Method for extracting mos device bsim4 model parameters based on reinforcement learning

Abstract

The application belongs to the technical field of semiconductor device modeling and parameter extraction, and discloses a MOS device BSIM4 model parameter extraction method based on reinforcement learning. The method comprises the following steps: acquiring measured I-V and C-V curve data of a MOS device; constructing a reinforcement learning environment based on the data, defining a state space, an action space adopting a grouping cooperative mechanism, and a composite reward function based on fitting error; constructing a reinforcement learning intelligent agent adopting Dueling DQN and a priority experience replay mechanism; iteratively training the intelligent agent in the environment, and the intelligent agent optimizes the strategy by observing the state, outputting the action of adjusting the parameter group, and obtaining the reward; an improved ε-greedy and optimal step exploration mechanism is adopted during training; when the convergence condition is met, the optimized BSIM4 model parameter set is output, the automation of the parameter extraction process is realized, and the extraction efficiency and precision are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device modeling and parameter extraction technology, and in particular to a method for extracting parameters from the BSIM4 model of a MOS device based on reinforcement learning. Background Technology

[0002] As semiconductor process nodes continue to evolve, integrated circuit design places increasingly stringent demands on the accuracy and reliability of device models. The BSIM4 model, as an industry-standard compact MOSFET model, contains over 200 physical parameters. Accurately extracting these parameters is fundamental for performing precise circuit simulation, performance prediction, and yield analysis.

[0003] Currently, the extraction of BSIM4 model parameters mainly relies on a manual or semi-automatic process combining traditional optimization algorithms (such as gradient descent and genetic algorithms) with commercial simulation software (such as Hspice). A typical process includes: engineers setting initial parameter values ​​based on experience; adjusting parameters through iterative simulation to make the simulation curves approximate the measured current-voltage (IV) and capacitance-voltage (CV) curves; and finally using analytical equations to assist convergence.

[0004] However, existing technical solutions have significant drawbacks: low efficiency: due to the extremely high dimensionality of the parameter space, traditional optimization algorithms require a large number of simulation iterations, resulting in the extraction process taking several days or even weeks.

[0005] Prone to getting trapped in local optima: Algorithms such as gradient descent are sensitive to initial parameter values, while stochastic optimization methods such as genetic algorithms have unstable convergence, making it difficult to guarantee obtaining the globally optimal parameter solution. Over-reliance on human experience: Key steps such as parameter initialization, weight setting, and constraint adjustment are highly dependent on the engineer's experience, resulting in strong subjectivity and poor repeatability and consistency of the extracted results.

[0006] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0007] The main objective of this invention is to provide a method for extracting parameters of the BSIM4 model of MOS devices based on reinforcement learning, aiming to solve the technical problems of low efficiency and reliance on human experience.

[0008] To achieve the above objectives, this invention provides a method for extracting BSIM4 model parameters of MOS devices based on reinforcement learning. The method includes the following steps: The state space of the reinforcement learning environment is constructed based on the measured electrical characteristic curve data. The measured curve is discretized and encoded together with the current BSIM4 model parameter values ​​to form the state space. The adjustment instructions of the model parameters are defined as the action space, and the actions are generated by a group collaboration mechanism. The reward function is designed based on the fitting error between the simulation curve and the measured curve. Construct and initialize a reinforcement learning agent, wherein the agent adopts a Dueling DQN network structure as the decision network and is configured with a priority experience replay mechanism for sampling and learning from historical data; The agent is iteratively trained in the reinforcement learning environment using the measured electrical characteristic curve data. In each training step, the agent observes the current state of the environment and outputs an action. The agent executes the action to update the BSIM4 model parameters and performs a simulation based on the updated parameters to obtain a simulation curve. The agent calculates the reward value between the simulation curve and the measured curve according to the reward function and uses the reward value to update the decision network parameters of the agent. During the iterative training process, an improved ε-greedy strategy is adopted in combination with an optimal step exploration mechanism to balance the exploration of new actions with the utilization of existing knowledge; When the reward function value of the iterative training meets the preset convergence condition or reaches the maximum number of training steps, training stops and the current optimized BSIM4 model parameter set is output.

[0009] In one embodiment, constructing the state space of the reinforcement learning environment based on the measured electrical characteristic curve data includes: Multi-dimensional sampling and normalization are performed on the IV curve and CV curve. The processed data point sequence is then concatenated with the current key BSIM4 model parameter values ​​to be optimized to form a state vector. The state space of the reinforcement learning environment is obtained by fitting the state vector.

[0010] In one embodiment, the grouping and coordination mechanism of the action space is to divide the BSIM4 model parameters into a threshold voltage-related group, a mobility-related group, and a channel modulation effect-related group based on the correlation of the physical model; each group of parameters is assigned a coordinated adjustment action as a whole.

[0011] In one embodiment, the reward function is calculated in a composite form, as shown in the formula: R = - (α · RMS_IV + β · RMS_CV) + γ · R_progress Where RMS_IV and RMS_CV are the root mean square errors of the IV and CV curves, respectively, and α and β are the weighting coefficients of the error term; R_progress is the progress reward based on the degree of reduction of the fitting error, and γ is the weighting coefficient of R_progress.

[0012] In one embodiment, the improved ε-greedy strategy employs an exploration rate that dynamically decays with the number of training steps. The optimal step exploration mechanism is as follows: periodically storing the action sequence with the highest cumulative reward in the current evaluation period into the experience pool for priority replay in the future.

[0013] Furthermore, to achieve the above objectives, this invention also proposes a reinforcement learning-based BSIM4 model parameter extraction device for MOS devices. This reinforcement learning-based BSIM4 model parameter extraction device is applied to the reinforcement learning-based BSIM4 model parameter extraction method for MOS devices described above. The device includes: The acquisition module is used to acquire measured electrical characteristic curve data of MOS devices, including IV curves and CV curves. The construction module is used to construct the state space of the reinforcement learning environment based on the measured electrical characteristic curve data. The measured curve is discretized and encoded together with the current BSIM4 model parameter values ​​to form the state space. The adjustment instructions of the model parameters are defined as the action space, and the actions are generated by a group collaboration mechanism. The reward function is designed according to the fitting error between the simulation curve and the measured curve. The building module is used to build and initialize a reinforcement learning agent, wherein the agent adopts a Dueling DQN network structure as the decision network and is configured with a priority experience replay mechanism for sampling and learning from historical data. The training module is used to iteratively train the agent in the reinforcement learning environment using the measured electrical characteristic curve data. In each training step, the agent observes the current state of the environment and outputs an action, executes the action to update the BSIM4 model parameters, and performs simulation based on the updated parameters to obtain a simulation curve. The reward value between the simulation curve and the measured curve is calculated according to the reward function, and the decision network parameters of the agent are updated using the reward value. An additional module is added to employ an improved ε-greedy strategy combined with an optimal step exploration mechanism during the iterative training process to balance the exploration of new actions with the utilization of existing knowledge. The output module is used to stop training and output the current optimized BSIM4 model parameter set when the reward function value of the iterative training meets the preset convergence condition or reaches the maximum number of training steps.

[0014] In one embodiment, the construction module is used to perform multi-dimensional sampling and normalization processing on the IV curve and CV curve, and to concatenate the processed data point sequence with the current key BSIM4 model parameter values ​​to be optimized to form a state vector; The state space of the reinforcement learning environment is obtained by fitting the state vector.

[0015] In one embodiment, the grouping and coordination mechanism of the action space is to divide the BSIM4 model parameters into a threshold voltage-related group, a mobility-related group, and a channel modulation effect-related group based on the correlation of the physical model; each group of parameters is assigned a coordinated adjustment action as a whole.

[0016] Furthermore, to achieve the above objectives, this invention also proposes a reinforcement learning-based BSIM4 model parameter extraction device for MOS devices. The reinforcement learning-based BSIM4 model parameter extraction device includes: a memory, a processor, and a reinforcement learning-based BSIM4 model parameter extraction program stored in the memory and executable on the processor. The reinforcement learning-based BSIM4 model parameter extraction program is configured to implement the steps of the reinforcement learning-based BSIM4 model parameter extraction method for MOS devices as described above.

[0017] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a reinforcement learning-based BSIM4 model parameter extraction program for MOS devices. When the reinforcement learning-based BSIM4 model parameter extraction program is executed by a processor, it implements the steps of the reinforcement learning-based BSIM4 model parameter extraction method for MOS devices as described above.

[0018] This invention acquires measured IV and CV curve data of MOS devices; based on this data, a reinforcement learning environment is constructed, defining a state space, an action space using a group collaboration mechanism, and a composite reward function based on fitting error; a reinforcement learning agent using Dueling DQN and a priority experience replay mechanism is constructed; the agent is iteratively trained in the environment, optimizing the strategy by observing the state, outputting actions to adjust parameter sets, and obtaining rewards; an improved ε-greedy and optimal step exploration mechanism is used during training; when the convergence condition is met, the optimized BSIM4 model parameter set is output, realizing the automation of the parameter extraction process and significantly improving extraction efficiency and accuracy. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the reinforcement learning-based BSIM4 model parameter extraction method for MOS devices according to the present invention. Figure 2This is a structural block diagram of the first embodiment of the BSIM4 model parameter extraction device for MOS devices based on reinforcement learning according to the present invention.

[0020] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] This invention provides a method for extracting parameters of the BSIM4 model of MOS devices based on reinforcement learning, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the method for extracting parameters of the BSIM4 model of a MOS device based on reinforcement learning according to the present invention.

[0023] In this embodiment, the method for extracting parameters of the BSIM4 model of MOS devices based on reinforcement learning includes the following steps: Step S10: Obtain the measured electrical characteristic curve data of the MOS device.

[0024] In this embodiment, the execution entity is a reinforcement learning-based BSIM4 model parameter extraction device for MOS devices. This reinforcement learning-based BSIM4 model parameter extraction device has functions such as data processing, data communication, and program execution. The reinforcement learning-based BSIM4 model parameter extraction device for MOS devices can be a computer terminal device or other network device, or other devices with similar functions. This embodiment does not limit the scope of such devices.

[0025] It should be noted that currently, the extraction of BSIM4 model parameters mainly relies on a manual or semi-automatic process combining traditional optimization algorithms (such as gradient descent and genetic algorithms) with commercial simulation software (such as Hspice). A typical process includes: engineers setting initial parameter values ​​based on experience; adjusting parameters through iterative simulation to make the simulation curves approximate the measured current-voltage (IV) and capacitance-voltage (CV) curves; and finally using analytical equations to assist convergence. However, existing solutions have significant drawbacks: low efficiency: due to the extremely high dimensionality of the parameter space, traditional optimization algorithms require numerous simulation iterations, resulting in extraction processes that can take days or even weeks. Prone to local optima: algorithms such as gradient descent are sensitive to initial parameter values, while stochastic optimization methods such as genetic algorithms have unstable convergence, making it difficult to guarantee a globally optimal parameter solution. Over-reliance on human experience: key steps such as parameter initialization, weight setting, and constraint adjustment are highly dependent on the engineer's experience, leading to strong subjectivity and poor repeatability and consistency of the extracted results.

[0026] To address the aforementioned technical issues, this embodiment acquires measured IV and CV curve data of MOS devices; based on this data, a reinforcement learning environment is constructed, defining a state space, an action space employing a group collaboration mechanism, and a composite reward function based on fitting error; a reinforcement learning agent employing Dueling DQN and a priority experience replay mechanism is constructed; the agent is iteratively trained in the environment, optimizing its strategy by observing states, outputting actions to adjust parameter sets, and obtaining rewards; an improved ε-greedy and optimal step exploration mechanism is used during training; when the convergence condition is met, the optimized BSIM4 model parameter set is output, automating the parameter extraction process and significantly improving extraction efficiency and accuracy. Specifically, this can be implemented as follows.

[0027] In this embodiment, the measured electrical characteristic curve data includes IV curves and CV curves. This data must include at least current-voltage (IV) curves and capacitance-voltage (CV) curves measured under different bias conditions. These curves are the gold standard for evaluating the fitting accuracy of the BSIM4 model.

[0028] Step S20: Construct the state space of the reinforcement learning environment based on the measured electrical characteristic curve data.

[0029] In the specific implementation, the measured curve is discretized and encoded together with the current BSIM4 model parameter values ​​into a state space. The adjustment instructions of the model parameters are defined as the action space, and a grouping and collaboration mechanism is used to generate actions. The reward function is designed based on the fitting error between the simulation curve and the measured curve.

[0030] It should be noted that the state in the state space construction of this embodiment must fully reflect the current fitting progress. Specifically, the measured IV and CV curves are sampled and normalized in multiple dimensions to obtain a discrete data point sequence. At the same time, the key BSIM4 model parameter values ​​to be optimized in the current iteration (e.g., threshold voltage VTH0, mobility U0, saturation velocity VSAT, etc.) are composed into a parameter vector. The above curve data sequence and parameter vector are concatenated to encode a high-dimensional state vector, which is the state of the current environment. The action space is defined as actions, which are defined as adjustment instructions for the BSIM4 model parameters. To reduce the search dimensionality and utilize physical correlation, this invention adopts a grouped collaborative optimization mechanism. First, according to the physical meaning of the parameters, the BSIM4 model parameters are divided into several groups, such as: threshold voltage related parameter group (e.g., VTH0, K1, K2), mobility related parameter group (e.g., U0, UA, UB), channel modulation effect related parameter group (e.g., PCLM, PDIBL1, PDIBL2), etc. In each training step, the agent's output action is not to adjust a single parameter, but to apply a coordinated adjustment direction and step size to the entire parameter set (e.g., "increase the threshold voltage set parameters by 0.5%"). Reward function design: The reward is the guiding principle for the agent's learning. The reward function R of this invention is calculated in a composite form, aiming to simultaneously pursue high accuracy and fast convergence: the reward function is calculated in a composite form, with the formula: R = - (α · RMS_IV + β · RMS_CV) + γ · R_progress Where RMS_IV and RMS_CV are the root mean square errors of the IV and CV curves, respectively, and α and β are the weighting coefficients of the error term; R_progress is the progress reward based on the degree of reduction of the fitting error, and γ is the weighting coefficient of R_progress.

[0031] In this embodiment, the grouping and coordination mechanism of the action space is to divide the BSIM4 model parameters into threshold voltage related group, mobility related group, and channel modulation effect related group according to the correlation of the physical model; each group of parameters is assigned a coordinated adjustment action as a whole.

[0032] Furthermore, the state space of the reinforcement learning environment is constructed based on the measured electrical characteristic curve data, including: performing multi-dimensional sampling and normalization processing on the IV curve and CV curve, concatenating the processed data point sequence with the current key BSIM4 model parameter values ​​to be optimized to form a state vector; and fitting the state vector to obtain the state space of the reinforcement learning environment.

[0033] Step S30: Construct and initialize the reinforcement learning agent.

[0034] It should be noted that the agent uses a Dueling DQN network structure as the decision network and is configured with a priority experience replay mechanism for sampling and learning from historical data.

[0035] In its implementation, a deep reinforcement learning agent using the Dueling Deep Q-Network (DQN) architecture is employed. Its decision network (Q-network) can separately estimate the state value and the advantage value of each action, thus more robustly evaluating action value in a complex parameter space. Simultaneously, the agent is configured with a Prioritized Experience Replay mechanism. This mechanism does not sample uniformly from historical interaction data (experience pool), but instead prioritizes replaying experiences with higher training value (such as larger TD errors), thereby significantly improving learning efficiency and convergence speed.

[0036] Step S40: Use the measured electrical characteristic curve data to iteratively train the agent in the reinforcement learning environment.

[0037] It should be noted that in each training step, the agent observes the current state of the environment and outputs an action, executes the action to update the BSIM4 model parameters, and performs simulation based on the updated parameters to obtain a simulation curve. The reward value between the simulation curve and the measured curve is calculated according to the reward function, and the decision network parameters of the agent are updated using the reward value.

[0038] In the specific implementation, each training step (Step) contains the following sub-loops: Observation: The agent observes the current state S_t of the environment. Decision and Execution: The agent outputs an action A_t (e.g., "increase the mobility group parameter by 1%) according to its policy (initially a random policy or an ε-greedy policy). The environment executes this action, i.e., updates the BSIM4 model parameters according to the action instruction. Simulation and Evaluation: Based on the updated BSIM4 parameters, a circuit simulator (e.g., Hspice) is invoked to perform simulation, obtaining the corresponding simulation IV and CV curves. Reward Calculation: The environment calculates the reward value R_t between the simulation curve and the measured curve according to the reward function designed in step S102. Observe New State: The environment transitions to the next state S_{t+1} (i.e., the state encoded based on the new parameters and the new simulation curve). Storage and Learning: The experience tuple (S_t, A_t, R_t, S_{t+1}) of this interaction is stored in the experience pool. The agent samples a batch of experience from the experience pool, calculates the target Q value using R_t and S_{t+1}, and updates the parameters of its Dueling DQN decision network through backpropagation, thereby learning a better policy.

[0039] Step S50: During the iterative training process, an improved ε-greedy strategy is adopted in combination with an optimal step exploration mechanism to balance the exploration of new actions with the utilization of existing knowledge.

[0040] In the specific implementation, an improved ε-greedy strategy is adopted. Specifically, a high exploration rate (ε) is set in the early stage of training to encourage the agent to try different parameter adjustments. As the number of training steps increases, the exploration rate ε gradually and dynamically decays, causing the agent to gradually tend to choose the action currently considered optimal. In addition, an optimal step exploration mechanism is introduced: the continuous action sequence with the highest cumulative reward in the current cycle is periodically evaluated and recorded, and stored as a special "successful experience" in the experience pool for subsequent priority replay and learning, thereby accelerating convergence.

[0041] It should be noted that the improved ε-greedy strategy employs an exploration rate that dynamically decays with the number of training steps; that is, a higher exploration rate is used initially, and then gradually reduced. The optimal step exploration mechanism is as follows: the action sequence with the highest cumulative reward in the current evaluation period is periodically stored in the experience pool for priority replay later.

[0042] Step S60: When the reward function value of the iterative training satisfies the preset convergence condition or reaches the maximum number of training steps, stop training and output the current optimized BSIM4 model parameter set. In the specific implementation, training stops when the reward function value stabilizes at a high level for a continuous period of time (meeting the preset convergence condition) or reaches the preset maximum number of training steps. Finally, the optimized set of BSIM4 model parameters corresponding to the agent at this point is output. This embodiment can shorten the parameter extraction time from several weeks in traditional methods to several hours, and achieve high-precision results with both IV and CV curve fitting errors less than 5%.

[0043] In this embodiment, measured IV and CV curve data of MOS devices are acquired. Based on this data, a reinforcement learning environment is constructed, defining a state space, an action space using a group collaboration mechanism, and a composite reward function based on the fitting error. A reinforcement learning agent using Dueling DQN and a priority experience replay mechanism is constructed. The agent is iteratively trained in the environment, where it optimizes its strategy by observing states, outputting actions to adjust parameter sets, and obtaining rewards. An improved ε-greedy and optimal step exploration mechanism is used during training. When the convergence condition is met, the optimized BSIM4 model parameter set is output, automating the parameter extraction process and significantly improving extraction efficiency and accuracy.

[0044] Furthermore, this embodiment of the invention also proposes a storage medium storing a reinforcement learning-based BSIM4 model parameter extraction program for MOS devices. When the reinforcement learning-based BSIM4 model parameter extraction program is executed by a processor, it implements the steps of the reinforcement learning-based BSIM4 model parameter extraction method for MOS devices as described above.

[0045] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the BSIM4 model parameter extraction device for MOS devices based on reinforcement learning according to the present invention.

[0046] like Figure 2 As shown, the reinforcement learning-based BSIM4 model parameter extraction device for MOS devices proposed in this embodiment of the invention includes: The acquisition module 10 is used to acquire the measured electrical characteristic curve data of the MOS device, including the IV curve and the CV curve. The construction module 20 is used to construct the state space of the reinforcement learning environment based on the measured electrical characteristic curve data. The measured curve is discretized and encoded together with the current BSIM4 model parameter values ​​to form the state space. The adjustment instructions of the model parameters are defined as the action space, and the actions are generated by a group collaboration mechanism. The reward function is designed according to the fitting error between the simulation curve and the measured curve. The construction module 20 is used to construct and initialize a reinforcement learning agent, wherein the agent adopts a Dueling DQN network structure as the decision network and is configured with a priority experience replay mechanism for sampling and learning from historical data. Training module 30 is used to iteratively train the agent in the reinforcement learning environment using the measured electrical characteristic curve data. In each training step, the agent observes the current state of the environment and outputs an action, executes the action to update the BSIM4 model parameters, and performs simulation based on the updated parameters to obtain a simulation curve. The agent calculates the reward value between the simulation curve and the measured curve according to the reward function, and uses the reward value to update the decision network parameters of the agent. Add module 40 to employ an improved ε-greedy strategy combined with an optimal step exploration mechanism during the iterative training process to balance the exploration of new actions with the utilization of existing knowledge; The output module 50 is used to stop training and output the current optimized BSIM4 model parameter set when the reward function value of the iterative training meets the preset convergence condition or reaches the maximum number of training steps.

[0047] In this embodiment, measured IV and CV curve data of MOS devices are acquired. Based on this data, a reinforcement learning environment is constructed, defining a state space, an action space using a group collaboration mechanism, and a composite reward function based on the fitting error. A reinforcement learning agent using Dueling DQN and a priority experience replay mechanism is constructed. The agent is iteratively trained in the environment, where it optimizes its strategy by observing states, outputting actions to adjust parameter sets, and obtaining rewards. An improved ε-greedy and optimal step exploration mechanism is used during training. When the convergence condition is met, the optimized BSIM4 model parameter set is output, automating the parameter extraction process and significantly improving extraction efficiency and accuracy.

[0048] In one embodiment, the construction module 20 is used to perform multi-dimensional sampling and normalization processing on the IV curve and CV curve, and to concatenate the processed data point sequence with the current key BSIM4 model parameter values ​​to be optimized to form a state vector; The state space of the reinforcement learning environment is obtained by fitting the state vector.

[0049] In one embodiment, the grouping and coordination mechanism of the action space is to divide the BSIM4 model parameters into a threshold voltage-related group, a mobility-related group, and a channel modulation effect-related group based on the correlation of the physical model; each group of parameters is assigned a coordinated adjustment action as a whole.

[0050] This application embodiment also provides a reinforcement learning-based BSIM4 model parameter extraction device for MOS devices, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other through the communication bus. The memory is used to store the reinforcement learning-based BSIM4 model parameter extraction program for MOS devices. When the processor executes the program stored in the memory, it implements the above-mentioned reinforcement learning-based BSIM4 model parameter extraction method for MOS devices.

[0051] The communication bus mentioned in the reinforcement learning-based BSIM4 model parameter extraction device for MOS devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0052] The communication interface is used for communication between the aforementioned reinforcement learning-based MOS device BSIM4 model parameter extraction device and other devices.

[0053] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0054] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0055] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0057] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0059] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0060] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0061] In addition, for technical details not described in detail in this embodiment, please refer to the reinforcement learning-based BSIM4 model parameter extraction method for MOS devices provided in any embodiment of the present invention, which will not be repeated here.

[0062] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0063] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0064] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0065] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

[0066] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above method.

Claims

1. A method for extracting parameters of the BSIM4 model of a MOS device based on reinforcement learning, characterized in that, The reinforcement learning-based BSIM4 model parameter extraction method for MOS devices includes: Obtain the measured electrical characteristic curve data of the MOS device, including the IV curve and CV curve; The state space of the reinforcement learning environment is constructed based on the measured electrical characteristic curve data. The measured curve is discretized and encoded together with the current BSIM4 model parameter values ​​to form the state space. The adjustment instructions of the model parameters are defined as the action space, and the actions are generated by a group collaboration mechanism. The reward function is designed based on the fitting error between the simulation curve and the measured curve. Construct and initialize a reinforcement learning agent, wherein the agent adopts a Dueling DQN network structure as the decision network and is configured with a priority experience replay mechanism for sampling and learning from historical data; The agent is iteratively trained in the reinforcement learning environment using the measured electrical characteristic curve data. In each training step, the agent observes the current state of the environment and outputs an action. The agent executes the action to update the BSIM4 model parameters and performs a simulation based on the updated parameters to obtain a simulation curve. The agent calculates the reward value between the simulation curve and the measured curve according to the reward function and uses the reward value to update the decision network parameters of the agent. During the iterative training process, an improved ε-greedy strategy is adopted in combination with an optimal step exploration mechanism to balance the exploration of new actions with the utilization of existing knowledge; When the reward function value of the iterative training meets the preset convergence condition or reaches the maximum number of training steps, training stops and the current optimized BSIM4 model parameter set is output.

2. The method for extracting parameters of the BSIM4 model of MOS devices based on reinforcement learning as described in claim 1, characterized in that, The state space of the reinforcement learning environment constructed based on the measured electrical characteristic curve data includes: Multi-dimensional sampling and normalization are performed on the IV curve and CV curve. The processed data point sequence is then concatenated with the current key BSIM4 model parameter values ​​to be optimized to form a state vector. The state space of the reinforcement learning environment is obtained by fitting the state vector.

3. The method for extracting parameters of the BSIM4 model of MOS devices based on reinforcement learning as described in claim 1, characterized in that, The grouping and coordination mechanism of the action space is based on the correlation of the physical model, dividing the BSIM4 model parameters into threshold voltage related group, mobility related group, and channel modulation effect related group; each group of parameters is assigned a coordinated adjustment action as a whole.

4. The method for extracting parameters of the BSIM4 model of MOS devices based on reinforcement learning as described in claim 1, characterized in that, The reward function is calculated in a composite form, and the formula is as follows: R = - (α · RMS_IV + β · RMS_CV) + γ · R_progress Where RMS_IV and RMS_CV are the root mean square errors of the IV and CV curves, respectively, and α and β are the weighting coefficients of the error term; R_progress is the progress reward based on the degree of reduction of the fitting error, and γ is the weighting coefficient of R_progress.

5. The method for extracting parameters of the BSIM4 model of MOS devices based on reinforcement learning as described in claim 1, characterized in that, The improved ε-greedy strategy employs an exploration rate that dynamically decays with the number of training steps. The optimal step exploration mechanism is as follows: periodically store the action sequence with the highest cumulative reward in the current evaluation period into the experience pool for priority replay in the future.

6. A device for extracting parameters of a BSIM4 model of a MOS device based on reinforcement learning, characterized in that, The reinforcement learning-based BSIM4 model parameter extraction device for MOS devices is applied to the reinforcement learning-based BSIM4 model parameter extraction method for MOS devices as described in any one of claims 1 to 5, and the device comprises: The acquisition module is used to acquire measured electrical characteristic curve data of MOS devices, including IV curves and CV curves. The construction module is used to construct the state space of the reinforcement learning environment based on the measured electrical characteristic curve data. The measured curve is discretized and encoded together with the current BSIM4 model parameter values ​​to form the state space. The adjustment instructions of the model parameters are defined as the action space, and the actions are generated by a group collaboration mechanism. The reward function is designed according to the fitting error between the simulation curve and the measured curve. The building module is used to build and initialize a reinforcement learning agent, wherein the agent adopts the DuelingDQN network structure as the decision network and is configured with a priority experience replay mechanism for sampling and learning from historical data. The training module is used to iteratively train the agent in the reinforcement learning environment using the measured electrical characteristic curve data. In each training step, the agent observes the current state of the environment and outputs an action, executes the action to update the BSIM4 model parameters, and performs simulation based on the updated parameters to obtain a simulation curve. The reward value between the simulation curve and the measured curve is calculated according to the reward function, and the decision network parameters of the agent are updated using the reward value. An additional module is added to employ an improved ε-greedy strategy combined with an optimal step exploration mechanism during the iterative training process to balance the exploration of new actions with the utilization of existing knowledge. The output module is used to stop training and output the current optimized BSIM4 model parameter set when the reward function value of the iterative training meets the preset convergence condition or reaches the maximum number of training steps.

7. The BSIM4 model parameter extraction device for MOS devices based on reinforcement learning as described in claim 6, characterized in that, The construction module is used to perform multi-dimensional sampling and normalization processing on the IV curve and CV curve, and to concatenate the processed data point sequence with the current key BSIM4 model parameter values ​​to be optimized to form a state vector. The state space of the reinforcement learning environment is obtained by fitting the state vector.

8. The BSIM4 model parameter extraction device for MOS devices based on reinforcement learning as described in claim 6, characterized in that, The grouping and coordination mechanism of the action space is based on the correlation of the physical model, dividing the BSIM4 model parameters into threshold voltage related group, mobility related group, and channel modulation effect related group; each group of parameters is assigned a coordinated adjustment action as a whole.

9. A device for extracting parameters of a BSIM4 model of a MOS device based on reinforcement learning, characterized in that, The reinforcement learning-based MOS device BSIM4 model parameter extraction device includes: a memory, a processor, and a reinforcement learning-based MOS device BSIM4 model parameter extraction program stored in the memory and executable on the processor. The reinforcement learning-based MOS device BSIM4 model parameter extraction program is configured to implement the steps of the reinforcement learning-based MOS device BSIM4 model parameter extraction method as described in any one of claims 1 to 5.

10. A storage medium, characterized in that, The storage medium stores a BSIM4 model parameter extraction program for MOS devices based on reinforcement learning. When the BSIM4 model parameter extraction program for MOS devices based on reinforcement learning is executed by the processor, it implements the steps of the BSIM4 model parameter extraction method for MOS devices based on reinforcement learning as described in any one of claims 1 to 5.

Images