A device model parameter automatic extraction method based on a hybrid algorithm

Abstract

The present application belongs to the technical field of compact model parameter extraction, and specifically relates to a device model parameter automatic extraction method based on a hybrid algorithm.The present application combines a HiMOSS algorithm and a reinforcement learning algorithm, first uses the HiMOSS algorithm to perform preliminary exploration and positioning in a parameter search space, and then uses the reinforcement learning algorithm to perform further optimization and parameter extraction.The present application can realize automatic extraction of compact model parameters, and can be used for accurately and efficiently creating a compact model in analog circuit design.

Description

Technical Field

[0001] This invention belongs to the field of compact model parameter extraction technology, specifically, it relates to an automatic extraction method for device model parameters based on a hybrid algorithm. Background Technology

[0002] With the rapid development of the integrated circuit field, the feature size of metal-oxide-semiconductor (MOSFET), a key device used in integrated circuits, has entered the nanometer scale, and the physical effects of these devices are becoming increasingly complex. Compact models are used to simulate the electrical characteristics of these devices. To achieve more accurate modeling of the corresponding physical devices, the number of parameters contained in compact models has gradually increased to hundreds or even thousands, making traditional manual parameter extraction based on engineer experience difficult. Accurate and efficient parameter extraction from compact models has become a crucial step in integrated circuit design, and the development of automated parameter extraction methods is an inevitable trend. Currently, algorithms such as genetic algorithms, difference algorithms, Bayesian optimization algorithms, neural networks, and reinforcement learning have been applied to parameter extraction. However, these solutions still suffer from limitations in general applicability and difficulty in handling high-dimensional challenges. Therefore, new breakthroughs are still needed in automated parameter extraction methods. Summary of the Invention

[0003] This invention aims to provide an automatic extraction method for device model parameters based on a hybrid algorithm to solve the problem of accurately and efficiently extracting device model parameters in the parameter space, thereby assisting in analog circuit design. The automatic extraction method for device model parameters based on the HiRL hybrid algorithm proposed in this invention has faster convergence and better performance.

[0004] The technical solution of the present invention is described in detail below.

[0005] This invention provides an automatic parameter extraction method for device models based on a hybrid algorithm. It combines the HiMOSS algorithm and reinforcement learning algorithm to simulate the manual parameter extraction process, achieving a synergy between global exploration and local optimization in the parameter search space. The method outputs the optimal extracted model parameters to achieve fitting of the device's electrical characteristics. The specific steps are as follows:

[0006] (1) Collect the electrical characteristic curves of the target;

[0007] (2) Select a compact model based on the device characteristics to simulate the electrical characteristics of the device, and define the range of values ​​for the parameters to be extracted in the compact model.

[0008] (3) Use the HiMOSS algorithm to conduct preliminary exploration and localization in the parameter search space to generate a new range of model parameters.

[0009] The parameter extraction problem is described as a mathematical problem of maximizing the objective function R(x). The objective function R(x) is defined, and the algorithm optimizes by maximizing R(x). In each optimization cycle, sampling points (i.e., model parameters x) are selected within the initially defined range of model parameter values ​​according to the subspace sampling strategy. The corresponding reward value is calculated using the objective function R(x), and both the sample and the corresponding reward are recorded. Then, the subspace is updated. After repeating this process for several cycles, the loop ends. Based on the recorded sampling data, a new range of model parameter values ​​is generated.

[0010] (4) Use reinforcement learning algorithms to optimize and extract parameters within the new model parameter range.

[0011] Within each optimization cycle, first observe the environment and confirm the state s at that moment. t According to state s t The action a that determines the output at this moment t This refers to a compact model parameter sequence; subsequently, action a is calculated using the objective function. t The corresponding reward r t and obtain the action a to be executed. t The new state s t+1 This will include the state, action, reward, and state of the next moment. t ,a t ,r t ,s t+1 Add them together to the experience replay pool;

[0012] After optimization, the optimal compact model parameters are output to achieve fitting of the electrical characteristics of the target device.

[0013] In this invention, in step (1), electrical characteristic curves are acquired by simulating with TCAD software or by testing real devices; the electrical characteristic curves include IV characteristic curves and CV characteristic curves.

[0014] In this invention, in step (2), the compact model is either the BSIMCMG compact model or the BSIMSOI compact model.

[0015] In this invention, in step (3), the initial settings of the HiMOSS algorithm are first performed, defining the initial range of model parameter x, the maximum number of samplings T, and the hyperparameter β used to control the range of newly generated model parameters; then the optimization cycle is cyclical.

[0016] In this invention, in step (3), the parameter extraction problem is as follows: search for suitable compact model parameters so that when these model parameters are applied to the compact model, the electrical characteristic curve expressed by the compact model is as close as possible to the desired electrical characteristic curve.

[0017] In this invention, in steps (3) and (4), when there are M data points to be fitted, the expression of the objective function R(x) is as follows:

[0018]

[0019] Where x represents the current model parameter values, H(i,x) represents the squared error between the fitted value and the target value for the i-th data point, and w i The weight assigned to the i-th data point.

[0020] In this invention, in step (4), the agent of the reinforcement learning algorithm is an agent comprising an actor network and a critic network; the input of the actor network is state s. t The output is action a. t The input to the critic network is state s t and action a t The output is the action value function.

[0021] In this invention, in step (4), when the agent and the environment interact a predetermined number of times, the optimization ends, and the optimal compact model parameter is the reward r. t Action a corresponding to the maximum t .

[0022] The HiMOSS algorithm, as described above, is an improved Bayesian optimization algorithm that constructs a surrogate model of the objective function through a Gaussian process. It boasts high sampling efficiency and can quickly locate the potential optimal region of the model parameters. Reinforcement learning is a machine learning method parallel to supervised and unsupervised learning, where the agent learns through interaction with the environment. Its adaptability in high-dimensional spaces allows it to effectively capture complex state-action mappings. Compared to existing technologies, the advantages of this invention are:

[0023] 1) The hybrid algorithm HiRL proposed in this invention combines the advantages of HiMOSS algorithm and reinforcement learning algorithm, simulates the manual parameter extraction process, and achieves the synergy of global exploration and local optimization in the parameter search space, which can fit the electrical characteristic curve of the device more efficiently and accurately.

[0024] 2) This invention can realize the automatic extraction of compact model parameters of devices in a continuous parameter space.

[0025] 3) This invention selects the BSIMSOI model and the BSIMCMG model, and conducts simulation comparison experiments using TCAD simulation data and actual device measurement data. The results show that the proposed method outperforms other comparative algorithms, proving the effectiveness of this invention. This invention provides an effective method for the automation of compact model parameter extraction, which can promote the advancement of parameter extraction automation. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a compact model parameter automatic extraction framework based on the HiRL hybrid algorithm.

[0027] Figure 2 This is a performance comparison of the BSIMCMG model for extracting 100 parameters based on the HiRL algorithm, HiMOSS algorithm, reinforcement learning (RL) algorithm, and Bayesian optimization (BO) algorithm.

[0028] Figure 3 This is a performance comparison of the BSIMSOI model for extracting 104 parameters using the HiRL algorithm, HiMOSS algorithm, reinforcement learning algorithm, and Bayesian optimization algorithm. Detailed Implementation

[0029] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] This invention provides a method for automatically extracting device model parameters based on a hybrid algorithm, the specific steps of which are as follows:

[0031] (1) Use TCAD software to simulate or collect the target electrical characteristic curves by testing real devices;

[0032] (2) Select a suitable BSIM compact model based on the device characteristics, define the range of values ​​for the model parameters, and build a SPICE simulation environment;

[0033] (3) The parameter extraction problem is described as a mathematical problem of maximizing the objective function. The objective function R(x) is defined, and the hyperparameters in the HiMOSS and reinforcement learning algorithms are defined. The hyperparameters in the HiMOSS and reinforcement learning algorithms are used to control the range of model parameters and the optimal values ​​of the model parameters, respectively. When there are M data points to be fitted, the expression for R(x) is as follows:

[0034]

[0035] Where x represents the current model parameter value, H(i,x) represents the square of the relative error between the fitted value and the target value of the i-th data point, and is the weight assigned to the i-th data point;

[0036] (4) Initial optimization is performed using the HiMOSS algorithm. Within each optimization cycle, sampling points x are selected based on the surrogate model and the acquisition function. i Based on the objective function, calculate the selected sampling point x. i The corresponding reward value R is recorded in the dataset for sample x. i and reward value R;

[0037] (5) Generate a new range of model parameters based on the recorded sampling information and hyperparameter β;

[0038] (6) Optimize the new model parameters using reinforcement learning algorithms. During each optimization cycle, the agent interacts with the SPICE simulation environment, adjusting the current state s... t Select the action to perform, a. t Then, action a is calculated using the objective function. t The corresponding reward r t and obtain the action a to be executed. t The new state s t+1 . will (s t ,a t ,r t ,s t+1 Stored in the experience replay pool;

[0039] (7) After optimization, output the optimal compact model parameters and fit the electrical characteristics of the target device through SPICE simulation.

[0040] In step (1) above, the electrical characteristic curves include the IV characteristic curve and the CV characteristic curve.

[0041] In step (2) above, the BSIM compact model is a field-effect transistor model designed by the BSIM team at the University of California, Berkeley, comprising the BSIMCMG model (BSIM-CMG[M\OL].2020.https: / / bsim.berkeley.edu / models / bsimcmg / ) and the BSIMSOI model (BSIM-SOI[M\OL].2017.https: / / bsim.berkeley.edu / models / bsimsoi / ). The compact model contains multiple parameters to be extracted, and the value range of each parameter is defined separately.

[0042] In step (3) above, the parameter extraction problem is specifically defined as: searching for suitable compact model parameters such that when these model parameters are applied to the compact model, the electrical characteristic curve expressed by the compact model is as close as possible to the desired electrical characteristic curve. The HiMOSS (High-Dimensional Multi-Objective Subspace Sampling) algorithm mentioned above refers to the high-dimensional multi-objective subspace sampling algorithm applied to the field of circuit simulation by Zeng Xuan's team at Fudan University (Gu T, Lyu R, Bi Z, et al. HiMOSS: A Novel High-dimensional Multi-objective Optimization Method via Adaptive Gradient-Based Subspace Sampling for Analog Circuit Sizing[C]. DAC'24:61st ACM / IEEE Design Automation Conference,2024,233:1-6.), which is an improved Bayesian optimization algorithm.

[0043] In step (4) above, sampling point x i Equivalent to the model parameters selected in this round, and x i It is a vector containing all the model parameter values ​​to be extracted.

[0044] In step (5) above, the hyperparameter β is used to control the range of newly generated parameters. The larger the value of hyperparameter β, the wider the new range; the smaller the value of hyperparameter β, the narrower the new range.

[0045] In step (6) above, the agent comprises an actor network and a critic network. The input to the actor network is state s. t The output is action a t The input to the critic network is state s t and action a t The output is the action value function.

[0046] In step (7) above, the optimization termination means that the agent has completed the set number of simulations or met the termination condition set according to the calculation result of the objective function, and the optimal model parameter is the reward r. t Action a corresponding to the maximum t .

[0047] The following example illustrates the process of extracting compact model parameters from TCAD simulation data.

[0048] Figure 1This paper demonstrates the detailed process of parameter extraction based on the HiRL algorithm. The input of this embodiment includes the target IV and CV electrical characteristic data to be fitted, and the final output is the required model parameters. The target electrical characteristic curves are obtained through simulation using TCAD software. First, the HiMOSS algorithm needs to be initialized, defining the initial range of model parameter x, the maximum number of samplings T, and the hyperparameter β. Then, an optimization cycle is performed. In the t-th optimization cycle, sampling points are selected within the interval according to the subspace sampling strategy to determine the values ​​of the model parameters, denoted as x. t The corresponding reward value is then calculated using the objective function R(x), and both the sample and the corresponding reward are recorded. Then, a subspace update is performed. This process is repeated for T cycles, after which the loop ends. Based on the recorded sampled data, a new lower bound x for the model parameters is generated. L 'and new upper limit x U The reinforcement learning algorithm is optimized within the new scope. The agent will repeat the following operations: first observe the environment and confirm the current state s. t According to state s t The action a that determines the output at this moment t This refers to a compact model parameter sequence. The agent then performs action a. t The reward r at this moment is calculated using SPICE simulation software based on the objective function R(x). t Action a t Execution will generate a new state s for the next moment. t+1 This includes the state, action, reward at this moment, and the state at the next moment (s). t ,a t ,r t ,s t+1 Together, these are added to the experience replay pool. Simultaneously, the critic network and the actor network are updated. Both the critic and actor networks employ three fully connected layers. The critic network's inputs are states and actions, i.e., IV and CV characteristic curves and model parameter values. It quantifies the long-term reward of state-action pairs using the Q-function, providing optimization direction for the actor network. The actor network's inputs are states, i.e., IV and CV characteristic curves, and its output is actions, i.e., model parameter values. The loop exits when the agent's reward meets the requirements. The extracted optimal model parameters are output to fit the device's electrical characteristics, and training ends.

[0049] like Figure 2 and Figure 3As shown, two classic compact models, BSIMCMG and BSIMSOI, were selected, and a simulation experiment was conducted to compare the parameter extraction performance of TCAD-acquired data. Compared with the HiMOSS algorithm, reinforcement learning (RL) algorithm, and Bayesian optimization (BO) algorithm, the automatic parameter extraction method for device models based on the HiRL hybrid algorithm proposed in this invention converges faster and performs better.

Claims

1. A method for automatically extracting device model parameters based on a hybrid algorithm, characterized in that, By combining the HiMOSS algorithm and reinforcement learning algorithm, the manual parameter extraction process is simulated. Global exploration and local optimization are coordinated in the parameter search space, and the extracted optimal model parameters are output to achieve fitting of the device's electrical characteristics. The specific steps are as follows: (1) Collect the electrical characteristic curves of the target; (2) Select a compact model based on the device characteristics to simulate the electrical characteristics of the device, and define the range of values ​​for the parameters to be extracted in the compact model. (3) Use the HiMOSS algorithm to conduct preliminary exploration and localization in the parameter search space to generate a new range of model parameters. The parameter extraction problem is described as a mathematical problem of maximizing the objective function R(x). The objective function R(x) is defined, and the algorithm optimizes by maximizing R(x). In each optimization cycle, the HiMOSS algorithm selects sampling points within the initially defined range of model parameter values ​​according to the subspace sampling strategy, determining the current model parameter value x. Subsequently, the corresponding reward value is calculated using the objective function R(x), and both the sample and the corresponding reward are recorded. Then, the subspace is updated. After repeating this process for several cycles, the loop ends. Based on the recorded sampling data, a new range of model parameter values ​​is generated. (4) Use reinforcement learning algorithms to optimize and extract parameters within the new model parameter range. Within each optimization cycle, first observe the environment and confirm the state s at that moment. t According to state s t The action a that determines the output at this moment t This is the compact model parameter sequence; subsequently, action a is calculated using the objective function R(x). t The corresponding reward r t and obtain the action a to be executed. t The new state s t+1 This will include the state, action, reward, and state of the next moment. t ,a t ,r t ,s t+1 Add them together to the experience replay pool; After optimization, the optimal compact model parameters are output to achieve fitting of the electrical characteristics of the target device.

2. The automatic extraction method for device model parameters according to claim 1, characterized in that, In step (1), electrical characteristic curves are acquired by simulating with TCAD software or by testing real devices; the electrical characteristic curves include IV characteristic curves and CV characteristic curves.

3. The automatic extraction method for device model parameters according to claim 1, characterized in that, In step (2), the compact model is either the BSIMCMG compact model or the BSIMSOI compact model.

4. The automatic extraction method for device model parameters according to claim 1, characterized in that, In step (3), the initial settings of the HiMOSS algorithm are first performed, defining the initial range of model parameter x, the maximum number of samplings T, and the hyperparameter β used to control the range of newly generated model parameters; then the optimization cycle is cyclical.

5. The automatic extraction method for device model parameters according to claim 1, characterized in that, In step (3), the parameter extraction problem is as follows: search for suitable compact model parameters so that when these model parameters are applied to the compact model, the electrical characteristic curve expressed by the compact model is as close as possible to the desired electrical characteristic curve.

6. The automatic extraction method for device model parameters according to claim 1, characterized in that, In steps (3) and (4), when there are M data points to be fitted, the expression for the objective function R(x) is as follows: Where x represents the current model parameter values, H(i,x) represents the squared error between the fitted value and the target value for the i-th data point, and w i The weight assigned to the i-th data point.

7. The automatic extraction method for device model parameters according to claim 1, characterized in that, In step (4), the agent of the reinforcement learning algorithm includes an actor network and a critic network; the input of the actor network is state s. t The output is action a. t The input to the critic network is state s t and action a t The output is the action value function.

8. The method for automatically extracting device model parameters according to claim 1, characterized in that, In step (4), the optimization ends when the agent and the environment interact a predetermined number of times or when the conditions set in the objective function are met. At this point, the optimal compact model parameters are the reward r. t Action a corresponding to the maximum t .

Images