Sectional matching circuit automatic design method and system based on Bayesian optimization and sectional variable geometric parameter radio frequency transformer

Abstract

The invention discloses a Bayesian optimization-based segmented matching circuit automatic design method and system and a segmented variable geometric parameter radio frequency transformer. The radio-frequency transformer comprises a main coil and a secondary coil, the main coil and the secondary coil are sectional type coils with variable geometric parameters, the opening direction of the main coil and the opening direction of the secondary coil form an angle of 180 degrees, and the central axes of the main coil and the secondary coil coincide; the main coil is located on a top metal layer of the radio frequency transformer, the secondary coil is located on the next layer of the top metal layer, or the main coil is located on the ninth layer, and the secondary coil is located on the eighth layer. According to the method, the transformer matching circuit with the performance exceeding that of a traditional structure can be designed, meanwhile, the design period is shortened to be within several hours from several days, and the dependence on the personal experience of engineers is reduced.

Description

Technical Field

[0001] This invention relates to the field of radio frequency integrated circuit design technology, and in particular to an automated design method and system for segmented matching circuits based on Bayesian optimization, as well as a segmented radio frequency transformer with variable geometric parameters. Background Technology

[0002] In the design and manufacture of modern high-performance radio frequency integrated circuits (RFICs), the performance of passive components has a significant impact on key system parameters such as power, efficiency, and operating bandwidth. On-chip transformers are a common type of passive component, frequently used in the design of broadband matching circuits and signal differential conversion. In traditional design flows, senior engineers typically first design an initial structure based on experience, then simulate and observe the performance data. Next, engineers manually adjust some dimensions in the graphical interface of electromagnetic simulation software, or select a few design variables and use the built-in algorithms of the electromagnetic simulation software for automatic iterative optimization. Both of these methods have inherent limitations.

[0003] Manual adjustments are limited by the engineers' design capabilities. Since there is no precise and universally applicable formula for calculating the dimensions and performance indicators of transformer matching circuits, engineers can only improve matching performance based on their experience and continuous experimentation. When considering improvement directions, engineers can usually only weigh a few key variables simultaneously, unable to systematically think about and evaluate a complex structure with dozens of independent geometric variables, let alone find the optimal solution in such a high-dimensional space. Once engineers find a barely satisfactory design, they often stop further exploration, even though this solution is, in most cases, only a local optimum in the vast design space, leading to unnecessary compromises in the chip's final performance and power consumption. Furthermore, design experience highly relies on the engineer's personal "feel" and memory, and this valuable knowledge is difficult to quantify, solidify, and pass on. This requires new engineers to undergo a lengthy learning process to master this "skill," resulting in project progress and quality being highly dependent on the availability of a few senior experts.

[0004] The approach of using built-in algorithms in electromagnetic industry software for automatic iterative optimization is constrained by the dimensionality of variables and the efficiency of the algorithms. Specifically, the spiral coils of traditional transformers are designed with symmetrical and regular geometric structures, such as standard circles or octagons, with uniform linewidths. This results in very few design variables available for optimization, leading to a very narrow design space. For octagonal inductors, there are typically only a few degrees of freedom, such as coil radius, linewidth, and spacing. The scarcity of design variables makes it impossible for optimization algorithms to perform fine-tuning of current density, parasitic capacitance, and inductance at different locations of the coil, and it is difficult to effectively compensate for high-frequency non-ideal effects such as the skin effect and proximity effect. As a result, the final design always involves a rough, global compromise between various performance indicators. On the other hand, the optimization algorithms in existing EDA tools are mainly traditional algorithms, such as stochastic methods and gradient methods. Among them, stochastic methods lack learning mechanisms and blindly search in optimization problems, resulting in extremely low efficiency; gradient methods are prone to getting trapped in local optima and are very sensitive to initial values. These traditional optimization algorithms themselves have serious limitations and are completely unable to cope with more complex, high-dimensional design explorations.

[0005] Therefore, existing technologies urgently need a completely new transformer structure that can break through performance bottlenecks at the physical level, as well as an advanced design method that can handle such a complex structure. Summary of the Invention

[0006] The purpose of this invention is to provide an automated design method and system for segmented matching circuits based on Bayesian optimization, as well as a segmented radio frequency transformer with variable geometric parameters, to overcome the limitations of existing design schemes in terms of engineer design capabilities, variable dimensions, and algorithm efficiency. The numerous technical effects of the preferred solutions among the many technical solutions provided by this invention are detailed below.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an automated design method for piecewise matching circuits based on Bayesian optimization, comprising the following steps: S100. Divide the components in the matching circuit into several consecutive segments according to their geometry, and set corresponding geometric parameters for each segment. S200, Do the geometric parameters of each segment have historical data? S300. If so, load historical data, generate a candidate value for the geometric parameters of each segment based on the historical data, and execute step S600. S400 If not, generate n initial values ​​for the geometric parameters of each segment, and evaluate each initial value corresponding to the component in turn; S500: Generate a candidate value for the geometric parameters of each segment based on the evaluation data. S600. When each candidate value satisfies the constraint range of the corresponding segment, a three-dimensional model of the component is constructed based on all candidate values. A full-wave electromagnetic simulation is performed based on the three-dimensional model to generate a scattering matrix. Circuit-level simulation is performed based on the scattering matrix, and the error between the simulation value and the input target variable value is recorded. When at least one candidate value does not satisfy the constraint range of the corresponding segment, a penalty value is recorded. S700. When the termination condition is met, output the evaluation result; otherwise, generate a candidate value for the geometric parameters of each segment based on historical data or the data corresponding to the evaluation and return to step S600. In this process, a candidate value is generated for the geometric parameters of each segment using a Bayesian optimization algorithm.

[0008] In one or more embodiments, when performing circuit-level simulation based on the scattering matrix, the method further includes: The differential signal is converted into a single-ended signal by a balun. One end is connected to a load with a set resistance value, and the real and imaginary parts of the input impedance at the other end are simulated.

[0009] In one or more embodiments, n is twice the total number of geometric parameters corresponding to all segments of the component.

[0010] In one or more embodiments, if the candidate or initial values ​​of the generated geometric parameters do not meet the corresponding constraint range, the set penalty value is fed back to the Bayesian optimization algorithm.

[0011] In one or more embodiments, each initial value corresponding to a component is evaluated sequentially, including the following steps: Each segment of the component is assigned an initial value corresponding to a geometric parameter, and a total of n initial values ​​are assigned. Does each initial value satisfy the constraint range of the corresponding segment? Thus, a three-dimensional model of the component is constructed based on each set of initial values. A full-wave electromagnetic simulation is performed based on the three-dimensional model to generate a scattering matrix. Circuit-level simulation is performed based on the scattering matrix. The error between the simulation value and the input target variable value is recorded to obtain the evaluation data for each set of initial values. Otherwise, record the penalty value to obtain the evaluation data for each set of initial values.

[0012] In one or more embodiments, the components include a transformer, the main coil and secondary coil of the transformer having an octagonal coil body, each coil being symmetrical about the central axis; each coil is divided into 11 consecutive segments and numbered sequentially, the width and length of the first to fifth segments corresponding to each number, the width of the sixth segment, and the center distance between the main coil and the secondary coil are used as a geometric parameter to form 23 geometric parameters of the transformer.

[0013] According to another aspect of the present invention, a segmented matching circuit automated design system based on Bayesian optimization is also provided, which is applied to the segmented matching circuit automated design method based on Bayesian optimization described above, including a central control module, a component segmentation and variable constraint module, a Bayesian optimization module, a three-dimensional electromagnetic simulator, and a circuit analysis module. The central control module is used for process scheduling, data processing, intelligent decision-making, and communication with the simulator in the automated design method of segmented matching circuits based on Bayesian optimization. The component segmentation and variable constraint module is used to divide the components in the matching circuit into several continuous segments according to their geometry, set corresponding geometric parameters for each segment, and transmit the segmentation results to the central control module; it is also used to configure the corresponding constraint range for each segment and accept the input target variable value for scheduling by the central control module. The Bayesian optimization module, under the control of the central control module, determines whether historical data exists for the geometric parameters of each segment. If historical data exists, it loads the historical data and uses the Bayesian optimization algorithm to generate a candidate value for each segment's geometric parameters. If no historical data exists, it generates n initial values ​​for each segment's geometric parameters, evaluates each initial value corresponding to the component in turn, and uses the Bayesian optimization algorithm to generate candidate values ​​for each segment's geometric parameters based on the evaluation data. It then checks whether each candidate value satisfies the constraint range of the corresponding segment. When at least one candidate value does not satisfy the constraint, a penalty value is recorded, and the evaluation result is obtained. The 3D electromagnetic simulator constructs a 3D model of the component based on all candidate values ​​when each candidate value is satisfied, and then performs full-wave electromagnetic simulation based on the 3D model to generate a scattering matrix. The circuit analysis module is used to perform circuit-level simulation based on the scattering matrix, record the error between the simulation value and the input target variable value, and obtain the evaluation result.

[0014] According to another aspect of the present invention, a segmented variable geometric parameter coil is also provided, which is designed according to the segmented matching circuit automated design system based on Bayesian optimization described above. The coil includes taps and a coil body. The coil body includes a first metal wire, a second metal wire, a third metal wire, a fourth metal wire, a fifth metal wire, a sixth metal wire, a seventh metal wire, an eighth metal wire, a ninth metal wire, a tenth metal wire and a last metal wire connected in sequence. The second, third, fourth, fifth, sixth, seventh, eighth, ninth, and tenth metal wires form an octagonal structure, with the first and last metal wires forming an opening. The first and last metal wires are differential signal input and output lines, with the taps leading out from the sixth metal wire.

[0015] According to another aspect of the present invention, a segmented variable geometry radio frequency transformer is also provided, including a main coil and a secondary coil. Both the main coil and the secondary coil are segmented variable geometry coils as described above. The opening direction of the main coil is 180 degrees different from the opening direction of the secondary coil, and their central axes coincide. The main coil is located on the top metal layer of the RF transformer, and the secondary coil is located on the layer below the top metal layer; or the main coil is located on the ninth layer, and the secondary coil is located on the eighth layer.

[0016] In one or more embodiments, the segmented variable geometry radio frequency transformer is characterized in that it further includes a ground plane placed around the transformer matching circuit, the distance between the ground plane and the coil being adjustable; when the main coil is located on the ninth layer and the secondary coil is located on the eighth layer, the ground plane is placed on the fourth layer.

[0017] Implementing one of the above-described technical solutions of the present invention has the following advantages or beneficial effects: This invention proposes a segmented transformer matching network model with variable geometric parameters. The coil is decomposed into multiple geometric segments, each with independently variable line width and length, thereby increasing the design degrees of freedom from single digits to dozens. A built-in geometric topology constraint system is established to support this complex structure, ensuring its physical realizability under parameter variations. This invention employs a high-sample-efficiency Bayesian optimization algorithm and deeply integrates it with the constraint system and simulation software to construct a fully automated intelligent design method, efficiently exploring the vast design space brought about by the new structure.

[0018] This invention enables the design of transformer matching circuits with performance exceeding that of traditional structures, while reducing the design cycle from several days to within several hours and decreasing reliance on the personal experience of engineers. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart of an automated design method for segmented matching circuits based on Bayesian optimization, according to an embodiment of the present invention. Figure 2 This is a structural block diagram of a segmented matching circuit automated design system based on Bayesian optimization, according to an embodiment of the invention. Figure 3(a) is a schematic diagram showing how the absolute error of the real and imaginary parts of the input impedance changes with the number of iterations in an embodiment of the invention; Figure 3(b) is a two-dimensional error distribution diagram automatically generated after the optimization of the embodiment of the invention; Figure 3(c) is a schematic diagram of the final output of the central control module in the embodiment of the invention; Figure 4 This is a schematic diagram of a segmented coil structure with variable geometric parameters according to an embodiment of the invention; Figure 5 This is a three-dimensional view of a segmented radio frequency transformer structure with variable geometric parameters according to an embodiment of the invention (other layers are hidden).

[0020] In the diagram: 1. Coil body; 10. First metal wire segment; 11. Second metal wire segment; 12. Third metal wire segment; 13. Fourth metal wire segment; 14. Fifth metal wire segment; 15. Sixth metal wire segment; 16. Seventh metal wire segment; 17. Eighth metal wire segment; 18. Ninth metal wire segment; 19. Tenth metal wire segment; 20. Last metal wire segment; 2. Tap; 3. Opening; 4. Main coil; 5. Secondary coil; 6. Ground plane. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, various exemplary embodiments described below will be referenced to the accompanying drawings, which form part of the exemplary embodiments, illustrating various exemplary embodiments that may be used to implement the present invention. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. It should be understood that they are merely examples of processes, methods, and apparatuses consistent with some aspects of the present invention disclosed as detailed in the appended claims, and other embodiments may be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and spirit of the present invention.

[0022] In the description of this invention, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. The term "a plurality of" means two or more. The term "and / or" includes any and all combinations of one or more of the associated listed items. Those skilled in the art will be able to understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0023] To illustrate the technical solution described in this invention, specific embodiments are described below, showing only the parts related to the embodiments of this invention.

[0024] Example 1: like Figure 1As shown, this invention provides an automated design method for piecewise matching circuits based on Bayesian optimization, comprising the following steps: S100. Divide the components in the matching circuit into several consecutive segments according to their geometry, and set corresponding geometric parameters for each segment.

[0025] In one or more embodiments, the components include a transformer, wherein the main coil and secondary coil of the transformer have octagonal coil bodies.

[0026] Understandably, in the 65nm process commonly used in the radio frequency (RF) field, design rules require all transmission lines to be horizontal, vertical, or tilted at 45 degrees, thus making it impossible to design circular coils. This embodiment employs an octagonal inductor coil design, which complies with the design rules and has good compatibility with other processes. The 65nm process, commonly used in the RF field, is a semiconductor manufacturing process using CMOS technology that supports high-frequency applications. Its transistor gate minimum linewidth is 65nm, featuring high integration and low power consumption, making it suitable for wireless communication, radar, and other scenarios.

[0027] Furthermore, each coil is divided into 11 consecutive segments and numbered sequentially (see Example 3 for specific numbering). The width and length of the first to fifth segments corresponding to each coil, the width of the sixth segment, and the center distance between the main coil and the secondary coil are used as a geometric parameter to form 23 geometric parameters of the transformer.

[0028] Understandably, each coil is symmetrical about the central axis, meaning the first (first segment in Example 3) to fifth segments are the same as the eleventh (last segment in Example 3) to seventh segments. The spacing between the first and last metal segments is adjustable, but is fixed at 20µm by default. The length of the sixth segment is not adjustable because it can be calculated from the lengths of the first five segments and the spacing between the first and last segments. Furthermore, the distance between the centers of the two coils is also an important independent optimization variable; therefore, the entire transformer has a total of 23 independent optimization variables.

[0029] After completing the geometric segmentation and parameter setting of the components, users can input their design requirements and optimization strategies, including the range requirements for each geometric parameter, the target variable value, and the total number of evaluations of the optimization algorithm, the penalty value for constraint or simulation failure, etc., according to actual needs.

[0030] Furthermore, some combinations of the 23 independent optimization parameters do not meet the design requirements and must be eliminated. The corresponding constraints include: a. DRC (Design Rule Check) requirements: The design rules of the process impose requirements on the linewidth of each line segment. For example, the 65nm process requires the linewidth to be within the range of 2~12um.

[0031] b. Miniaturization requirements: Integrated circuits are very expensive to manufacture, and the device area should be as small as possible. Therefore, it is necessary to set an upper limit on the length of each metal line segment. In this embodiment, the upper limit is 60µm.

[0032] c. Topology Requirements: When a metal wire segment is short, if the segment or the segments on either side are too wide, the segments cannot be properly connected, leading to a topology error. To avoid this problem, this solution establishes a built-in geometric topology constraint system for this complex structure, consisting of 20 constraint conditions, to ensure the physical realizability of the transformer under parameter variations.

[0033] In a specific embodiment, the following are 20 constraints: -length12-width12 / 2+width13 / sqrt(2)-width11 / 2<0 -length13 / sqrt(2)+width14 / 2-width13 / sqrt(2)+width12 / 2<0 -width14-length14+width15 / sqrt(2)+width13 / sqrt(2)<0 -length15 / sqrt(2)+width16 / 2-width15 / sqrt(2)+width14 / 2<0 -gap1 / 2-length12-length13 / sqrt(2)+length15 / sqrt(2)-width16 / 2+width15 / sqrt(2)<0 -gap1 / 2-length12-length13 / sqrt(2)+length15 / sqrt(2)+width16 / 2-width15 / sqrt(2)<0 width14 / 2-width15 / sqrt(2)+length15 / sqrt(2)+width16 / 2>0 -width14+width13 / sqrt(2)+width15 / sqrt(2)+length14>0 width12 / 2+length13 / sqrt(2)+width14 / 2-width13 / sqrt(2)>0 -width11 / 2+length12-width12 / 2+width13 / sqrt(2)>0 -length22-width22 / 2+width23 / sqrt(2)-width21 / 2<0 -length23 / sqrt(2)+width24 / 2-width23 / sqrt(2)+width22 / 2<0 -width24-length24+width25 / sqrt(2)+width23 / sqrt(2)<0 -length25 / sqrt(2)+width26 / 2-width25 / sqrt(2)+width24 / 2<0 -gap2 / 2-length22-length23 / sqrt(2)+length25 / sqrt(2)-width26 / 2+width25 / sqrt(2)<0 -gap2 / 2-length22-length23 / sqrt(2)+length25 / sqrt(2)+width26 / 2-width25 / sqrt(2)<0 width24 / 2-width25 / sqrt(2)+length25 / sqrt(2)+width26 / 2>0 -width24+width23 / sqrt(2)+width25 / sqrt(2)+length24>0 width22 / 2+length23 / sqrt(2)+width24 / 2-width23 / sqrt(2)>0 -width21 / 2+length22-width22 / 2+width23 / sqrt(2)>0 (1); Wherein, width11 refers to the width of the first segment of the primary coil, length12 refers to the length of the second segment of the primary coil, width21 refers to the width of the first segment of the secondary coil, length23 refers to the length of the third segment of the secondary coil, and so on; gap1 refers to the distance between the centers of the two ports of the primary coil, gap2 refers to the distance between the centers of the two ports of the secondary coil, and sqrt(2) refers to the square root of 2.

[0034] Therefore, the entire transformer has 23 independent optimization parameters. This means that transformer design is no longer about adjusting a few macroscopic parameters, but rather about fine-tuning every part of the structure. This design philosophy exponentially expands the total design freedom and greatly broadens the optimization space.

[0035] It is understood that this embodiment illustrates an octagonal structure, but the "segmentation" concept of this embodiment can be applied to any topology, such as interleaved transformers, stacked multilayer transformers, or even asymmetric or arbitrary-shaped transformers.

[0036] S200, Do the geometric parameters of each segment have historical data?

[0037] If available historical simulation records, this method will automatically use these data for initial evaluation, enabling "continued testing" and saving time.

[0038] S300. If so, load historical data; based on the historical data, use the Bayesian optimization algorithm to generate a candidate value for the geometric parameters of each segment, and execute step S600.

[0039] In this step, the Bayesian optimization algorithm intelligently generates a new set of the most promising candidate geometric parameters based on existing data, and then iterates repeatedly until the preset number of evaluations or error convergence is reached.

[0040] The above group consists of the geometric parameters corresponding to all segments. For example, each of the 23 geometric parameters of the electromagnetic coil has a candidate value, and the 23 candidate values ​​form a group.

[0041] S400 If not, generate n initial values ​​for the geometric parameters of each segment, and evaluate each initial value corresponding to the component in turn.

[0042] Based on the above embodiments, 46 sets of design variables can be generated in a quasi-random manner (Sobol sequence) and prepared for simulation and evaluation.

[0043] In one or more embodiments, each initial value corresponding to a component is evaluated sequentially, including the following steps: Each segment of the component is assigned an initial value corresponding to a geometric parameter, and a total of n initial values ​​are assigned. Does each initial value satisfy the constraint range of the corresponding segment? Thus, a three-dimensional model of the component is constructed based on each set of initial values. A full-wave electromagnetic simulation is performed based on the three-dimensional model to generate a scattering matrix. Circuit-level simulation is performed based on the scattering matrix. The error between the simulation value and the input target variable value is recorded to obtain the evaluation data for each set of initial values. Otherwise, record the penalty value to obtain the evaluation data for each set of initial values.

[0044] Furthermore, n is twice the total number of geometric parameters corresponding to all segments of the component. Correspondingly, 23 geometric parameters generate 46 initial values ​​(i.e., 46 initial points as described below), and the transformer corresponds to 2 sets of initial values.

[0045] It should be noted that, according to the rules of thumb for Bayesian optimization algorithms, the initial number of points is usually set to twice the number of optimization variables. This helps the model to have a preliminary, general understanding of the entire design space. In this embodiment, there are 23 optimization geometric parameters, so the initial number of points is set to 46.

[0046] S500. Based on the evaluation data, a candidate value is generated for the geometric parameters of each segment using a Bayesian optimization algorithm.

[0047] S600. When each candidate value satisfies the constraint range of the corresponding segment, a three-dimensional model of the component is constructed based on all candidate values. A full-wave electromagnetic simulation is performed based on the three-dimensional model to generate a scattering matrix. Circuit-level simulation is performed based on the scattering matrix, and the error between the simulation value and the input target variable value is recorded. When at least one candidate value does not satisfy the constraint range of the corresponding segment, a penalty value is recorded.

[0048] In this embodiment, the input target variables include a combination of impedance, quality factor, coupling coefficient, insertion loss, self-resonant frequency, or one or more.

[0049] In one or more embodiments, when performing circuit-level simulation based on the scattering matrix, the method further includes: The differential signal is converted into a single-ended signal by a balun. One end is connected to a load with a set resistance value, and the real and imaginary parts of the input impedance at the other end are simulated.

[0050] In this step, the values ​​of the target variable and the corresponding total error will be recorded. Taking impedance as the input target variable as an example, the total error is defined as: (2); in, This refers to the target input impedance. These are simulation values, where Re represents the real part and Im represents the imaginary part.

[0051] If a set of design variables does not meet the constraints, the optimization module will receive a larger error value as a "penalty" than the normal result. If the user also includes the output impedance in the design objective, the total error will increase the simulated value. and The error is also calculated together. To facilitate user observation of specific error changes, this method uses the program to draw a line graph of "error - number of iterations" in real time on a single window, displaying the error for both the real and imaginary parts. Results that do not meet the constraints will not be displayed in the line graph.

[0052] In one or more embodiments, if the candidate or initial values ​​of the generated geometric parameters do not meet the corresponding constraint range, the set penalty value is fed back to the Bayesian optimization algorithm.

[0053] It is understandable that the penalty value is fed back to the Bayesian optimization algorithm so that it learns that this set of values ​​is invalid.

[0054] In this embodiment, the penalty value is set to 100.

[0055] It should be noted that if a set of design variables does not meet the constraints, a suitable penalty value needs to be fed back to the Bayesian optimization algorithm (the Bayesian optimization module described below). If the penalty value is less than the actual error value, the optimization model will consider the combination of variables that does not meet the constraints to be acceptable, and therefore continue to explore nearby areas; if the penalty value is too large, the penalty value will greatly increase the mean and standard deviation, causing all normal error values ​​to be "compressed" into almost zero, indistinguishable decimals after standardization.

[0056] It should be noted that all simulation records will be recorded, which can provide a large amount of reference data for the next optimization, allowing for quick location of the optimal range.

[0057] S700: When the termination condition is met, output the evaluation result; otherwise, based on historical data or the data corresponding to the evaluation, use the Bayesian optimization algorithm to generate a candidate value for the geometric parameters of each segment, and return to step S600. The error or penalty value is used as the output evaluation data.

[0058] In this step, the termination conditions include, but are not limited to, the number of evaluations or error convergence mentioned above.

[0059] In this embodiment, each effective performance evaluation requires a full-wave electromagnetic simulation that takes several minutes or even longer. Furthermore, there are as many as 23 design variables, and the relationship between the simulation results and parameters is complex, nonlinear, and non-convex, making it impossible to write analytical expressions or directly obtain gradients. In traditional optimization algorithms of electromagnetic industrial software, the hit rate of stochastic methods decreases exponentially with increasing dimensionality, resulting in extremely low efficiency and complete lack of engineering practicality. Gradient methods or quasi-Newton methods require finite difference estimation of gradients in high-dimensional space, which is computationally expensive and inevitably leads to getting trapped in one of countless local optima in high-dimensional space. This embodiment creatively introduces Bayesian optimization, designed to solve the problem of "expensive high-dimensional black boxes," offering unparalleled sample efficiency. Instead of blind searching, it uses Gaussian processes to build a cheap, global statistical surrogate model for the expensive simulation function. This model can not only predict the performance of any unevaluated point, but more importantly, it can also provide the uncertainty of the prediction. This function skillfully balances "exploitation" and "exploration" at each decision step, achieving the optimization effect that traditional algorithms require tens of thousands or even more evaluations with only hundreds of simulation evaluations, thus realizing an efficiency improvement of at least one order of magnitude.

[0060] In summary, the segmented structure proposed in this embodiment, with its extremely high degree of design freedom, can precisely control the local electromagnetic field, effectively compensate for high-frequency effects, and thus design a transformer with higher Q value, lower loss, and more accurate matching. Its overall performance exceeds that of any traditional uniform linewidth structure.

[0061] The Bayesian optimization algorithm used in this embodiment improves sample efficiency by at least an order of magnitude compared to traditional algorithms (such as gradient methods and stochastic methods). Combined with a fast constraint pre-screening mechanism, the design cycle, which originally required several weeks, is shortened to less than a day. Moreover, the fully automated process solidifies the optimization ideas of senior engineers into algorithms, reducing reliance on personal experience and ensuring the complete reproducibility of the design process.

[0062] Example 2: like Figure 2 As shown, this embodiment provides an automated design system for segmented matching circuits based on Bayesian optimization, which is applied to the automated design method for segmented matching circuits based on Bayesian optimization described in Embodiment 1. The system includes a central control module, a Bayesian optimization module, a component segmentation and variable constraint module, a three-dimensional electromagnetic simulator, and a circuit analysis module.

[0063] The central control module is used for process scheduling, data processing, intelligent decision-making, and communication with the simulator in the automated design method of segmented matching circuits based on Bayesian optimization.

[0064] Based on the above embodiments, users input their design requirements and optimization strategies into the central control module. For example, the optimal load impedance for the power amplifier designed by the user is 40+J. For 60 ohms, you can enter "TARGET_RE_ZIN=40" and "TARGET_IM_ZIN=60" in the central control program. If there are requirements for the design area, the user can adjust the range of line length values ​​in the optimization variables. The user can also set the total number of evaluations of the optimization algorithm, the penalty value for constraints or simulation failures, etc., according to actual needs.

[0065] The component segmentation and variable constraint module is used to divide the components in the matching circuit into several continuous segments according to their geometry, set corresponding geometric parameters for each segment, and transmit the segmentation results to the central control module; it is also used to configure the corresponding constraint range for each segment and accept the input target variable value for scheduling by the central control module.

[0066] The Bayesian optimization module, under the control of the central control module, determines whether historical data exists for the geometric parameters of each segment. If historical data exists, it loads the historical data and uses the Bayesian optimization algorithm to generate a candidate value for the geometric parameters of each segment. If no historical data exists, it generates n initial values ​​for the geometric parameters of each segment, evaluates each initial value corresponding to the component in turn, and uses the Bayesian optimization algorithm to generate candidate values ​​for the geometric parameters of each segment based on the evaluation data. It then checks whether each candidate value meets the constraint range of the corresponding segment. When at least one candidate value does not meet the constraint, a penalty value is recorded, and the evaluation result is obtained.

[0067] The 3D electromagnetic simulator constructs a 3D model of the component based on all candidate values ​​when each candidate value is satisfied, and then performs full-wave electromagnetic simulation based on the 3D model to generate a scattering matrix.

[0068] In this specific implementation, ANSYS's HFSS software is selected for electromagnetic simulation, and the simulation is started through the central control module.

[0069] The circuit analysis module is used to perform circuit-level simulation based on the scattering matrix, record the error between the simulation value and the input target variable value, and obtain the evaluation result.

[0070] In this specific implementation, ANSYS's Circuit software is selected for circuit analysis, and the central control module imports the scattering matrix into the Circuit software.

[0071] It should be noted that the specific method and process involved in this embodiment are the same as those in Embodiment 1, and for details, please refer to Embodiment 1.

[0072] It should be noted that the above modules may include: a configuration module, used to configure and edit the protocol file, wherein the protocol file is binary format data; specifically, the steps for configuring and editing the protocol file are as described in Example 1. A parsing module, used to parse the protocol file into JSON format data; specifically, the steps for parsing the protocol file into JSON format data are as described in Example 1.

[0073] Those skilled in the art will understand that all or part of the features / steps of the above-described method embodiments can be implemented by methods, data processing systems, or computer programs. These features can be implemented entirely in software or by a combination of hardware and software. The aforementioned computer program can be stored in one or more computer-readable storage media. When the computer program is executed (e.g., by a processor), it performs the steps of the above-described Bayesian optimization-based segmented matching circuit automated design method.

[0074] The aforementioned storage media capable of storing program code include: static disks, solid-state drives, random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), optical storage devices, magnetic storage devices, flash memory, magnetic disks or optical disks, and / or combinations of the above devices, that is, they can be implemented by any type of volatile or non-volatile storage devices or combinations thereof.

[0075] As shown in Figure 3, Figure 3(a) illustrates how the absolute errors of the real and imaginary parts of the input impedance decrease with the number of iterations. Subsequently, other parameter combinations were explored, ultimately finding a combination of variables that closely approximates the target. Figure 3(b) shows the automatically generated two-dimensional error distribution plot after optimization, visually illustrating the Pareto front of the optimization results under two mutually influencing optimization objectives. Figure 3(c) shows the detailed results of the final output of the central control program.

[0076] In summary, the system in this embodiment displays the optimization process from Embodiment 1 to the user in real time, giving the user a clear understanding of the optimization progress. After optimization is complete, a two-dimensional error distribution plot is generated, showing the Pareto front of the optimization results under two mutually influencing optimization objectives.

[0077] Example 3: like Figure 4 As shown, the present invention also provides a segmented variable geometric parameter coil, designed according to an automated design system for segmented matching circuits based on Bayesian optimization as described in Embodiment 2. It includes a tap 2 and a coil body 1. The coil body 1 includes a first metal wire 10, a second metal wire 11, a third metal wire 12, a fourth metal wire 13, a fifth metal wire 14, a sixth metal wire 15, a seventh metal wire 16, an eighth metal wire 17, a ninth metal wire 18, a tenth metal wire 19, and a final metal wire 20 connected sequentially. The second metal line 11, the third metal line 12, the fourth metal line 13, the fifth metal line 14, the sixth metal line 15, the seventh metal line 16, the eighth metal line 17, the ninth metal line 18, and the tenth metal line 19 form an octagonal structure. The first metal line 10 and the last metal line 20 form an opening 3. The first metal line 10 and the last metal line 20 are differential signal input and output lines. Tap 2 is led out from the sixth metal line 15.

[0078] Furthermore, the spacing between the first and last metal wires is adjustable, that is, the opening 3 is adjustable, with a default fixed value of 20um.

[0079] In this embodiment, according to the Bayesian optimization-based segmented matching circuit automated design system in Example 2, the Bayesian optimization-based segmented matching circuit automated design method in Example 1 is executed, and the optimized geometric parameter values ​​of the main coil and secondary coil in this embodiment are shown in Table 1 below. The distance between the main coil and the secondary coil is 15µm.

[0080] Table 1. Optimized values ​​of geometric parameters for the primary and secondary coils. Example 4: like Figure 5 As shown, the present invention also provides a segmented variable geometry radio frequency transformer, including a main coil 4 and a secondary coil 5. Both the main coil 4 and the secondary coil 5 are the segmented variable geometry coils described in claim 9. The opening 3 of the main coil 4 is 180 degrees away from the opening 3 of the secondary coil 5, and their central axes coincide.

[0081] Furthermore, to reduce ohmic losses, the main coil 4 is located on the top metal layer of the RF transformer, and the secondary coil 5 is located on the layer below the top metal layer. For the 65nm process, the main coil 4 is located on the ninth layer and the secondary coil 5 is located on the eighth layer.

[0082] This embodiment of a segmented variable geometry radio frequency transformer also includes a ground plane 6 placed around the transformer matching circuit, the distance between the ground plane 6 and the coil being adjustable.

[0083] Furthermore, when the main coil 4 is located on the ninth layer and the secondary coil 5 is located on the eighth layer, the ground plane 6 is placed on the fourth layer.

[0084] Understandably, the distance between ground plane 6 and the coil is adjustable. Its purpose is twofold: a smaller distance results in a smaller footprint, while a larger distance improves transformer coupling. Users can choose the appropriate setting based on their needs. This variable is not involved in optimization; it is manually adjusted by the user in the central control module before optimization.

[0085] It should be noted that, for the purpose of better illustrating the structure of this embodiment, Figure 5 The structure conceals other layers besides the main coil, secondary coil, and ground plane, and these other layers can be consistent with existing technologies.

[0086] It should be understood that the above embodiments are only special cases and do not indicate that the present invention is implemented in such a way.

[0087] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0088] The above description is merely a preferred embodiment of the present invention. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An automated design method for piecewise matching circuits based on Bayesian optimization, characterized in that, Includes the following steps: S100. Divide the components in the matching circuit into several consecutive segments according to their geometry, and set corresponding geometric parameters for each segment. S200, Do the geometric parameters of each segment have historical data? S300. If so, load historical data, generate a candidate value for the geometric parameters of each segment based on the historical data, and execute step S600. S400 If not, generate n initial values ​​for the geometric parameters of each segment, and evaluate each initial value corresponding to the component in turn; S500: Generate a candidate value for the geometric parameters of each segment based on the evaluation data. S600. When each candidate value satisfies the constraint range of the corresponding segment, a three-dimensional model of the component is constructed based on all candidate values. A full-wave electromagnetic simulation is performed based on the three-dimensional model to generate a scattering matrix. Circuit-level simulation is performed based on the scattering matrix, and the error between the simulation value and the input target variable value is recorded. When at least one candidate value does not satisfy the constraint range of the corresponding segment, a penalty value is recorded. S700. When the termination condition is met, output the evaluation result; otherwise, generate a candidate value for the geometric parameters of each segment based on historical data or the data corresponding to the evaluation and return to step S600. In this process, a candidate value is generated for the geometric parameters of each segment using a Bayesian optimization algorithm.

2. The automated design method for piecewise matching circuits based on Bayesian optimization according to claim 1, characterized in that, When performing circuit-level simulation based on the scattering matrix, it also includes: The differential signal is converted into a single-ended signal by a balun. One end is connected to a load with a set resistance value, and the real and imaginary parts of the input impedance at the other end are simulated.

3. The automated design method for piecewise matching circuits based on Bayesian optimization according to claim 1, characterized in that, n is twice the total number of geometric parameters corresponding to all segments of the component.

4. The automated design method for piecewise matching circuits based on Bayesian optimization according to claim 1, characterized in that, The initial values ​​for each component are evaluated sequentially, including the following steps: Each segment of the component is assigned an initial value corresponding to a geometric parameter, and a total of n initial values ​​are assigned. Does each initial value satisfy the constraint range of the corresponding segment? Thus, a three-dimensional model of the component is constructed based on each set of initial values. A full-wave electromagnetic simulation is performed based on the three-dimensional model to generate a scattering matrix. Circuit-level simulation is performed based on the scattering matrix. The error between the simulation value and the input target variable value is recorded to obtain the evaluation data for each set of initial values. Otherwise, record the penalty value to obtain the evaluation data for each set of initial values.

5. The automated design method for piecewise matching circuits based on Bayesian optimization according to claim 4, characterized in that, If the generated candidate or initial values ​​of geometric parameters do not meet the corresponding constraint range, the set penalty value will be fed back to the Bayesian optimization algorithm.

6. The automated design method for piecewise matching circuits based on Bayesian optimization according to claim 1, characterized in that, The components include a transformer, whose main coil and secondary coil are octagonal in shape, and each coil is symmetrical about the central axis. Each coil is divided into 11 consecutive segments and numbered sequentially. The width and length of the first to fifth segments corresponding to each number, the width of the sixth segment, and the center-to-center distance between the main coil and the secondary coil are used as geometric parameters to form 23 geometric parameters of the transformer.

7. An automated design system for piecewise matching circuits based on Bayesian optimization, characterized in that, The method for automated design of segmented matching circuits based on Bayesian optimization as described in any one of claims 1-6 includes a central control module, a component segmentation and variable constraint module, a Bayesian optimization module, a three-dimensional electromagnetic simulator, and a circuit analysis module. The central control module is used for process scheduling, data processing, intelligent decision-making, and communication with the simulator in the automated design method of segmented matching circuits based on Bayesian optimization. The component segmentation and variable constraint module is used to divide the components in the matching circuit into several continuous segments according to their geometry, set corresponding geometric parameters for each segment, and transmit the segmentation results to the central control module; it is also used to configure the corresponding constraint range for each segment and accept the input target variable value for scheduling by the central control module. The Bayesian optimization module, under the control of the central control module, determines whether historical data exists for the geometric parameters of each segment. If historical data exists, it loads the historical data and uses the Bayesian optimization algorithm to generate a candidate value for each segment's geometric parameters. If no historical data exists, it generates n initial values ​​for each segment's geometric parameters, evaluates each initial value corresponding to the component in turn, and uses the Bayesian optimization algorithm to generate candidate values ​​for each segment's geometric parameters based on the evaluation data. It then checks whether each candidate value satisfies the constraint range of the corresponding segment. When at least one candidate value does not satisfy the constraint, a penalty value is recorded, and the evaluation result is obtained. The 3D electromagnetic simulator constructs a 3D model of the component based on all candidate values ​​when each candidate value is satisfied, and then performs full-wave electromagnetic simulation based on the 3D model to generate a scattering matrix. The circuit analysis module is used to perform circuit-level simulation based on the scattering matrix, record the error between the simulation value and the input target variable value, and obtain the evaluation result.

8. A segmented coil with variable geometric parameters, characterized in that, According to claim 7, an automated design system for segmented matching circuits based on Bayesian optimization is designed, comprising taps and a coil body. The coil body comprises a first metal wire, a second metal wire, a third metal wire, a fourth metal wire, a fifth metal wire, a sixth metal wire, a seventh metal wire, an eighth metal wire, a ninth metal wire, a tenth metal wire, and a last metal wire connected in sequence. The second, third, fourth, fifth, sixth, seventh, eighth, ninth, and tenth metal wires form an octagonal structure, with the first and last metal wires forming an opening. The first and last metal wires are differential signal input / output lines, with the taps leading out from the sixth metal wire.

9. A segmented radio frequency transformer with variable geometric parameters, characterized in that, It includes a main coil and a secondary coil, both of which are segmented variable geometric parameter coils as described in claim 8. The opening direction of the main coil is 180 degrees different from the opening direction of the secondary coil, and their central axes coincide. The main coil is located on the top metal layer of the RF transformer, and the secondary coil is located on the layer below the top metal layer; or the main coil is located on the ninth layer, and the secondary coil is located on the eighth layer.

10. A segmented variable geometry radio frequency transformer according to claim 9, characterized in that, It also includes a ground plane placed around the transformer matching circuit, and the distance between the ground plane and the coil is adjustable; When the main coil is located on the ninth layer and the secondary coil is located on the eighth layer, the ground plane is placed on the fourth layer.

Images