CMOS intensive model parameter extraction method based on residual network model
Through the parameter extraction method based on the residual network model, the Monte Carlo method is used to generate parameter sets and train the model, which solves the problems of complex, slow speed and low accuracy of manual fitting, and realizes the fast and accurate extraction of CMOS intensive model parameters. , suitable for large-scale integrated circuit design.
Patent Information
- Application Number
- CN202411088409.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-12
AI Technical Summary
In large-scale integrated circuit design, manual fitting method extraction of CMOS intensive models is complex, slow, and low precision, which cannot meet automation requirements.
The parameter extraction method based on the residual network model is adopted, and the parameter set is generated using the Monte Carlo method for SPICE simulation, and the residual network model is trained to achieve fast and accurate parameter extraction.
The parameter extraction process is simplified, speed and accuracy are improved, and the experience requirements for operators are reduced. It is suitable for large-scale integrated circuit design and shortened the R&D cycle.
Smart Images

Figure CN118940691A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of integrated circuits, and in particular relates to a CMOS intensive model parameter extraction method based on a residual network (Residual Network, ResNet) model. Background Art
[0002] In the process of designing ultra-large-scale integrated circuit chips using EDA software, it is often necessary to use a compact model of transistors. It has the advantages of strong operability and fast simulation speed. It is a bridge from single transistor to large-scale integrated circuit design and plays a vital role in the design and verification of large-scale integrated circuits. Before simulation, it is necessary to generate a compact model for the target device. At this time, it is necessary to extract some key parameters corresponding to the device in the compact model. At present, parameter extraction is mainly achieved through manual fitting. In manual fitting, it is necessary to manually adjust the values of key parameters of the compact model, run SPICE simulation, compare the error between the electrical characteristic curve output by SPICE simulation and the target electrical characteristic curve obtained by experiment, and then change the parameter value according to the existing experience and error, and repeat this process to achieve parameter extraction. Manual fitting is not only very complicated in operation and has high requirements on the operator's experience, but also has slow extraction speed and low accuracy, which cannot meet the automation requirements of large-scale integrated circuit design.
[0003] The residual network model is a deep learning model. By introducing residual blocks and adding cross-layer connections in the training process of the deep learning model, it can effectively solve the gradient vanishing and gradient exploding problems in the deep neural network training process. The present invention proposes a CMOS intensive model parameter extraction method based on the residual network model. The Monte Carlo method can be used to generate a parameter set, and then a model training data set can be obtained through SPICE simulation for training and verifying the residual network model. The trained residual network model can realize fast and accurate CMOS transistor intensive model parameter extraction, overcoming the defects of complex manual fitting and extraction parameter operation, low speed and accuracy. Summary of the invention
[0004] The purpose of the present invention is to overcome the defects of complex operation, low speed and precision in extracting intensive model parameters by manual fitting. The present invention constructs a CMOS intensive model parameter extraction method based on a residual network model, generates a sufficient parameter set by using the Monte Carlo method for SPICE simulation, converts the corresponding relationship between the parameter value and the IV curve into a data set that can be used for model training, and finally trains the residual network model. The intensive model parameters corresponding to the experimental data or TCAD simulation data can be directly extracted according to the residual network model. After completing the construction of the residual network model, the method can realize fast and accurate CMOS transistor intensive model parameter extraction through the residual network model, effectively solve the operational problems in manual fitting, help predict the yield of integrated circuit products and shorten the research and development cycle, and provide assistance for the development and improvement of EDA software.
[0005] To achieve the above object, the technical solution adopted by the present invention is: S1. Generate CMOS intensive model parameter set using Monte Carlo method: Determine the key parameters to be extracted and the extraction order in the intensive model, and adjust the key parameter values by manual fitting so that the error between the electrical characteristic curve obtained by SPICE simulation and the target electrical characteristic curve obtained from TCAD simulation or transistor experimental data is controlled within a small range. After determining the key parameter values, use the Monte Carlo method to generate a large number of new parameter sets around the determined values.
[0006] S2. Obtain the training data set required for the residual network model: Use SPICE software to import the new parameter set generated in S1 into the intensive model for simulation, and obtain the electrical characteristic curve of the intensive model corresponding to each set of new parameters. Use the electrical characteristic curve data as the output layer of the residual network model, and the parameter group corresponding to the characteristic curve as the input layer of the residual network model to obtain the data set for model training.
[0007] S3. Training the residual network model: Divide the model training data set obtained in S2 into two parts, with the majority of the data set used as the training set and the minority of the data set used as the validation set. After using the training set to train the residual network model, the validation set is used to evaluate the model training effect.
[0008] S4. Extracting intensive model parameters using the residual network model: After the model is trained, the IV electrical characteristic curve obtained from TCAD simulation or transistor experimental data is used as the input of the residual network model. The model can extract the intensive model parameters corresponding to the electrical characteristic curve in a very short time.
[0009] Furthermore, in step S1, the intensive model is any intensive model that complies with industrial standards, such as BSIM, HiSIM, and PSP.
[0010] Further, in step S1, the electrical characteristic curve obtained by the SPICE simulation is compared with the target electrical characteristic curve obtained from the TCAD simulation or transistor experimental data, including the I of the transistor at high drain voltage (High Drain) and low drain voltage (Low Drain). D -V G and I D -V D Two types of electrical characteristic curves. The error calculation method between the electrical characteristic curve obtained by SPICE simulation and the target electrical characteristic curve is: (error under high drain voltage + error under low drain voltage) × 0.5.
[0011] Preferably, the error limit range between the electrical characteristic curve obtained by SPICE simulation and the target electrical characteristic curve obtained from TCAD simulation or transistor experimental data is a specific value less than 5%, the value of which is determined according to the trade-off between speed and accuracy when extracting intensive model parameters and can be adjusted according to actual conditions.
[0012] Furthermore, in step S1, when generating a large number of new parameter sets, for each key parameter, a range needs to be determined around its value, for example: ±50%, and a large number of parameter values are generated within this range using the Monte Carlo method, thereby forming a large number of new parameter sets. The purpose is to improve the extraction accuracy and ensure that the residual network model can be widely applied to device data containing statistical fluctuations.
[0013] Preferably, when the Monte Carlo method generates a large number of new parameter sets, the fluctuation range around the key parameter values is determined according to factors such as the value range of each key parameter, the ability to influence the electrical characteristics of the device, and the degree of fluctuation due to statistical fluctuations.
[0014] Furthermore, in step S1, the selection of key parameters adopts the principle of "few and simple", and selects some important physical parameters and a small number of fitting parameters for adjusting the simulation results from the parameters that have a crucial impact on the electrical characteristics of the device. By changing several key parameters, the output performance of the entire device can be affected, so that the electrical characteristic curve of the intensive model obtained by SPICE simulation fits the target electrical characteristic curve, and the selected parameters are extracted.
[0015] Further, in step S2, the simulation software used in the extraction process is any SPICE simulation software.
[0016] Further, in step S3, the ratio of the training set to the validation set is any ratio that avoids overfitting and underfitting.
[0017] The present invention has at least the following beneficial technical effects: 1. Fast and accurate CMOS transistor intensive model parameter extraction through residual network model.
[0018] The present invention realizes the process of extracting intensive model parameters based on the residual network model, avoiding the complex steps of cyclically adjusting parameters and simulating and comparing the fitting degree to realize parameter extraction in the traditional mode. After the model is established, TCAD data or transistor experimental data is input, and the corresponding extracted intensive model parameters can be output immediately.
[0019] 2. The residual network model is portable and the parameter extraction process has low operator requirements.
[0020] After the residual network model is built, the model is portable. To apply the trained residual network model, you only need to input the experimental data, which avoids the problem of poor extraction effect due to insufficient operator experience.
[0021] 3. The residual network model is more optimized than other deep learning models.
[0022] The residual network model can effectively solve the problems of gradient disappearance and gradient explosion in classic deep learning models. It can also solve the network degradation effect that occurs as the number of hidden layers in deep learning models increases by connecting the hidden layers through the "residual blocks" in the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 This is an example of a parameter set generated by the Monte Carlo method of the present invention. Figure 2 Flowchart for implementing the present invention DETAILED DESCRIPTION
[0024] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with the accompanying drawings.
[0025] The parameter extraction method of the present invention is implemented by the following steps: S1. Generate CMOS intensive model parameter set using Monte Carlo method: Determine the key parameters to be extracted and the extraction order in the intensive model, and adjust the key parameter values by manual fitting so that the error between the electrical characteristic curve obtained by SPICE simulation and the target electrical characteristic curve obtained from TCAD simulation or transistor experimental data is controlled within a small range, generally within 5%. After determining the key parameter values, use the Monte Carlo method to generate a large number of new parameter sets around the determined values.
[0026] S2. Obtain the training data set required for the residual network model: Use SPICE software to import the new parameter set generated in S1 into the intensive model for simulation, and obtain the electrical characteristic curve of the intensive model corresponding to each set of new parameters. Use the electrical characteristic curve data as the output layer of the residual network model, and the parameter group corresponding to the characteristic curve as the input layer of the residual network model to obtain the data set for model training.
[0027] S3. Training the residual network model: Divide the model training data set obtained in S2 into two parts, with the majority of the data set used as the training set and the minority of the data set used as the validation set. After using the training set to train the residual network model, the validation set is used to evaluate the model training effect.
[0028] S4. Extracting intensive model parameters using the residual network model: After the model is trained, the IV electrical characteristic curve obtained from TCAD simulation or transistor experimental data is used as the input of the residual network model. The model can extract the intensive model parameters corresponding to the electrical characteristic curve in a very short time.
[0029] In the step S1, the intensive model is any intensive model that complies with industrial standards, such as BSIM, HiSIM, and PSP.
[0030] In step S1, the electrical characteristic curve obtained by the SPICE simulation is compared with the target electrical characteristic curve obtained from the TCAD simulation or transistor experimental data, including the I D -V G and I D -V D Two types of electrical characteristic curves. The error calculation method between the electrical characteristic curve obtained by SPICE simulation and the target electrical characteristic curve is: (error under high drain voltage + error under low drain voltage) × 0.5.
[0031] Preferably, the error limit range between the electrical characteristic curve obtained by SPICE simulation and the target electrical characteristic curve obtained from TCAD simulation or transistor experimental data is a specific value less than 5%, the value of which is determined according to the trade-off between speed and accuracy when extracting intensive model parameters and can be adjusted according to actual conditions.
[0032] In step S1, when generating a large number of new parameter sets, for each key parameter, a range needs to be determined around its value, for example: ±50%, and a large number of parameter values are generated within this range using the Monte Carlo method, thereby forming a large number of new parameter sets. The purpose is to improve the extraction accuracy and ensure that the residual network model can be widely applied to device data containing statistical fluctuations.
[0033] Preferably, when the Monte Carlo method generates a large number of new parameter sets, the fluctuation range around the key parameter values is determined according to factors such as the value range of each key parameter, the ability to influence the electrical characteristics of the device, and the degree of fluctuation due to statistical fluctuations.
[0034] In step S1, the selection of key parameters adopts the principle of "few and simple", and selects some important physical parameters and a small number of fitting parameters for adjusting the simulation results from the parameters that have a crucial impact on the electrical characteristics of the device. By changing several key parameters, the output performance of the entire device can be affected, so that the electrical characteristic curve of the intensive model obtained by SPICE simulation fits the target electrical characteristic curve, and the selected parameters are extracted.
[0035] In step S2, the simulation software used in the extraction process is any SPICE simulation software.
[0036] In step S3, the ratio of the training set to the validation set is any ratio that avoids overfitting and underfitting.
Claims
1. A CMOS intensive model parameter extraction method based on a residual network model, characterized in that: Follow the steps below to achieve this: S1. Generate CMOS intensive model parameter set using Monte Carlo method: Determine the key parameters to be extracted and the extraction order in the intensive model, adjust the key parameter values by manual fitting, so that the error between the electrical characteristic curve obtained by SPICE simulation and the target electrical characteristic curve obtained from TCAD simulation or transistor experimental data is controlled within 3%. After determining the key parameter values, use the Monte Carlo method to generate a large number of new parameter sets around the determined values; S2. Obtain the training data set required for the residual network model: The new parameter set generated in S1 is imported into the intensive model, and the SPICE software is used for simulation to obtain the electrical characteristic curve of the intensive model corresponding to each set of new parameters; the electrical characteristic curve data is used as the output layer of the residual network model, and the parameter group corresponding to the characteristic curve is used as the input layer of the residual network model to obtain the data set for model training; S3. Training residual network model: The model training data set obtained in S2 is divided into two parts, where most of the data set is used as the training set and a small part of the data set is used as the validation set; after the residual network model is trained with the training set, the model training effect is evaluated with the validation set; S4. Extracting intensive model parameters using residual network model: After the model is trained, the IV electrical characteristic curve obtained from TCAD simulation or transistor experimental data is used as the input of the residual network model. The model can extract the intensive model parameters corresponding to the electrical characteristic curve in a very short time.
2. The CMOS intensive model parameter extraction method based on the residual network model as claimed in claim 1, characterized in that: In step S1, the intensive model may be any one of the intensive models of industrial standards such as BSIM, HiSIM, PSP, etc.
3. The CMOS intensive model parameter extraction method based on the residual network model according to claim 1, characterized in that: In step S1, the electrical characteristic curve obtained by the SPICE simulation is compared with the target electrical characteristic curve obtained from the TCAD simulation or transistor experimental data, including the I of the transistor at high drain voltage (High Drain) and low drain voltage (Low Drain). D -V G and I D -V D Two types of electrical characteristic curves.
4. The CMOS intensive model parameter extraction method based on the residual network model as claimed in claim 3, characterized in that: The error limit range between the electrical characteristic curve obtained by SPICE simulation and the target electrical characteristic curve obtained from TCAD simulation or transistor experimental data is a specific value less than 5%. The value is determined according to the trade-off between speed and accuracy when extracting intensive model parameters and can be adjusted according to actual conditions.
5. The CMOS intensive model parameter extraction method based on the residual network model as claimed in claim 1, characterized in that: In step S1, when a large number of new parameter sets are generated, for each key parameter, a range needs to be determined around its value, and a large number of parameter values are generated within the range using the Monte Carlo method, thereby forming a large number of new parameter sets.
6. The CMOS intensive model parameter extraction method based on the residual network model as claimed in claim 1, characterized in that: In step S1, the key parameters are selected based on the principle of "few and simple". Some important physical parameters and a small number of fitting parameters for adjusting the simulation results are selected from the parameters that have a crucial impact on the electrical characteristics of the device. By changing several key parameters, the output performance of the entire device can be affected, so that the electrical characteristic curve of the intensive model obtained by SPICE simulation fits the target electrical characteristic curve, thereby realizing the extraction of the selected parameters.
7. The CMOS intensive model parameter extraction method based on the residual network model as claimed in claim 1, characterized in that: In step S2, the simulation software used in the extraction process is any SPICE simulation software.
8. The CMOS intensive model parameter extraction method based on the residual network model as claimed in claim 1, characterized in that: In step S3, the ratio of the training set to the validation set is any ratio that avoids overfitting and underfitting.
Citation Information
Cited By
BSIM-BULK model parameter extraction method based on neural network
CN119830708A
A neural network-based BSIM-BULK model parameter extraction method
CN119830708B
