Parasitic parameter extraction method based on neural network
Through the parasitic parameter extraction method based on convolutional neural network, the problems of high computational cost and low accuracy in the traditional method are solved, and efficient and accurate parasitic parameter extraction is achieved, which is suitable for various circuit layouts and improves the design of integrated circuits. efficiency.
Patent Information
- Application Number
- CN202411018138.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-12-03
AI Technical Summary
The traditional parasitic parameter extraction method has insufficient calculation cost and efficiency. The field solver has high computational cost and complex operation, while the model based on the lookup table is not accurate and has low flexibility, making it difficult to be between accuracy, calculation cost and flexibility. Find balance.
Using a method based on convolutional neural network, the two-dimensional structural pattern of the integrated circuit is generated, grid-based feature representation is performed, and the convolutional neural network model is trained using the mean square error and Hessian non-diagonal element squares. and as a loss function, combined with the Adam optimizer, to achieve fast and accurate parasitic parameter extraction.
It realizes high-precision parasitic parameter extraction, shortens the design cycle, adapts to various circuit layouts, improves design efficiency, and reduces calculation costs.
Smart Images

Figure CN119067045A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of electronic design automation, and in particular to a parasitic parameter extraction method based on a neural network. Background Art
[0002] In integrated circuit design, parasitic parameters are key indicators in integrated circuit design, affecting the performance of the circuit, such as parasitic capacitance, parasitic inductance, etc. Accurately extracting parasitic parameters is crucial to ensure circuit performance. As technology nodes continue to shrink, these tasks become more and more complex, and traditional calculation methods face the challenges of high computational cost and low efficiency.
[0003] Traditional parasitic parameter extraction methods mainly include field solvers and models based on lookup tables. Field solvers calculate the electromagnetic field distribution by solving Maxwell's equations, providing high-precision parasitic parameters, but the calculation cost is high, requiring a lot of computing resources and time, and the operation is complex. The model based on the lookup table pre-calculates and stores the parasitic parameters of different circuit structures to form a lookup table. It has fast calculation speed, is easy to implement and integrate, but the accuracy may not be as good as the field solver, relies on pre-calculated data, and has low flexibility. It is crucial to choose an appropriate parasitic parameter extraction method based on the trade-off between accuracy, computational cost, and flexibility. Summary of the invention
[0004] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a method for extracting parasitic parameters in integrated circuits based on neural networks. The present invention aims to combine the advantages of high precision and high efficiency, and quickly and accurately extract parasitic parameters (parasitic capacitance, parasitic inductance, parasitic resistance, etc.) through a convolutional neural network model, thereby reducing calculation costs and improving design efficiency.
[0005] The technical solution of the present invention is specifically described as follows.
[0006] The present invention provides a parasitic parameter extraction method based on a neural network, which comprises the following steps:
[0007] 1) Randomly generate two-dimensional structural patterns of integrated circuits to simulate integrated circuit layouts. The integrated circuit wiring patterns include axial symmetry, diagonal symmetry, and asymmetry. The other characteristics are the same, but the wiring with different metal numbers in each layer belongs to the same pattern; for each pattern, generate sample structures with different numbers of conductors, and train a neural network for each pattern;
[0008] 2) gridding the density feature representation of the two-dimensional structure pattern of the integrated circuit structure, converting the layout into a high-dimensional vector, and obtaining a two-dimensional matrix of the integrated circuit structure;
[0009] 3) Model training: Using the two-dimensional matrix of the integrated circuit structure as input and the corresponding parasitic parameters as output, a convolutional neural network model based on the ResNet architecture is trained;
[0010] 4) Prediction: Input the new two-dimensional pattern structure pattern into the trained convolutional neural network model to achieve rapid prediction of the parasitic parameters of the integrated circuit structure to be tested.
[0011] In the present invention, in step 1), the two-dimensional structure pattern follows defined process rules, including the minimum width and spacing of the metal layer.
[0012] In the present invention, in step 1), the two-dimensional pattern of the integrated circuit structure is evenly divided into a number of grid units according to the width of the metal layer, and each grid unit is assigned a different weight according to the different dielectric or metal layers it occupies, thereby obtaining a feature vector of each grid unit.
[0013] In the present invention, in step 3), the convolutional neural network model is a ResNet50 architecture.
[0014] In the present invention, in step 3), the training process includes:
[0015] a. Use the mean square error MSE and the sum of the non-diagonal elements of the Hessian as the loss function for optimization; the loss function loss formula is as follows:
[0016]
[0017] Where α is a hyperparameter, C is a parasitic parameter, and C predict is the predicted parasitic function, x ij is the element of the Hessian matrix;
[0018] b. Use Adam optimizer to update model parameters.
[0019] In the present invention, in step 3), the parasitic parameters used for training are obtained by professional EDA software such as Mentor Calibre.
[0020] In the present invention, in step 4), the new two-dimensional structure pattern is gridded and represented by density features.
[0021] In the present invention, in step 4), during prediction, the trained model runs on the GPU server.
[0022] The present invention also provides a computer-readable storage medium on which instructions are stored, and the above-mentioned parasitic parameter extraction method is implemented when the instructions are executed.
[0023] As described above, the present invention uses a neural network model to predict the parameters of a model with a variable number of conductors and dielectrics. During the training and prediction process of the neural network model, the input data is a feature vector processed by a gridded density feature representation, and the mean square error and the sum of the squares of the Hessian non-diagonal elements are used as loss functions for training. The Adam optimizer is used to update the model parameters. Compared with the prior art, the present invention has the following advantages:
[0024] 1) Compared with the field solver, the extraction accuracy of parasitic parameters of the present invention can be maintained at the same order of magnitude.
[0025] 2) The present invention can improve the solution speed and shorten the design cycle.
[0026] 3) The present invention can adapt to various circuit layouts and has good versatility. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 : Example diagram of 2D integrated circuit layout.
[0028] Figure 2 : Gridded data representation of two-dimensional structural patterns.
[0029] Figure 3 : The architecture of the model.
[0030] Figure 4 : Optimization of loss function during training. DETAILED DESCRIPTION
[0031] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.
[0032] The present invention provides a parasitic parameter extraction method based on a neural network, the method comprising:
[0033] Step 1: Meshed Data Representation: Converting the 2D structural pattern into a meshed, dense feature representation allows the pattern to contain varying numbers of conductors, thus significantly reducing the number of patterns that need to be modeled.
[0034] Step 2: Neural Network Model: A CNN neural network model is proposed to predict the parasitic parameters of modes with varying numbers of conductors and dielectrics. The model is able to capture spatial information and improve the accuracy of parasitic parameter prediction through training skills.
[0035] Step 3: Training, including a special loss function for more accurate convergence, makes the model suitable for modeling parasitic parameters.
[0036] Step 4: Inference, the model runs much faster on GPU servers than the field solver, and its memory consumption is negligible compared to the parasitic parameter model based on the lookup table.
[0037] The specific implementation method is as follows.
[0038] Data preparation: Figure 1 , according to a specific process technology, such as 55nm or 15nm process, the geometric parameters of the metal layer are defined, including the minimum width (wmin), minimum spacing (smin) and thickness.
[0039] Pattern generation: Randomly generate 2D pattern structures. We randomly generate 2D structures with random number of metals per layer, random positions, and random size of each metal. These structures follow the pattern requirements (such as symmetry vs. asymmetry), and the foundry's process rules, including minimum width and spacing. For each pattern, sample structures with different numbers of conductors are generated.
[0040] Integrated circuit wiring patterns include symmetry and asymmetry, the amount of metal in each layer, etc., which patterns are generally determined by the foundry, and then the estimated parasitic parameters are obtained by using a lookup table through pattern matching. Traditional EDA requires a fixed number of conductors in each layer of the pattern, while the present invention allows the number of conductors to be changed, thereby greatly reducing the number of patterns that need to be modeled (merging multiple foundry patterns), and obtains parasitic parameters by calling a neural network instead of a lookup table through pattern matching.
[0041] Grid-based data representation: Figure 2 , each two-dimensional pattern is evenly divided into grid units. Different weights are assigned to different grid units according to the different dielectric or metal layers they occupy, and the corresponding feature vector is generated for each grid.
[0042] Model architecture: Figure 3 , implement a neural network model based on the ResNet architecture, use convolutional layers to process the input vector data, and insert batch normalization and ReLU activation layers after each layer. Use convolutional layers and connection layers to improve learning capabilities.
[0043] Model training: Figure 4 , use the optimizer to train the model, determine the learning rate and batch size. Use the mean square error (MSE) and the sum of the squares of the non-diagonal elements of the Hessian matrix as the loss function.
[0044]
[0045] where x ijis the element of the Hessian matrix, C is the parasitic parameter, such as parasitic capacitance, parasitic inductance, parasitic resistance, etc. The electromagnetic field solver is used to obtain the accurate parasitic parameters, and α is a hyperparameter that needs to be set manually.
[0046] Hyperparameter Tuning: Tune the model’s hyperparameters, including learning rate, batch size, alpha, number and size of convolutional layers, to achieve the best predictive performance.
[0047] Model evaluation: The trained model is evaluated using the validation dataset and the error between the predicted parasitic parameter values and the actual parasitic parameter values is calculated.
[0048] Prediction process: In the prediction stage, the new two-dimensional structural pattern is input into the trained model to quickly predict the parasitic parameters.
[0049] Example 1
[0050] In the 15nm process technology data set, a 15nm process technology based on FreePDK15 was tested. For each process technology, some three-layer metal layer combinations were selected to generate 2-D patterns, with the metal medium between the same layer as medium 1 and the metal medium between different layers as medium 2. For each pattern, several two-dimensional patterns were taken to form a data set, and then the data set was randomly divided into a training subset (accounting for 90% of the samples) and a test subset (accounting for 10% of the samples). The settings of the neural network are: skeleton Resnet50, metal is 999 in the grid weight distribution, the medium is the corresponding dielectric coefficient, if there are multiple materials in a grid, they are weighted according to the material area, the hyperparameter α is 0.9, the learning rate is 10-4, the batch size is 64, and the training time is about 1h on an A100.
[0051] The results show that the accuracy of the proposed model on the capacitance test data set is less than 8% in 99% of the cases, while the original loss function accuracy is less than 10% in 99% of the cases, and it can provide accurate parasitic parameter prediction for 2-D mode structures with a variable number of conductors. In addition, the proposed model has a significant improvement in running time compared to traditional 2-D field solvers, while consuming negligible memory, which further proves its effectiveness and practicality in full-chip parasitic parameter extraction.
Claims
1. A parasitic parameter extraction method based on neural network, characterized in that: The following steps are involved: 1) Randomly generate two-dimensional structural patterns of integrated circuits to simulate integrated circuit layouts. The integrated circuit wiring patterns include axial symmetry, diagonal symmetry, and asymmetry. The other characteristics are the same, but the wiring with different metal numbers in each layer belongs to the same pattern; for each pattern, generate sample structures with different numbers of conductors, and train a neural network for each pattern; 2) gridding the density feature representation of the two-dimensional structure pattern of the integrated circuit structure, converting the layout into a high-dimensional vector, and obtaining a two-dimensional matrix of the integrated circuit structure; 3) Model training: Using the two-dimensional matrix of the integrated circuit structure as input and the corresponding parasitic parameters as output, a convolutional neural network model based on the ResNet architecture is trained; 4) Prediction: Input the new two-dimensional structure pattern into the trained convolutional neural network model to achieve rapid prediction of the parasitic parameters of the integrated circuit structure to be tested.
2. The method according to claim 1, characterized in that In step 1), the two-dimensional structure pattern follows defined process rules, including the minimum width and spacing of the metal layers.
3. The method according to claim 1, characterized in that In step 1), the two-dimensional structure pattern of the integrated circuit structure is evenly divided into a number of grid units according to the width of the metal layer. Different weights are assigned to each grid unit according to the different dielectric or metal layers it occupies, and a feature vector of each grid unit is obtained.
4. The method according to claim 1, characterized in that: In step 3), the parasitic parameters used for training are obtained by the mentorcalibr tool.
5. The method according to claim 1, characterized in that In step 3), the convolutional neural network model adopts the ResNet50 architecture.
6. The method according to claim 1, characterized in that In step 3), the training process includes: a. Use the mean square error MSE and the sum of the non-diagonal elements of the Hessian as the loss function for optimization; the loss function loss formula is as follows: Where α is a hyperparameter, C is a parasitic parameter, and C predict is the predicted parasitic function, x ij is the element of the Hessian matrix; b. Use Adam optimizer to update model parameters.
7. The method according to claim 1, characterized in that In step 4), the new two-dimensional pattern structure pattern is gridded and represented by density features.
8. The method according to claim 1, characterized in that In step 4), during prediction, the trained model runs on the GPU server.
9. A computer-readable storage medium having instructions stored thereon, wherein the instructions, when executed, implement the method according to any one of claims 1 to 4.
