, and a linear layer, wherein the GELU activation function can be expressed as: x*sigmoid(1.703x).

7. The RDL circuit composite material equivalent method based on the ConvNeXt model according to claim 5, characterized in that: 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 sequence.

8. The RDL circuit composite material equivalent method based on the ConvNeXt model according to claim 5, characterized in that: The input channels and output channels of the first downsampling layer, the second downsampling layer and the third downsampling layer increase in sequence.

9. The RDL circuit composite material equivalent method based on the ConvNeXt model according to claim 1, characterized in that: Step S5 includes the following steps: S5a: The RDL image is input into the constructed ConvNeXt neural network model to predict the Young's modulus and thermal expansion coefficient in the three directions of X, Y, and Z, and the shear modulus on the three planes of XY, YZ, and XZ respectively; S5b: Import the Young's modulus and thermal expansion coefficient in the three directions of X, Y, and Z and the predicted values ​​of shear modulus on the three planes of XY, YZ, and XZ into the finite element software to form the RDL equivalent material for subsequent thermodynamic simulation.

10. An RDL circuit composite material equivalent system based on the ConvNeXt model, characterized in that: The invention comprises a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the RDL circuit composite material equivalent method based on the ConvNeXt model as described in any one of claims 1 to 9.

Images