A device model parameter automatic extraction method based on residual multilayer perception
By using a residual multilayer perceptron-based method, device model parameters are automatically extracted, solving the problem of relying on human experience in existing technologies. This achieves efficient and accurate extraction of device model parameters across the entire size range, improving the stability and consistency of electrical behavior and shortening the development cycle.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- EAST CHINA NORMAL UNIV
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-31
AI Technical Summary
Existing methods for extracting device model parameters rely on human experience, which is inefficient and has limited ability to extract models across the entire size range, making it difficult to guarantee the accuracy and consistency of electrical behavior.
A residual multilayer perceptron-based approach is adopted. By constructing a nonlinear mapping relationship between device model parameters and electrical characteristics, and combining forward prediction and backward optimization, device model parameters are automatically extracted. The Latin hypercube sampling method is used to generate training set samples, and a residual block structure and a comprehensive loss function are introduced to improve training stability and fitting accuracy.
It reduces reliance on human experience, improves the efficiency and accuracy of parameter optimization, ensures consistent and accurate electrical behavior across the entire size range, and shortens the PDK development cycle.
Smart Images
Figure CN122491028A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semiconductor device modeling and parameter extraction technology, specifically a method for automatic extraction of device model parameters based on residual multilayer perceptron. Background Technology
[0002] In the Device Design and Process Co-optimization (DTCO) workflow, device model parameter extraction occurs after device optimization and before circuit simulation, serving as a crucial step in transitioning device characteristics to circuit-level applications. Its core task is to optimize the undetermined parameters in the model based on the device's electrical characteristic curves and relevant simulation or test results, obtaining a set of model parameters that accurately reflect the device's electrical behavior. The accuracy of the model fitting is critical to the accuracy of subsequent circuit simulation results.
[0003] However, existing methods for extracting device model parameters still rely heavily on human experience. In practice, engineers often need to pre-define parameter ranges, select optimization paths, and continuously adjust parameter settings through multiple rounds of fitting and result comparison. This approach not only has a long development cycle but also, when the parameter dimensionality is high and the coupling relationships between parameters are complex, is easily limited by the selection of initial values and local optima, thus affecting the efficiency and stability of parameter extraction.
[0004] Furthermore, many existing research methods only focus on model parameter fitting for devices of a single size, primarily optimizing electrical characteristics under a fixed size, while neglecting the uniform applicability of the model across different size ranges. However, in actual PDK development, the more challenging and time-consuming task is not local fitting for a single-size device, but rather global parameter extraction across the entire size range. This process requires ensuring the consistency and accuracy of the device's electrical behavior description under different channel lengths, channel widths, and other structural dimensions.
[0005] Therefore, how to reduce the reliance on engineers' experience in the parameter extraction process, reduce the time cost of manual iteration, and improve the efficiency of global modeling across the entire size range has become a key issue that urgently needs to be addressed in the field of device model parameter extraction. Summary of the Invention
[0006] The purpose of this invention is to address the problems of existing technologies, such as the high dependence on human experience in device model parameter extraction, low optimization efficiency, and limited model extraction capabilities across the entire size range. This invention proposes an automatic device model parameter extraction method based on a residual multilayer perceptron. This method automatically extracts device model parameters after the device optimization step and before circuit simulation. By constructing a nonlinear mapping relationship between device model parameters and electrical characteristics, and combining forward prediction and backward optimization processes, it achieves rapid and accurate extraction of device model parameters. This reduces the reliance on human experience in the parameter extraction process, improves the efficiency, accuracy, and stability of parameter optimization, and enhances the level of automation. This provides a more reliable model foundation for subsequent circuit simulation and performance evaluation, further shortening the PDK development cycle and demonstrating promising application prospects.
[0007] The specific technical solution for achieving the objective of this invention is as follows:
[0008] An automatic parameter extraction method for device models based on residual multilayer perceptrons, the method comprising the following specific steps:
[0009] Step 1: Determine the type of CV and IV characteristic curves to be fitted under multiple bias conditions and the voltage sweep range;
[0010] Step 2: Determine the device model parameters to be extracted, and select the parameter range within the range specified in the model manual to ensure the model fitting effect;
[0011] Step 3: Use the Latin hypercube sampling method to generate a uniformly distributed combination of device model parameters in the parameter space; run the required simulation netlist file to obtain the corresponding CV and IV characteristic curves; combine the two and use them as training set samples for subsequent neural network training.
[0012] Step 4: Train the residual multilayer perceptron using the device model parameter combination as network input and the CV and IV characteristic curves as network output;
[0013] Step 5: Backfill the measured data or simulation reference data of the fitted device into the output of the neural network, fix the parameters in the trained neural network, and fix the device process size related parameters in the network input according to the actual process conditions. Use reverse optimization to optimize the remaining device model parameters so that the final simulation results are close to the measured characteristics; finally, obtain the model parameter file corresponding to the device.
[0014] Furthermore, step 4 specifically includes:
[0015] 4-1: By introducing a cascaded residual block structure into a conventional multilayer perceptron, the network maintains its overall feedforward architecture while adding shortcut connections between input features and subsequent layer outputs. Compared to traditional neural networks, this structure provides a more direct propagation path for error backpropagation, effectively alleviating the gradient decay and optimization difficulties that easily occur during deep network training, thus improving the stability and convergence performance of model training. Simultaneously, the residual blocks do not learn the target mapping. It is not itself, but the corresponding residual term. ;
[0016] in, This represents the input feature vector of the residual block; Represents the expected output learned by the residual block; represents the residual function. This describes the incremental result of input features after passing through several fully connected layers, normalization layers, and nonlinear activation functions. Compared to directly fitting complex mapping relationships, learning the residual term... This structure typically has lower optimization difficulty, thus it can improve the network's ability to express complex nonlinear relationships and its fitting effect while being relatively simple to implement.
[0017] 4-2: To accurately characterize the various electrical characteristics of the device under different bias conditions, a comprehensive loss function is constructed based on the fitting characteristics of different types of electrical characteristic curves; the comprehensive loss function simultaneously considers two types of electrical characteristic curves: one is different drain-source voltages. Gate total capacitance curve under the condition Coupling capacitance from device gate to channel Secondly, they are different. Leakage current curve under certain conditions Specifically, for the IV curve, the loss is further calculated in both the linear and logarithmic coordinate domains to simultaneously consider the fitting accuracy of the device in both the strong inversion region and the subthreshold region; the loss functions are then defined as follows:
[0018]
[0019] in, , and These represent the fitting error of the IV curve in linear coordinates, the fitting error of the IV curve in logarithmic coordinates, and the fitting error of the CV curve in linear coordinates, respectively. This represents the output value of the neural network. Represents the measured value of the device; Indicates the first The number of sampling points for each IV curve. Indicates the first The number of sampling points for the CV curve; This indicates the number of device dimensions involved in the model extraction;
[0020] The total loss function can then be written as:
[0021]
[0022] in, , and These are the linear domain loss weights of the IV curve, the logarithmic domain loss weights of the IV curve, and the linear domain loss weights of the CV curve, respectively.
[0023] 4-3: Optimize the neural network structure based on residual multilayer perceptron; improve the configurability, training stability and representation ability of complex nonlinear characteristics of the model structure by opening up hyperparameters including hidden layer dimension, number of residual blocks, number of layers within a block, activation function and normalization method.
[0024] Furthermore, step 5 specifically includes:
[0025] 5-1: For a trained model, freeze the parameters of each layer;
[0026] 5-2: Based on the actual process conditions or structural dimensions of the device, fix the corresponding parameters at the input of the neural network; initialize the parameter combinations of the remaining device models in the parameter space;
[0027] 5-3: Compare the network output obtained by forward propagation of the parameter combination with the measured data of the device to obtain the fitting error. Then, use the fitting error to back-optimize the parameter combination at the network input. The fitting error includes the linear domain loss of the IV curve, the logarithmic domain loss of the IV curve, the linear domain loss of the CV curve, and the soft constraint introduced for the fitting error of the device's electrical characteristics, which is the fitting error of the threshold voltage and on-state current in the linear transition curve and the saturation transition curve. According to the fitting requirements of the device's electrical characteristics, additional penalty terms are applied for cases where the threshold voltage error exceeds 10mV and the on-state current error exceeds 3%, specifically described by the following formula:
[0028]
[0029] in, This indicates an additional penalty item. This indicates the number of device sizes involved in the fitting process. Indicates the first Each size device in parameter combination The threshold voltage fitting error is below. Indicates the first Each size device in parameter combination Current fitting error under the given conditions and This is the penalty coefficient.
[0030] Beneficial effects
[0031] 1) This invention combines a residual multilayer perceptron with a parameter back-optimization process to establish a complex nonlinear mapping relationship between device model parameters and electrical characteristics. Unlike some existing methods that directly use the device's electrical characteristic curve as the neural network input and the model parameters as the output, this invention adopts a modeling approach of "model parameter input, electrical characteristic output". This is because the mapping from device model parameters to electrical characteristic curves is usually closer to injective mapping, while the back-mapping from electrical characteristic curves to model parameters easily leads to multiple solutions for multiple sets of parameters corresponding to similar curve features, which is detrimental to stable network training and improved prediction accuracy.
[0032] 2) Compared to traditional Bayesian optimization methods, this invention eliminates the need for frequent construction and updating of surrogate models in high-dimensional parameter spaces, avoiding point-by-point search and iterative updates. This results in higher search efficiency and better scalability, making it more suitable for parameter extraction tasks with high parameter dimensions and complex coupling relationships. Compared to conventional neural network structures, this invention improves network training stability and nonlinear representation capabilities by introducing residual connections and mechanisms such as normalization and regularization. It also mitigates degradation problems and overfitting risks in deep network training, making it more suitable for parameter extraction from device models across the entire size range. Therefore, while ensuring fitting accuracy, it further improves parameter extraction efficiency and reduces reliance on human experience.
[0033] 3) The loss function designed in this invention is used for target optimization and error constraint in the automatic extraction of device model parameters. This loss function simultaneously includes the CV curve fitting error, the linear domain fitting error of the IV curve, and the logarithmic domain fitting error of the IV curve, to achieve a comprehensive evaluation of the multi-dimensional electrical characteristics of the device; and it sets penalty terms for threshold voltage errors greater than 10 mV and current errors greater than 3%, thereby strengthening the constraints on key electrical characteristic indicators. With the above-mentioned loss function design, the accuracy, stability, and size applicability of device model parameter extraction can be effectively improved, and a more reliable model foundation can be provided for subsequent circuit simulation. Attached Figure Description
[0034] Figure 1 This is a flowchart of the present invention;
[0035] Figure 2 This is a schematic diagram of CV curve fitting when performing instance fitting using the present invention;
[0036] Figure 3 When performing instance fitting using this invention Under the conditions Schematic diagram of curve fitting;
[0037] Figure 4 When performing instance fitting using this invention Under the conditions Schematic diagram of curve fitting. Detailed Implementation
[0038] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0039] See Figure 1 The present invention provides an automatic parameter extraction method for device models based on residual multilayer perceptrons, which specifically includes the following steps:
[0040] 1) Determine the type of CV and IV characteristic curves to be fitted under multiple bias conditions and the voltage sweep range;
[0041] 2) Determine the device model parameters to be extracted, and select the parameter range within the range specified in the model manual to ensure the model fitting effect;
[0042] 3) Generate a uniformly distributed combination of device model parameters in the parameter space using the Latin hypercube sampling method; run the required simulation netlist file to obtain the corresponding CV and IV characteristic curves. The combined netlist and IV curves are then used as the training set samples for subsequent neural network training.
[0043] 4) Train the residual multilayer perceptron using parameter combinations as network input and CV and IV characteristic curves as network output, specifically including:
[0044] 4.1: By introducing a cascaded residual block structure into a conventional multilayer perceptron, the network maintains its overall feedforward architecture while adding shortcut connections between input features and subsequent layer outputs. Compared to traditional neural networks, this structure provides a more direct propagation path for error backpropagation, effectively alleviating the gradient decay and optimization difficulties that easily occur during deep network training, and improving the stability and convergence performance of model training. Meanwhile, the residual blocks do not learn the target mapping. It is not itself, but the corresponding residual term. Compared to directly fitting complex mapping relationships, learning residual terms is usually easier to optimize. Therefore, this structure can improve the network's ability to express complex nonlinear relationships and its fitting effect while being relatively simple to implement.
[0045] 4.2: To accurately characterize the various electrical characteristics of the device under different bias conditions, this scheme constructs a comprehensive loss function based on the fitting characteristics of different types of electrical characteristic curves. The comprehensive loss function simultaneously considers two types of electrical characteristic curves: one is different drain-source voltages. Gate total capacitance curve under the condition Coupling capacitance from device gate to channel Secondly, they are different. Leakage current curve under certain conditions Specifically, for the IV curve, the loss is further calculated in both the linear and logarithmic coordinate domains to simultaneously consider the fitting accuracy of the device in both the strong inversion region and the subthreshold region; the loss functions are then defined as follows:
[0046]
[0047] in, , and These represent the fitting error of the IV curve in linear coordinates, the fitting error of the IV curve in logarithmic coordinates, and the fitting error of the CV curve in linear coordinates, respectively. This represents the output value of the neural network. Represents the measured value of the device; Indicates the first The number of sampling points for each IV curve. Indicates the first The number of sampling points for the CV curve; This indicates the number of device dimensions involved in the model extraction;
[0048] The total loss function can then be written as:
[0049]
[0050] in, , and These are the linear domain loss weights of the IV curve, the logarithmic domain loss weights of the IV curve, and the linear domain loss weights of the CV curve, respectively.
[0051] 4.3: The neural network structure based on residual multilayer perceptron was optimized. By adjusting key hyperparameters such as the hidden layer dimension, the number of residual blocks, the number of layers within a block, the activation function, and the normalization method, the configurability, training stability, and ability to represent complex nonlinear characteristics of the model structure were improved. This provides a more universal and scalable foundation for subsequent high-precision device modeling and parameter extraction.
[0052] 5) Backfill the output of the neural network with the measured or simulation reference data of the device to be fitted, fix the parameters in the trained neural network, and fix the device process dimension-related parameters in the network input according to the actual process conditions. Use backpropagation to optimize the remaining device model parameters so that the final simulation results are as close as possible to the measured characteristics. Finally, the model parameter file corresponding to the device is obtained. Specific steps include:
[0053] 5.1: For a trained model, freeze the parameters of each layer;
[0054] 5.2: Based on the actual process conditions or structural dimensions of the device, fix the size-related parameters at the neural network input. Initialize the parameter combinations of the remaining device models in the parameter space;
[0055] 5.3: The network output obtained by forward propagation of the parameter combination is compared with the measured device data to obtain the fitting error. This error is then used to back-optimize the parameter combination at the network input. In addition to the linear domain loss, logarithmic domain loss, and linear domain loss of the IV curve mentioned in 4.3, the fitting error also includes an additional soft constraint on the fitting error of the device's electrical characteristics. This mainly consists of the fitting errors of the threshold voltage and on-state current in the linear transition curve and saturation transition curve. Based on industry requirements for fitting the device's electrical characteristics, an additional penalty term is applied for threshold voltage errors exceeding 10mV and on-state current errors exceeding 3%, specifically described by the following formula:
[0056]
[0057] in, This indicates an additional penalty item. This indicates the number of device sizes involved in the fitting process. Indicates the first Each size device in parameter combination The threshold voltage fitting error is below. Indicates the first Each size device in parameter combination Current fitting error under the given conditions and This is the penalty coefficient.
[0058] Example: This example uses the parameter extraction process of a device model at a 40nm process node as an example. First, CV and IV characteristic curve data are collected for devices with a width of 0.5μm and channel lengths of 0.5 / 0.2 / 0.1 / 0.07μm. Then, more than 30 commonly used device model parameters are selected as parameters to be optimized, a parameter search space is constructed, and a residual multilayer perceptron is used to learn the mapping relationship between the model parameters and the device electrical characteristic curves. Based on the trained model, the device model parameters, which are used as inputs to the neural network, are optimized using backpropagation. During the optimization process, the IV and CV characteristic curves generated by the model under different parameter combinations are compared with the target data, the corresponding fitting errors are calculated, and a loss function is constructed by combining a set penalty term to evaluate the merits of the current parameter combination. After 8000 iterations of optimization, a set of device model parameters with good fitting effect is obtained, enabling the model to accurately characterize the IV and CV characteristics of the device, thus providing a reliable model foundation for subsequent circuit simulation and performance evaluation.
[0059] Figure 2 The image shown is extracted using the present invention. and The fitting results are shown in diagrams (a) and (b) in the figure, respectively. It can be seen that the fitting error is within the 10% standard. Figure 3 The demonstration shows devices with different channel lengths in Under the conditions The fitting results are shown in the diagrams (a) to (d) in the figure, indicating that the fitting error of the linear transfer characteristics is within the standard of 3%. Figure 4 The demonstration shows devices with different channel lengths in Under the conditions The fitting results are shown in the diagrams (a) to (d) in the figure, indicating that the fitting error of the linear transfer characteristics is within the 5% standard.
Claims
1. A method for automatically extracting device model parameters based on a residual multilayer perceptron, characterized in that, The method includes the following specific steps: Step 1: Determine the type of CV and IV characteristic curves to be fitted under multiple bias conditions and the voltage sweep range; Step 2: Determine the device model parameters to be extracted, and select the parameter range within the range specified in the model manual to ensure the model fitting effect; Step 3: Use the Latin hypercube sampling method to generate a uniformly distributed combination of device model parameters in the parameter space; run the required simulation netlist file to obtain the corresponding CV and IV characteristic curves; combine the two and use them as training set samples for subsequent neural network training. Step 4: Train the residual multilayer perceptron using the device model parameter combination as network input and the CV and IV characteristic curves as network output; Step 5: Backfill the measured data or simulation reference data of the fitted device into the output of the neural network, fix the parameters in the trained neural network, and fix the device process size related parameters in the network input according to the actual process conditions. Use reverse optimization to optimize the remaining device model parameters so that the final simulation results are close to the measured characteristics; finally, obtain the model parameter file corresponding to the device.
2. The device model parameter automatic extraction method according to claim 1, characterized in that, Step 4 specifically includes: 4-1: Introduce a residual block structure in series in a common multi-layer perceptron, the residual block learns the corresponding residual term ; wherein, represents an input feature vector of the residual block; represents an output expected to be learned by the residual block; represents a residual function, is used to describe the incremental result of the input feature after passing through several fully connected layers, normalization layers, and nonlinear activation functions; 4-2: Construct a comprehensive loss function based on the fitting characteristics of different types of electrical characteristic curves; the comprehensive loss function simultaneously considers two types of electrical characteristic curves: one is different drain-source voltages. Gate total capacitance curve under the condition Coupling capacitance from device gate to channel Secondly, they are different. Leakage current curve under certain conditions Specifically, for the IV curve, the loss is further calculated in both the linear and logarithmic coordinate domains to simultaneously consider the fitting accuracy of the device in both the strong inversion region and the subthreshold region; the loss functions are then defined as follows: ； in, and These represent the fitting error of the IV curve in linear coordinates, the fitting error of the IV curve in logarithmic coordinates, and the fitting error of the CV curve in linear coordinates, respectively. This represents the output value of the neural network. Represents the measured value of the device; Indicates the first The number of sampling points for each IV curve Indicates the first The number of sampling points for the CV curve; This indicates the number of device dimensions involved in the model extraction; The total loss function can then be written as: ； in, , and These are the linear domain loss weights of the IV curve, the logarithmic domain loss weights of the IV curve, and the linear domain loss weights of the CV curve, respectively. 4-3: Optimize the neural network structure based on residual multilayer perceptron; improve the configurability, training stability and representation ability of complex nonlinear characteristics of the model structure by opening up hyperparameters including hidden layer dimension, number of residual blocks, number of layers within a block, activation function and normalization method.
3. The automatic extraction method for device model parameters according to claim 1, characterized in that, Step 5 specifically includes: 5-1: For a trained model, freeze the parameters of each layer; 5-2: Based on the actual process conditions or structural dimensions of the device, fix the corresponding parameters at the input of the neural network; initialize the parameter combinations of the remaining device models in the parameter space; 5-3: Compare the network output obtained by forward propagation of the parameter combination with the measured data of the device to obtain the fitting error. Then, use the fitting error to back-optimize the parameter combination at the network input. The fitting error includes the linear domain loss of the IV curve, the logarithmic domain loss of the IV curve, the linear domain loss of the CV curve, and the soft constraint introduced for the fitting error of the device's electrical characteristics, which is the fitting error of the threshold voltage and on-state current in the linear transition curve and the saturation transition curve. According to the fitting requirements of the device's electrical characteristics, additional penalty terms are applied for cases where the threshold voltage error exceeds 10mV and the on-state current error exceeds 3%, specifically described by the following formula: ； in, This indicates an additional penalty item. This indicates the number of device sizes involved in the fitting process. Indicates the first Each size device in parameter combination The threshold voltage fitting error is below. Indicates the first Each size device in parameter combination Current fitting error under the given conditions and This is the penalty coefficient.