DTCO circuit parameter adjustment method and system integrating deep learning and reinforcement learning
The circuit modeling is carried out through deep learning and the reinforcement learning model is used to automatically adjust parameters, which solves the problems of modeling difficulties and low parameter adjustment efficiency in the DTCO method, and achieves rapid response and efficient design.
Patent Information
- Application Number
- CN202411932002.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-26
- Publication Date
- 2025-05-02
AI Technical Summary
The existing DTCO method has difficulty in modeling and low parameter adjustment efficiency in Spice Model generation and circuit parameter adjustment, which requires a lot of human intervention and time consumption.
Deep learning is used for circuit modeling, and parameters are adjusted through reinforcement learning models to automatically learn and discover the optimal circuit parameter settings.
It realizes rapid response and adaptation to process changes, meets the design requirements of high performance, low power consumption and fast iteration of integrated circuits, and improves parameter adjustment efficiency.
Smart Images
Figure CN119918487A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of integrated circuit technology, and in particular to a DTCO circuit parameter adjustment method and system integrating deep learning and reinforcement learning. Background Art
[0002] In the post-Moore era, the design and manufacture of integrated circuits face increasingly complex challenges. As an advanced methodology, DTCO aims to consider manufacturing process limitations and design requirements in the early stages of design to improve device performance, reduce manufacturing costs and maximize yield.
[0003] However, the traditional DTCO method has been challenged, especially in the generation of Spice Models and their subsequent application in circuit simulation: the physical basis modeling of Spice Models is difficult and time-consuming, while the empirical basis or equivalent circuit model can simplify the Spice Model modeling, but the model is not scalable; in the design of the circuit, the dedicated Spice Model requires multiple iterations, and a lot of human intervention and time-consuming parameter adjustment can meet the needs of dedicated chip PPAC.
[0004] Therefore, there is a need for a method that can solve the existing DTCO modeling difficulties and low parameter adjustment efficiency. Summary of the invention
[0005] To solve the above problems, the present invention provides a DTCO circuit parameter adjustment method and system that integrates deep learning and reinforcement learning. Deep learning is used for circuit modeling, and a reinforcement learning model is trained to perform parameter adjustment using the reinforcement learning model. This solves the problems of modeling difficulties in existing DTCO methods, the need for manual parameter adjustment, and low parameter adjustment efficiency.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] A DTCO circuit parameter adjustment method and system integrating deep learning and reinforcement learning, comprising the following steps:
[0008] S1. Obtaining a circuit data set, wherein the circuit data set includes several groups of circuit models, each group of circuit models corresponding to an electrical characteristic;
[0009] S2. Build a deep learning model, input a circuit data set into the deep learning model, train the deep learning model, and obtain a trained deep learning model;
[0010] S3. Construct a reinforcement learning model and perform simulation training on the reinforcement learning model, wherein the simulation training includes:
[0011] Inputting the simulation circuit parameters and the simulation results into the reinforcement learning model, the reinforcement learning model adjusts the simulation circuit according to the simulation circuit parameters and the simulation results, and optimizing the reinforcement learning model according to the loss of the simulation results of the simulation circuit after the parameter adjustment;
[0012] Repeat the simulation training of the reinforcement learning model until the simulation result loss of the simulation circuit is less than a preset threshold, and then end the simulation training;
[0013] S4. Input electrical characteristic requirements into the trained deep learning model, and the trained deep learning model outputs a circuit model that meets the electrical characteristic requirements. Input the circuit model that meets the electrical characteristic requirements into the reinforcement learning model, and the reinforcement learning model outputs the circuit model after parameter adjustment to complete the parameter adjustment.
[0014] Furthermore, the deep learning model includes a CNN model part and a RNN model part.
[0015] Furthermore, the CNN model part includes three convolutional layers connected in sequence, each convolutional layer is followed by a pooling layer, and the convolution expression of the CNN model part is:
[0016] a (l) =f(W (l) *x (l-1) +b (l) )
[0017] Among them, a (l) represents the activation value of the lth layer, f represents the activation function, W (l) is the weight of the lth layer, * represents the convolution operation, x (l-1) Represents the data of the previous layer, b (l) is the bias of the lth layer.
[0018] Furthermore, the RNN model part is a bidirectional LSTM model, and the mathematical expression of the RNN model part is:
[0019] h t =f(W hh h l-1 +W xh x t +b h )
[0020] y t =W hy h t +b y
[0021] Among them, h t is the hidden state at time step t, x t is the input at time step t, W hh , W xh, W hy is the weight matrix, b h 、b y is the bias term, y t is the output at time step t.
[0022] Furthermore, the connection relationship between the CNN model part and the RNN model part is: after the data is input into the CNN model part, the CNN model part outputs the spatial features of the data, and after the spatial features of the data are flattened, feature sequence data is obtained, and the feature sequence data is input into the RNN model part, and the RNN model part outputs a circuit model that meets the electrical characteristic requirements.
[0023] Furthermore, the reinforcement learning model is a Q learning model, and the Q learning model includes a value function Q. The value function Q is updated in each simulation training, and the updated mathematical expression is:
[0024] Q(s t ,a t )←Q(s t ,a t )+α[r t+1 +γmax a Q(s t+1 ,a)-Q(s t ,a t )]
[0025] Among them, s t represents the simulation result at time t, a t represents the parameter adjustment action for the simulation circuit, α is the learning rate, γ is the discount factor, and r t+1 It is the reward returned by the simulation circuit after adjusting the parameters of the simulation circuit.
[0026] Furthermore, during the simulation training process, the reinforcement learning model controls the learning rate in the following manner:
[0027] Q(a)=Q(a)+α(R(a)-Q(a))
[0028] Where a represents the parameter adjustment action on the simulation circuit, Q(a) represents the expected return after performing the parameter adjustment action a on the simulation circuit, R(a) represents the actual return of the simulation circuit after performing the parameter adjustment action on the simulation circuit, and α is the learning rate.
[0029] Furthermore, the simulation training also includes parallel training, and the specific implementation method of the parallel training is:
[0030] Construct several Q learning models, train several Q learning models in parallel during each training process, collect the training results of all Q learning models, and optimize the reinforcement learning model according to the training results of all Q learning models.
[0031] Furthermore, the specific implementation method of optimizing the reinforcement learning model according to the training results of all Q learning models is: constructing a policy network, the policy network includes all Q learning models and policy parameters, and for each simulation training, the update expression of the policy parameters is:
[0032]
[0033] Among them, Δθ represents the update amount of the policy parameters, N is the number of Q learning models, and π θ is the policy network, a i represents the parameter adjustment action of the i-th Q learning model on the simulation circuit, s i Indicates that the i-th Q learning model performs action a on the simulation circuit i The simulation results of the post-simulation circuit, R i Indicates that the i-th Q learning model performs action a on the simulation circuit i Post-simulation circuit returns.
[0034] Through the above technical scheme, the present invention has the following beneficial effects: automatically learning and discovering the optimal circuit parameter settings through a reinforcement learning algorithm, achieving rapid response and adaptation to process changes to meet the requirements of high performance, low power consumption and rapid iteration of integrated circuit design. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 The figure is a schematic diagram of the overall process of a DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning of the present invention.
[0036] Figure 2 Schematic diagram of the 6T TFET SRAM circuit parameter adjustment process in an embodiment of the present invention.
[0037] Figure 3 This is a structural diagram of a DTCO circuit parameter adjustment system that integrates deep learning and reinforcement learning in an embodiment of the present invention. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0040] Example 1
[0041] See also Figure 1 , a DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning, comprising the following steps:
[0042] S1. Obtaining a circuit data set, wherein the circuit data set includes several groups of circuit models, each group of circuit models corresponding to an electrical characteristic;
[0043] S2. Build a deep learning model, input a circuit data set into the deep learning model, train the deep learning model, and obtain a trained deep learning model;
[0044] S3. Construct a reinforcement learning model and perform simulation training on the reinforcement learning model, wherein the simulation training includes:
[0045] Inputting the simulation circuit parameters and the simulation results into the reinforcement learning model, the reinforcement learning model adjusts the simulation circuit according to the simulation circuit parameters and the simulation results, and optimizing the reinforcement learning model according to the loss of the simulation results of the simulation circuit after the parameter adjustment;
[0046] Repeat the simulation training of the reinforcement learning model until the simulation result loss of the simulation circuit is less than a preset threshold, and then end the simulation training;
[0047] S4. Input electrical characteristic requirements into the trained deep learning model, and the trained deep learning model outputs a circuit model that meets the electrical characteristic requirements. Input the circuit model that meets the electrical characteristic requirements into the reinforcement learning model, and the reinforcement learning model outputs the circuit model after parameter adjustment to complete the parameter adjustment.
[0048] Specifically, the circuit model is a Spice circuit model, and the analog circuit simulator is an analog circuit simulator, which is used in electronic design and can perform nonlinear DC analysis, nonlinear transient analysis and linear AC analysis on the circuit. By simply selecting the model level and giving appropriate parameters, the Spice model can accurately calculate various working characteristics of the system, such as static and dynamic.
[0049] The Spice model is a netlist file created based on the connection relationship between the components in the schematic diagram. The netlist consists of a series of sub-circuits. Users can simply create a simulation netlist by calling related sub-circuit modules. The parameters of each component in the schematic diagram mainly characterize the physical and electrical properties of the components. The accuracy of the simulation results is determined by adjusting the components and their connection relationships, as well as adjusting the parameters.
[0050] In an optional embodiment, the deep learning model includes a CNN model part and a RNN model part.
[0051] By adopting a neural network with a hybrid structure of CNN and RNN, the relationship between the device parameters of the circuit model and the electrical characteristics is learned. The device parameters of the circuit model include gate length, gate width, threshold voltage, oxide layer thickness, mobility, etc., and the electrical characteristics include drive current, leakage current, subthreshold swing, capacitance, etc. The CNN model is a convolutional neural network. The CNN model part is used to extract the characteristic data of the circuit device parameters, and the RNN part is used to capture the sequence association between the parameters.
[0052] In an optional embodiment, the CNN model part includes three convolutional layers connected in sequence, each convolutional layer is followed by a pooling layer, and the convolution expression of the CNN model part is:
[0053] a (l) =f(W (l) *x (l-1) +b (l) )
[0054] Among them, a (l) represents the activation value of the lth layer, f represents the activation function, W (l) is the weight of the lth layer, * represents the convolution operation, x (l-1) Represents the data of the previous layer, b (l) is the bias of the lth layer.
[0055] Specifically, the activation function used in the CNN model is the ReLU function.
[0056] In an optional embodiment, the RNN model part is a bidirectional LSTM model, and the mathematical expression of the RNN model part is:
[0057] h t =f(W hh h l-1 +W xh x t +b h )
[0058] y t =W hy h t +b y
[0059] Among them, h t is the hidden state at time step t, x t is the input at time step t, W hh , W xh , W hy is the weight matrix, b h , b y is the bias term, y t is the output at time step t.
[0060] The RNN model part adopts a bidirectional LTSM structure, that is, a bidirectional long short-term memory neural network. The bidirectional LSTM is a special RNN that can process input sequences and output sequences at the same time, thereby improving the expressiveness and generalization ability of the model and better learning the impact of parameters on electrical characteristics.
[0061] In an optional embodiment, the connection relationship between the CNN model part and the RNN model part is: after the data is input into the CNN model part, the CNN model part outputs the spatial features of the data, and after the spatial features of the data are flattened, feature sequence data is obtained, and the feature sequence data is input into the RNN model part, and the RNN model part outputs a circuit model that meets the electrical characteristics requirements.
[0062] Specifically, during each training process, the input data passes through the CNN, extracts spatial features and outputs them, and the output of the CNN is flattened as the input sequence of the RNN. The RNN processes sequence data and captures the timing and dependencies between parameters. The final hidden state of the RNN passes through the fully connected layer and outputs the prediction result.
[0063] Specifically, the update method of the CNN model part and the RNN model part is the gradient descent method, and the CNN model part and the RNN model part calculate the loss through the multi-classification cross entropy loss function.
[0064] In an optional embodiment, the deep learning model is a DBN (deep belief network) model, the DBN model includes a plurality of RBMs (restricted Boltzmann machines), each RBM includes a visible layer and a hidden layer, and the modeling method of the visible layer and the hidden layer is:
[0065]
[0066] Where p(h|v) represents the conditional probability of the hidden layer state h given the visible layer state v, E(v,h) is the energy function, and z is the normalization factor to ensure that the sum of the probabilities is 1.
[0067] In an optional embodiment, the reinforcement learning model is a Q learning model, and the Q learning model includes a value function Q. The value function Q is updated in each simulation training, and the updated mathematical expression is:
[0068] Q(s t ,a t )←Q(s t ,a t )+α[r t+1 +γmax a Q(s t+1 ,a)-Q(s t ,at )]
[0069] Among them, s t represents the simulation result at time t, a t represents the parameter adjustment action for the simulation circuit, α is the learning rate, γ is the discount factor, and r t+1 It is the reward returned by the simulation circuit after adjusting the parameters of the simulation circuit.
[0070] The Q-learning method is a value-based reinforcement learning algorithm, the core of which is to learn an action-value function Q to evaluate the expected utility of alternative model groups taking different key parameters under a given circuit simulation state.
[0071] Specifically, the parameters that the Q learning model mainly adjusts during the parameter adjustment process are the key parameters of the circuit. By generating a key parameter matrix, the key parameter matrix is input into the simulation circuit for parameter adjustment. Through the reinforcement learning algorithm, the DTCO circuit parameter adjustment method can automatically adjust the circuit parameters to achieve performance optimization of the circuit design. This method not only improves the performance of the circuit, but also helps to reduce the design cycle and cost.
[0072] During the training process, the reward of the simulation circuit is set according to the performance indicators of the simulation circuit, and the reward is evaluated with PPAC (i.e., the balance between performance, power consumption, area, and cost) as the goal. When the parameter adjustment action of the Q learning model improves the circuit performance, reduces power consumption, reduces area, and reduces cost, and does not violate the design constraints, the reward is positive. When the parameter adjustment action of the Q learning model leads to performance degradation, increased power consumption, increased area, increased cost, or violates the design constraints, the reward is negative.
[0073] In an optional embodiment, the reinforcement learning model controls the learning rate during simulation training in the following manner:
[0074] Q(a)=Q(a)+α(R(a)-Q(a))
[0075] Where a represents the parameter adjustment action on the simulation circuit, Q(a) represents the expected return after performing the parameter adjustment action a on the simulation circuit, R(a) represents the actual return of the simulation circuit after performing the parameter adjustment action on the simulation circuit, and α is the learning rate, which controls the speed at which the information of the new parameter matrix combination covers the old information.
[0076] In an optional embodiment, the simulation training further includes parallel training, and the specific implementation of the parallel training is as follows:
[0077] Construct several Q learning models, train several Q learning models in parallel during each training process, collect the training results of all Q learning models, and optimize the reinforcement learning model according to the training results of all Q learning models.
[0078] In an optional embodiment, the specific implementation method of optimizing the reinforcement learning model according to the training results of all Q learning models is: constructing a policy network, the policy network includes all Q learning models and policy parameters, and for each simulation training, the update expression of the policy parameters is:
[0079]
[0080] Among them, Δθ represents the update amount of the policy parameters, N is the number of Q learning models, and π θ is the policy network, a i represents the parameter adjustment action of the i-th Q learning model on the simulation circuit, s i Indicates that the i-th Q learning model performs action a on the simulation circuit i The simulation results of the post-simulation circuit, R i Indicates that the i-th Q learning model performs action a on the simulation circuit i Post-simulation circuit returns.
[0081] Multiple Q-learning models are trained in parallel to accelerate the optimization process and increase the diversity of solutions. Each Q-learning model adjusts the key parameter matrix in its own SPICE simulation environment to adjust the simulation circuit parameters and collect independent simulation result experience.
[0082] See also Figure 2 , Figure 2 The specific process of device parameter tuning based on the read access time and write delay in the PPAC indicators of the 6TTFET SRAM circuit is demonstrated. After the device parameter tuning by deep learning and reinforcement learning using Spice model simulation, the read access time is reduced due to the adjustment of the device group parameters in the 6TTFET SRAM circuit: VDD is 300Mv, and the read delay is reduced to 10 -7 s, the write latency is reduced to 5*10 -8 This improvement optimizes the overall performance of the SRAM circuit, making it more efficient in high-speed computing and big data processing based on SRAM circuits.
[0083] Example 2
[0084] See also Figure 3 , a DTCO circuit parameter adjustment system integrating deep learning and reinforcement learning, including:
[0085] A data acquisition module, used to acquire a circuit data set, wherein the circuit data set includes a plurality of groups of circuit models, each group of circuit models corresponds to an electrical characteristic;
[0086] A deep learning module is used to build a deep learning model, input a circuit data set into the deep learning model, train the deep learning model, and obtain a trained deep learning model;
[0087] The reinforcement learning module is used to build a reinforcement learning model and perform simulation training on the reinforcement learning model. The simulation training includes:
[0088] Inputting the simulation circuit parameters and the simulation results into the reinforcement learning model, the reinforcement learning model adjusts the simulation circuit according to the simulation circuit parameters and the simulation results, and optimizing the reinforcement learning model according to the loss of the simulation results of the simulation circuit after the parameter adjustment;
[0089] Repeat the simulation training of the reinforcement learning model until the simulation result loss of the simulation circuit is less than a preset threshold, and then end the simulation training;
[0090] The circuit parameter adjustment module is used to input electrical characteristic requirements into the trained deep learning model, and the trained deep learning model outputs a circuit model that meets the electrical characteristic requirements. The circuit model that meets the electrical characteristic requirements is input into the reinforcement learning model, and the reinforcement learning model outputs the adjusted circuit model to complete the parameter adjustment.
[0091] The embodiment disclosed in this specification is only an example of a unilateral feature of the present invention, and the protection scope of the present invention is not limited to this embodiment, and any other functionally equivalent embodiments fall within the protection scope of the present invention. For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present invention.
Claims
1. A DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning, characterized in that: The following steps are involved: S1. Obtaining a circuit data set, wherein the circuit data set includes several groups of circuit models, each group of circuit models corresponding to an electrical characteristic; S2. Build a deep learning model, input a circuit data set into the deep learning model, train the deep learning model, and obtain a trained deep learning model; S3. Construct a reinforcement learning model and perform simulation training on the reinforcement learning model, wherein the simulation training includes: Inputting the simulation circuit parameters and the simulation results into the reinforcement learning model, the reinforcement learning model adjusts the simulation circuit according to the simulation circuit parameters and the simulation results, and optimizing the reinforcement learning model according to the loss of the simulation results of the simulation circuit after the parameter adjustment; Repeat the simulation training of the reinforcement learning model until the simulation result loss of the simulation circuit is less than a preset threshold, and then end the simulation training; S4. Input electrical characteristic requirements into the trained deep learning model, and the trained deep learning model outputs a circuit model that meets the electrical characteristic requirements. Input the circuit model that meets the electrical characteristic requirements into the reinforcement learning model, and the reinforcement learning model outputs the circuit model after parameter adjustment to complete the parameter adjustment.
2. According to the DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning according to claim 1, it is characterized in that: The deep learning model includes a CNN model part and a RNN model part.
3. According to claim 2, a DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning is characterized in that: The CNN model part includes three convolutional layers connected in sequence, each convolutional layer is followed by a pooling layer, and the convolution expression of the CNN model part is: a (l) =f(W (l) *x (l-1) +b (l) ) Among them, a (l) represents the activation value of the lth layer, f represents the activation function, W (l) is the weight of the lth layer, * represents the convolution operation, x (l-1) Represents the data of the previous layer, b (l) is the bias of the lth layer.
4. According to the DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning according to claim 3, it is characterized in that: The RNN model part is a bidirectional LSTM model, and the mathematical expression of the RNN model part is: h t =f(W hh h l-1 +W xh x t +b h ) y t =W hy h t +b y Among them, h t is the hidden state at time step t, x t is the input at time step t, W hh , W xh , W hy is the weight matrix, b h 、b y is the bias term, y t is the output at time step t.
5. According to the DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning according to claim 4, it is characterized in that: The connection relationship between the CNN model part and the RNN model part is: after the data is input into the CNN model part, the CNN model part outputs the spatial features of the data, and after the spatial features of the data are flattened, feature sequence data is obtained, and the feature sequence data is input into the RNN model part, and the RNN model part outputs a circuit model that meets the electrical characteristic requirements.
6. According to a DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning according to claim 1, it is characterized in that: The reinforcement learning model is a Q learning model, which includes a value function Q. The value function Q is updated in each simulation training, and the updated mathematical expression is: Q(s t ,a t )←Q(s t ,a t )+α[r t+1 +γmax a Q(s t+1 ,a)-Q(s t ,a t )] Among them, s t represents the simulation result at time t, a t represents the parameter adjustment action for the simulation circuit, α is the learning rate, γ is the discount factor, and r t+1 It is the reward returned by the simulation circuit after adjusting the parameters of the simulation circuit.
7. The DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning according to claim 6 is characterized in that: During the simulation training process, the reinforcement learning model controls the learning rate in the following ways: Q(a)=Q(a)+α(R(a)-Q(a)) Where a represents the parameter adjustment action on the simulation circuit, Q(a) represents the expected return after performing the parameter adjustment action a on the simulation circuit, R(a) represents the actual return of the simulation circuit after performing the parameter adjustment action on the simulation circuit, and α is the learning rate.
8. The DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning according to claim 7, characterized in that: The simulation training also includes parallel training, and the specific implementation method of the parallel training is: Construct several Q learning models, train several Q learning models in parallel during each training process, collect the training results of all Q learning models, and optimize the reinforcement learning model according to the training results of all Q learning models.
9. The DTCO circuit parameter adjustment method integrating deep learning and reinforcement learning according to claim 8, characterized in that: The specific implementation method of optimizing the reinforcement learning model according to the training results of all Q learning models is: constructing a policy network, which includes all Q learning models and policy parameters. For each simulation training, the update expression of the policy parameters is: Among them, Δθ represents the update amount of the policy parameters, N is the number of Q learning models, and π θ is the policy network, a i represents the parameter adjustment action of the i-th Q learning model on the simulation circuit, s i Indicates that the i-th Q learning model performs action a on the simulation circuit i The simulation results of the post-simulation circuit, R i Indicates that the i-th Q learning model performs action a on the simulation circuit i Post-simulation circuit returns.
10. A DTCO circuit parameter adjustment system integrating deep learning and reinforcement learning, characterized in that: include: A data acquisition module, used to acquire a circuit data set, wherein the circuit data set includes a plurality of groups of circuit models, each group of circuit models corresponds to an electrical characteristic; A deep learning module is used to build a deep learning model, input a circuit data set into the deep learning model, train the deep learning model, and obtain a trained deep learning model; The reinforcement learning module is used to build a reinforcement learning model and perform simulation training on the reinforcement learning model. The simulation training includes: Inputting the simulation circuit parameters and the simulation results into the reinforcement learning model, the reinforcement learning model adjusts the simulation circuit according to the simulation circuit parameters and the simulation results, and optimizing the reinforcement learning model according to the loss of the simulation results of the simulation circuit after the parameter adjustment; Repeat the simulation training of the reinforcement learning model until the simulation result loss of the simulation circuit is less than a preset threshold, and then end the simulation training; The circuit parameter adjustment module is used to input electrical characteristic requirements into the trained deep learning model, and the trained deep learning model outputs a circuit model that meets the electrical characteristic requirements. The circuit model that meets the electrical characteristic requirements is input into the reinforcement learning model, and the reinforcement learning model outputs the adjusted circuit model to complete the parameter adjustment.