Radio frequency circuit schematic diagram-to-layout collaborative optimization method and system based on AI model

Abstract

The invention provides a radio frequency circuit schematic diagram-to-layout collaborative optimization method and system based on an AI model, and the method comprises the steps: building a mapping database from the geometric parameters of a passive device to corresponding electromagnetic characteristics through obtaining the geometric parameters of the passive device; based on a database, a corresponding AI agent model is trained for each passive device, through the AI agent models, millisecond-level model prediction is used for replacing time-consuming electromagnetic simulation to serve as design guidance, the calling frequency of calculation-intensive electromagnetic simulation is remarkably reduced, the layout parasitic effect is further considered in the schematic diagram design stage, and the design efficiency is improved. And a performance gap between a schematic diagram and a layout is effectively bridged. Then a target schematic diagram is obtained, passive devices in the target schematic diagram are automatically replaced based on a Bayesian optimization algorithm and an AI agent model, and the design period from the schematic diagram to a qualified layout is greatly shortened; by optimizing the layout effect in the schematic diagram design stage, it can be ensured that the simulation performance of the finally generated physical layout is consistent with a schematic diagram simulation result.

Description

Technical Field

[0001] This invention relates to the field of electronic design automation technology, and in particular to an artificial intelligence-based process design kit (PDK) modeling method and an automated collaborative optimization method and system for radio frequency circuits from schematic design to layout implementation. Background Technology

[0002] Electronic Design Automation (EDA) is the cornerstone of modern integrated circuit design. It utilizes Computer-Aided Design (CAD) tools to significantly improve the efficiency and standardization of circuit design. In the field of radio frequency (RF) circuit design, the design flow is typically divided into two core stages: schematic design and layout design.

[0003] During the schematic design phase, design engineers use ideal or simplified device models (such as lumped-parameter elements) for circuit design and simulation, aiming to quickly verify the circuit topology and achieve desired performance metrics, such as gain, noise figure, and linearity. However, these ideal models cannot reflect the true electromagnetic characteristics of the devices after physical implementation. Higher-precision layout design, on the other hand, requires more computational resources and time.

[0004] When a design transitions from schematic to layout, passive components in the circuit (such as inductors, capacitors, and transmission lines) are implemented as physical structures with specific geometries. These physical structures introduce complex parasitic effects, including parasitic capacitance and resistance within the components themselves, as well as electromagnetic coupling effects between components and between components and transmission lines. These parasitic effects are particularly pronounced in the radio frequency (RF) band, especially the millimeter-wave band, resulting in a significant gap between the "post-simulation" performance after layout completion and the "pre-simulation" performance of the schematic (schematic-layout discrepancy).

[0005] This performance gap has led to the following industry pain points:

[0006] 1. Numerous design iterations and long cycles: After the schematic design meets the requirements, the layout design often fails to meet the performance requirements. Engineers need to repeatedly modify the layout, perform time-consuming electromagnetic simulations, and even return to the schematic stage to redesign, which greatly extends the R&D cycle.

[0007] 2. Reliance on engineer experience: The ability to predict and compensate for layout parasitic effects is highly dependent on the experience of senior engineers, which makes RF design difficult and makes it hard to guarantee consistent design quality.

[0008] 3. Low optimization efficiency: Traditional post-simulation optimization is performed on the complete layout, and each simulation takes several hours or even days. With many optimization variables, it is difficult for automated optimization algorithms to converge within an acceptable time.

[0009] Therefore, how to bridge the design gap from schematic to layout, introduce the physical effects of the layout in the early stages of design, and achieve collaborative optimization between schematic and layout is a technical problem that urgently needs to be solved in the current EDA field. Summary of the Invention

[0010] To overcome the shortcomings of existing technologies, this invention provides a method and system for collaborative optimization of radio frequency circuit schematics to layouts based on AI models. The aim is to break down the design barriers between schematics and layouts, and improve the first-time success rate and automation level of radio frequency circuit design.

[0011] To achieve the above objectives, this invention provides a method for co-optimization of radio frequency circuit schematics to layouts based on AI models, which includes the following steps:

[0012] Obtain the geometric parameters of passive devices and establish a mapping database from the geometric parameters of passive devices to their corresponding electromagnetic characteristics (S-parameters);

[0013] Based on the database, a corresponding AI agent model is trained for each passive device; the model can quickly and accurately predict its corresponding S-parameters based on the input geometric parameters.

[0014] Obtain the target schematic diagram and replace the passive components in the schematic diagram one by one. During the replacement, use the Bayesian optimization algorithm with the AI ​​agent model as the core to search for the optimal layout geometry parameters to replace the passive components while maintaining the normal performance of the overall circuit. After all passive components have been replaced, output the set of geometry parameters used to generate the layout.

[0015] Preferably, when establishing the database, based on the passive components in the Target Process Design Kit (PDK), automated scripts are used to control EDA software to perform parametric scanning of the passive components. After electromagnetic (EM) simulation, the database is established. The specific steps are as follows:

[0016] Select a passive component from the target process design kit; such as a solenoid inductor or a planar capacitor.

[0017] Define its key geometric parameters; the geometric parameters of the solenoid inductor include: number of turns, inner diameter, line width, and line spacing; the geometric parameters of the capacitor include: plate length and width;

[0018] An automated script (such as a mouse script) is used to control the electromagnetic field circuit simulation module, which performs combined scanning of geometric parameters within a preset parameter range;

[0019] For each parameter combination, the device layout is automatically generated, simulation ports are added, and the electromagnetic simulation engine is called to perform S-parameter simulation to obtain the S-parameters (including real and imaginary parts) covering the target frequency band.

[0020] The combination of geometric parameters and its corresponding S-parameter simulation results are stored as a single record in the database.

[0021] Preferably, the AI ​​agent model is a neural network model, and its training steps specifically include:

[0022] Use the geometric parameters in the database as input data and the corresponding S parameters as label data;

[0023] Normalize the input data to improve the stability and convergence speed of model training;

[0024] Construct a multilayer perceptron (MLP) network, where the number of nodes in the input layer corresponds to the dimension of the geometric parameters, the number of nodes in the output layer corresponds to the total number of real and imaginary parts of the S-parameters for all target frequency points, and the intermediate layers are set with multiple hidden layers containing activation functions (such as ReLU or Tanh).

[0025] The mean squared error (MSE) is used as the loss function, and an optimizer (such as Adam) is used to iteratively update the network weights until the prediction error of the model on the validation set meets the preset accuracy requirements.

[0026] Preferably, the target schematic is a schematic circuit that has been designed in EDA software and has achieved the performance specifications.

[0027] Preferably, the steps for replacing passive components include:

[0028] Randomly select an ideal passive device (such as L1) from the target schematic diagram;

[0029] Replace the ideal passive device with an S-parameter file port, the S-parameter file being generated by its corresponding AI agent model;

[0030] Initiate the Bayesian optimization process, using the geometric parameters of the ideal device as the optimization variables;

[0031] The objective function of Bayesian optimization is set as: minimizing the difference between the performance index of the replaced circuit and the performance index of the original ideal circuit;

[0032] The Bayesian optimization algorithm forms a new combination of geometric parameters based on the objective function value, calls the AI ​​proxy model to generate S-parameters, and then performs circuit-level simulation until the iteration number or performance convergence condition is met.

[0033] The optimal combination of geometric parameters obtained from the Bayesian optimization search is solidified, and the electromagnetic layout simulation tool in the electromagnetic field circuit simulation module is called to perform electromagnetic simulation. The S-parameter file generated by the electromagnetic simulation of the optimal geometric parameters after solidification is used to permanently replace the ideal device in the schematic diagram.

[0034] Preferably, the objective function of Bayesian optimization is the weighted sum of squared errors of the key performance indicators (such as power, gain, return loss, noise figure, etc.) of the original circuit and the replaced circuit.

[0035] Preferably, the set of geometric parameters used to generate the layout is a list or file containing all replaced devices and their optimal layout geometric parameters, which is directly used to automatically generate the final circuit layout in the electromagnetic field simulation module.

[0036] This invention also provides a collaborative optimization system that uses an AI model-based radio frequency circuit schematic-to-layout collaborative optimization method for collaborative optimization, comprising:

[0037] The PDK data extraction module is used to obtain the geometric parameters of passive devices and establish a mapping database from the geometric parameters of passive devices to their corresponding electromagnetic properties.

[0038] The AI ​​agent model training module, based on a database, trains a corresponding AI agent model for each passive device.

[0039] The collaborative optimization and replacement module is used to obtain the target schematic and replace the passive components in the target schematic one by one. During the replacement, the Bayesian optimization algorithm is used with the AI ​​agent model as the core to search for the optimal layout geometry parameters to replace the passive components while maintaining the normal performance of the overall circuit. After all passive components have been replaced, the geometric parameter set used to generate the layout is output.

[0040] The AI-based method and system for co-optimizing RF circuit schematics to layouts provided by this invention have the following advantages compared to existing technologies:

[0041] Forward layout effect: By using an AI proxy model, millisecond-level model predictions are used instead of time-consuming electromagnetic simulations as design guidance, significantly reducing the number of calls to computationally intensive electromagnetic simulations. This allows for high-precision consideration of layout parasitic effects during the schematic design phase, effectively bridging the performance gap between schematics and layouts.

[0042] Improved design efficiency: Automated and intelligent component replacement processes transform the process that originally required engineers to make repeated manual adjustments and simulations into algorithm-driven collaborative optimization, which greatly shortens the design cycle from schematic to qualified layout.

[0043] Improved first-time success rate: Since the layout effect has been fully optimized in the schematic stage, the performance of the final generated layout simulation is highly consistent with the schematic simulation results, which significantly improves the first-time success rate of the design and reduces R&D costs and rework risks.

[0044] Lowering the design threshold: This method integrates the ability of senior engineers to predict parasitic effects into AI models and optimization algorithms, enabling engineers with less experience to design high-performance RF circuits. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall architecture of the AI-based radio frequency circuit schematic-to-layout collaborative optimization method and system provided by the present invention.

[0046] Figure 2 A structural diagram of an inductor provided as an example of the present invention;

[0047] Figure 3 A flowchart for constructing a database provided by this invention;

[0048] Figure 4 This is a schematic diagram of the structure of the AI ​​agent model (neural network) provided by the present invention;

[0049] Figure 5 A detailed flowchart of the schematic-to-layout co-optimization steps provided by this invention;

[0050] Figure 6 A schematic diagram illustrating the Bayesian optimization process for a single device provided by the present invention;

[0051] Figure 7 A schematic example (filter) of the schematic-to-layout co-optimization step provided by the present invention;

[0052] Figure 8 A layout example (filter) for the schematic-to-layout co-optimization step provided by this invention. Detailed Implementation

[0053] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0054] like Figure 1As shown, the present invention provides a method for co-optimization of radio frequency circuit schematic to layout based on AI model. The specific steps include: obtaining the geometric parameters of passive devices and establishing a mapping database of the geometric parameters of passive devices to their corresponding electromagnetic characteristics.

[0055] Based on the database, a corresponding AI agent model is trained for each passive device.

[0056] Obtain the target schematic diagram and replace the passive components in the schematic diagram one by one. During the replacement, use the Bayesian optimization algorithm with the AI ​​agent model as the core to search for the optimal layout geometry parameters to replace the passive components while maintaining the normal performance of the overall circuit. After all passive components have been replaced, output the set of geometry parameters used to generate the layout.

[0057] The system corresponding to this method includes: a PDK data extraction module (201), an AI agent model training module (202), and a collaborative optimization and replacement module (204).

[0058] To better explain and verify the feasibility of this method, the present invention provides the following embodiments, as detailed below:

[0059] Example 1: Lightweighting of PDK passive devices and database construction.

[0060] The goal of this step is to take a 0.25μm GaAs pHEMT process PDK as an example and select passive devices (such as a spiral inductor, for example). Figure 2 As shown, a "Geometric Parameters - S-Parameters" database is established, and 1000 geometric parameters are sampled.

[0061] like Figure 3 As shown, the specific process is as follows:

[0062] 1.1: Determining the Target Device and Parameters. A spiral inductor from a 0.25μm GaAs pHEMT process PDK was selected as the target. Its key geometric parameters were defined as follows:

[0063] a) Number of turns N: Integer, ranging from 2 to 6 turns, in increments of 1;

[0064] b) Inner diameter D: Continuous variable, ranging from 60μm to 120μm, with a step size of 5μm;

[0065] c) Line width W: a continuous variable, ranging from 6μm to 14μm, with a step size of 1μm;

[0066] d) Line spacing S: a continuous variable, ranging from 2μm to 8μm, with a step size of 1μm;

[0067] 1.2: Automated Parameter Scanning. An automated script (using a mouse and keyboard control script implemented with the Python-based PyAutoGUI library as an example) interacts with the electromagnetic field simulation module. This module can utilize software such as ADS, Cadence, Virtuoso, Platform, AWR, Microwave, and Office for electromagnetic field simulation; no specific limitations are specified here.

[0068] 1.2.1. Parameter combination for generating spiral inductance [W,D,S,N];

[0069] 1.2.2. Call the automation script and execute it sequentially:

[0070] a) Create a new Layout Cell named "Ind_Nx_Dxx_Wxx_Sxx" and open or activate the Layout Cell;

[0071] b) Instantiate the spiral inductor and modify it to the parameter combination [W,D,S,N];

[0072] c) Automatically create the Momentum simulation controller "EM Setup";

[0073] d) Set the frequency range: 0.1GHz~20GHz, linear scan, step 0.1GHz, total 201 points;

[0074] e) Submit the simulation task;

[0075] f) After the simulation is completed, extract the S-parameters, parse the real / imaginary parts, and the key name format is (W,D,S,N,frequency point,S11 real part,S11 imaginary part,S12 real part,S12 imaginary part,S21 real part,S21 imaginary part,S22 real part,S22 imaginary part).

[0076] g) Submit the parsed data to the database.

[0077] 1.2.3. If the data size is met, the "Geometric Parameters - S-Parameters" database construction is complete; otherwise, re-execute steps 1.2.1 to 1.2.2.

[0078] 1.3: Output database file (using HDF5 as an example).

[0079] Example 2: AI agent model training.

[0080] This step utilizes the database established in Example 1 to train a fast prediction model for the solenoid inductor.

[0081] like Figure 4 As shown, the specific process is as follows:

[0082] 2.1: Data preprocessing.

[0083] 2.1.1 Read the HDF5 file obtained in Example 1 and generate a DataFrame;

[0084] 2.1.2 Min-Max Scaling is performed on all continuous input features (D, W, S, freq) and all output labels (S-parameters) to scale the values ​​to the [0, 1] or [-1, 1] interval. For discrete turns N, one-hot encoding can be used.

[0085] a) Input features X: W, D, S (normalized to [0,1]), N (One-Hot 3D), freq (frequency, in GHz);

[0086] b) Output label Y: 4×2=8-dimensional vector (S11 real part, S11 imaginary part, S12 real part, S12 imaginary part, S21 real part, S21 imaginary part, S22 real part, S22 imaginary part);

[0087] 2.2: Constructing the Neural Network. A multilayer perceptron (MLP) structure is used. The specific structure is set as follows:

[0088] 2.2.1. Input layer (401): The number of nodes is equal to the processed geometric parameter dimension 3 + 3 + 1 = 7.

[0089] 2.2.2. Hidden layer (402):

[0090] a) First hidden layer: 256 nodes, ReLU activated, Dropout 0.1;

[0091] b) Second hidden layer: 256 nodes, activate ReLU, BatchNorm;

[0092] c) Third hidden layer: 128 nodes, activate ReLU;

[0093] 2.2.3. Output Layer (403): The number of nodes is equal to the sum of the real and imaginary parts of all S-parameters. For example, for simulation of a two-port passive device, the number of output nodes is 4*2=8.

[0094] 2.3: Model Training

[0095] 2.3.1. Divide the dataset into a training set (80%), a validation set (10%), and a test set (10%). Use the Adam optimizer and the mean squared error (MSE) loss function.

[0096] 2.3.2. Perform multiple epochs of training on the training set, and evaluate the model performance on the validation set after each epoch. Stop training when the loss on the validation set no longer decreases significantly to prevent overfitting.

[0097] 2.3.3. After training, the AI ​​agent model (222) can receive a set of geometric parameters and predict its complete S-parameters in milliseconds.

[0098] Example 3: Collaborative optimization and replacement from schematic to layout.

[0099] This step is the core application stage of the method of the present invention. This embodiment takes a bandpass filter operating in the 1.3GHz to 2.3GHz band as the optimization object, aiming to efficiently and accurately transform an ideal component schematic that meets the performance requirements into a physical layout with consistent performance that can be directly manufactured.

[0100] like Figure 5 , 6 As shown, the specific process is as follows:

[0101] 3.1: Input, Initialization, and Optimization Goal Definition:

[0102] Input an ideal schematic diagram of a bandpass filter that has already been designed in the electromagnetic field circuit simulation module, such as... Figure 7 As shown in the diagram, this schematic uses ideal lumped inductor (L) and capacitor (C) components. Through ideal simulation, its performance specifications meet the design requirements, and this performance will serve as a baseline for subsequent optimization.

[0103] Passband: 1.3GHz~2.3GHz;

[0104] Insertion loss S21 > -3dB;

[0105] Input return loss S11 < -15dB;

[0106] Lower stopband: DC ~ 0.7 GHz;

[0107] Out-of-band suppression S21 < -40 dB;

[0108] Upper stopband: 2.8GHz~10GHz;

[0109] Out-of-band suppression S21 < -40 dB.

[0110] 3.1.2. Target Device Identification:

[0111] The system automatically parses the input schematic netlist, identifying all passive components that need to be replaced in the layout and their ideal values, such as inductors L1, L2,... and capacitors C1, C2,... Each component will have an AI proxy model trained in Example 2.

[0112] 3.1.3. Quantitative definition of the optimization objective function (Cost Function):

[0113] To guide the optimization process, a target function f(x) needs to be constructed that can accurately quantify the "degree of performance degradation," where x represents the geometric parameter vector of the device being optimized. This function aims to penalize any deviation from the baseline performance. In this embodiment, f(x) is defined as the weighted sum of squares of violations of performance metrics across key frequency bands:

[0114] f(x) = w p21 *Cost p21 (x)+w p11 *Cost p11 (x)+w s21 *Cost s21 (x),

[0115] in:

[0116] Passband insertion loss penalty Cost p21 (x):

[0117] Cost p21 (x)=max(0,(min(S p21 )-(-3)) 2 ),

[0118] This penalty applies to frequencies where insertion loss is worse than -3dB within the passband (1.3GHz to 2.3GHz). p21 S21 (dB) is within the passband.

[0119] Passband return loss penalty Cost p11 (x):

[0120] Cost p11 (x)=max(0,(max(S p21 )-(-15)) 2 ),

[0121] This penalty applies to frequencies where insertion loss is worse than -15dB within the passband (1.3GHz to 2.3GHz). p11 S11 (dB) is within the passband.

[0122] Stopband suppression penalty Cost s21 (x):

[0123] Cost s21 (x)=max(0,(max(S s21 )-(-30)) 2 ),

[0124] This penalty applies to frequencies where insertion loss is worse than -30dB within the stopband (DC~0.7GHz, 2.8~10GHz). s21 S21 (dB) is within the stopband.

[0125] w p21 ,w p11 ,w s21 , which is a weighting coefficient used to adjust the optimization priority of different performance indicators.

[0126] 3.2: Iterative Replacement Process

[0127] The identified passive devices are replaced one by one in a context-aware manner.

[0128] a. Component Selection: Randomly select an ideal inductor L1 from the schematic diagram for replacement. In the schematic design of the electromagnetic field circuit simulation module, the ideal inductor L1 is replaced with an S2P (2-port S-parameter) component, whose data will be dynamically generated by the AI ​​proxy model.

[0129] b. Initiate Bayesian Optimization: For L1, initiate the Bayesian Optimization (BO) process (e.g.) Figure 6 (As shown).

[0130] Optimization variable x: L1's geometric parameter vector x = [N, D, W, S].

[0131] Detailed explanation of Bayesian optimization principles:

[0132] Bayesian optimization is a sequential optimization strategy based on a probabilistic surrogate model. It is particularly suitable for scenarios where evaluating f(x) is costly. In this invention, evaluating f(x) once requires one "AI model prediction + circuit-level simulation," which, although much faster than electromagnetic simulation, is still a huge overhead for traditional algorithms that require tens of thousands of iterations. BO intelligently selects a small number of information-rich points for evaluation, finding the global optimum with extremely high data efficiency.

[0133] c. Solidify the optimal solution: Optimization terminates when BO reaches a preset number of iterations (e.g., 50-100 times) or the expected value of EI falls below a certain threshold. At this point, the geometric parameter x* with the lowest f(x) value among all observed points is considered the optimal solution.

[0134] d. Use the layout field simulation tool of the electromagnetic field simulation module to perform electromagnetic simulation on the passive device corresponding to this set of optimal parameters. Permanently replace L1 in the filter schematic with the S-parameter file of the electromagnetic simulation. Verify the final electromagnetic simulation to minimize the accumulation of neural network errors.

[0135] e. Continue iterating: After locking L1, select the next passive device (such as capacitor C1) and repeat the replacement optimization process of ac. The key is that when optimizing C1, L1 is already its layout-based S-parameter model that includes parasitic effects. This means that the optimization of C1 is carried out in a context that is closer to the real physical circuit, thus ensuring the accuracy of the optimization and the reliability of the final result.

[0136] 3.3: Output the final result.

[0137] Once all target passive components (all L1 and C1) have been replaced one by one, the schematic of the entire filter is fully "layouted". At this point, the system outputs a list or script file containing all components (L1, L2, C1, ...) and their corresponding optimal layout geometry parameters. Design engineers can use this list to generate, with one click in the electromagnetic field simulation module or automatically via script, the final, co-optimized circuit layout (e.g., ...). Figure 8 (As shown). Since the parasitic effects of the layout have been fully considered and compensated for during the design phase using the method of this invention, its final electromagnetic post-simulation performance will be highly consistent with the performance of the optimized schematic, thereby greatly improving the success rate of the first design and significantly shortening the R&D cycle.

[0138] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. A method for collaborative optimization of RF circuit schematic to layout based on AI models, characterized in that, Includes the following steps: Obtain the geometric parameters of passive devices and establish a mapping database from the geometric parameters of passive devices to their corresponding electromagnetic properties; Based on the database, a corresponding AI agent model is trained for each passive device. Obtain the target schematic diagram and replace the passive components in the schematic diagram one by one. During the replacement, use the Bayesian optimization algorithm with the AI ​​agent model as the core to search for the optimal layout geometry parameters to replace the passive components while maintaining the normal performance of the overall circuit. After all passive components have been replaced, output the set of geometry parameters used to generate the layout.

2. The method for collaborative optimization of RF circuit schematic to layout based on AI model according to claim 1, characterized in that, When establishing the database, based on the passive components in the target process design kit, automated scripts are used to control EDA software to perform parametric scanning of the passive components. After electromagnetic simulation, the database is established. The specific steps are as follows: Select one passive device from the target process design kit; Define its key geometric parameters; An automated script is used to control the electromagnetic field circuit simulation module, which performs combined scanning of geometric parameters within a preset parameter range; For each parameter combination, the device layout is automatically generated, simulation ports are added, and the electromagnetic simulation engine is called to perform S-parameter simulation to obtain the S-parameters covering the target frequency band. The combination of geometric parameters and its corresponding S-parameter simulation results are stored as a single record in the database.

3. The method for collaborative optimization of RF circuit schematic to layout based on AI model according to claim 1, characterized in that, The AI ​​agent model is a neural network model, and its training steps specifically include: Use the geometric parameters in the database as input data and the corresponding S parameters as label data; Normalize the input data; Construct a multilayer perceptron network, where the number of nodes in the input layer corresponds to the dimension of the geometric parameters, the number of nodes in the output layer corresponds to the total number of real and imaginary parts of the S-parameters for all target frequency points, and multiple hidden layers containing activation functions are set in the intermediate layers. The mean squared error is used as the loss function, and the optimizer is used to iteratively update the network weights until the prediction error of the model on the validation set meets the preset accuracy requirements.

4. The AI-based schematic-to-layout co-optimization method for radio frequency circuits according to claim 1, characterized in that, The target schematic diagram is a schematic circuit that has been designed and achieved its performance specifications in EDA software.

5. The method for collaborative optimization of RF circuit schematic to layout based on an AI model according to claim 1, characterized in that, The steps for replacing passive components include: Randomly select an ideal passive device from the target schematic; The ideal passive device is replaced with an S-parameter file port, the S-parameter file being generated by its corresponding AI agent model; Initiate the Bayesian optimization process, using the geometric parameters of the ideal device as the optimization variables; The objective function of Bayesian optimization is set as: minimizing the difference between the performance index of the replaced circuit and the performance index of the original ideal circuit; The Bayesian optimization algorithm forms a new combination of geometric parameters based on the objective function value, calls the AI ​​proxy model to generate S-parameters, and then performs circuit-level simulation until the iteration number or performance convergence condition is met. The optimal geometric parameter combination obtained from the Bayesian optimization search is solidified, and the electromagnetic layout simulation tool in the electromagnetic field simulation module is called to perform electromagnetic simulation. The S-parameter file generated by the optimal geometric parameters of the electromagnetic simulation after the combination is solidified is used to permanently replace the ideal device in the schematic diagram.

6. The AI-based schematic-to-layout co-optimization method for radio frequency circuits according to claim 5, characterized in that, The objective function of the Bayesian optimization is the weighted sum of squared errors of the key performance indicators of the original circuit and the replaced circuit.

7. The AI-based schematic-to-layout co-optimization method for radio frequency circuits according to claim 5, characterized in that, The set of geometric parameters used to generate the layout is a list or file containing all replaced devices and their optimal layout geometric parameters, which is directly used to automatically generate the final circuit layout in the electromagnetic field simulation module.

8. A collaborative optimization system, employing the AI-based radio frequency circuit schematic-to-layout collaborative optimization method according to any one of claims 1 to 7 for collaborative optimization, characterized in that, include: The PDK data extraction module is used to obtain the geometric parameters of passive devices and establish a mapping database from the geometric parameters of passive devices to their corresponding electromagnetic properties. The AI ​​agent model training module, based on a database, trains a corresponding AI agent model for each passive device. The collaborative optimization and replacement module is used to obtain the target schematic and replace the passive components in the target schematic one by one. During the replacement, the Bayesian optimization algorithm is used with the AI ​​agent model as the core to search for the optimal layout geometry parameters to replace the passive components while maintaining the normal performance of the overall circuit. After all passive components have been replaced, the geometric parameter set used to generate the layout is output.

Images