A passive chip layout optimization method and device

Abstract

The application provides a passive chip layout optimization method and device. The passive chip layout optimization method provided by the application comprises the following steps: determining a coarse center frequency, a coarse relative bandwidth, a self-coupling coefficient and a coupling coefficient between adjacent components according to S-parameter models and topologies of the components in a current layout; predicting the coupling coefficient between non-adjacent components by using a prediction model according to the sizes and relative positions of any pair of component groups in the current layout; obtaining a coarse coupling matrix of the current layout according to the three types of coefficients of the current layout; correcting the coarse coupling matrix, the coarse center frequency and the coarse relative bandwidth by using a current correction model to obtain an accurate coupling matrix; taking the accurate coupling matrix as a prediction value of a surrogate model, selecting a next evaluation point based on the prediction value and its uncertainty, and generating a next generation layout; and taking the coarse coupling matrix, the coarse center frequency and the coarse relative bandwidth and a simulated electromagnetic response as samples to update the current correction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radio frequency chip optimization design, and in particular to a passive chip layout optimization method and device. BACKGROUND

[0002] With the advancement of miniaturization trends, passive chips play an increasingly important role in radio frequency and microwave systems, and their optimization design has become a key issue in the current industry. Especially in the application scenario of multi-element device integration, miniaturized passive chips have put forward more stringent requirements for design accuracy and efficiency. However, existing passive chip layout optimization techniques still face many challenges, especially in the processing and optimization of coupling matrices, and have not yet achieved efficient and accurate design.

[0003] Traditional layout optimization design usually adopts a circuit simulation method based on a schematic diagram for design, but these methods often cannot handle complex electromagnetic coupling effects, especially in microwave circuits, where the coupling effect between elements often has an important influence on the overall performance. Although the scheme based on field-circuit joint simulation involves the coupling between elements, it needs to repeatedly call complex and time-consuming electromagnetic full-wave simulation, which is too long in optimization time. Although some existing neural network schemes attempt to optimize layout design by constructing a mapping relationship to reduce the overall simulation times of the layout, these methods usually ignore the introduction of electromagnetic prior knowledge, resulting in insufficient accuracy and efficiency in handling complex electromagnetic coupling relationships. SUMMARY

[0004] Therefore, the present application provides a passive chip layout optimization method and device to accurately and efficiently optimize the layout of a passive chip.

[0005] Specifically, the present application is implemented by the following technical solutions:

[0006] The first aspect of the present application provides a passive chip layout optimization method, which comprises:

[0007] When iteratively optimizing the layout design parameters based on a Bayesian optimization algorithm, in the process of each iteration, when it is judged that the iteration has not terminated based on the simulation electromagnetic response corresponding to the current layout and the preset expected electromagnetic response, the rough center frequency, the rough relative bandwidth, the self-coupling coefficient and the coupling coefficient between adjacent elements of the current layout are determined according to the S-parameter model of each element in the current layout and the current layout topology.

[0008] For any one group of non-adjacent element pairs in the current layout, the coupling coefficient of the group of element pairs is predicted using the prediction model corresponding to the group of non-adjacent element pairs, according to the size of each element in the group of element pairs and the relative position of the group of element pairs.

[0009] According to the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout and the coupling coefficients of all non-adjacent component pairs in the current layout, a rough coupling matrix corresponding to the current layout is obtained;

[0010] The rough coupling matrix, the rough center frequency and the rough relative bandwidth are corrected by using a current correction model to obtain a precise coupling matrix, a precise center frequency and a precise relative bandwidth;

[0011] The precise coupling matrix, the precise center frequency and the precise relative bandwidth are taken as prediction values of the proxy model, and the next evaluation point is selected based on the prediction values and their uncertainties to generate the next generation layout;

[0012] The current correction model is updated by using the rough coupling matrix, the rough center frequency, the rough relative bandwidth and the simulation result.

[0013] The second aspect of the application provides a passive chip layout optimization device, the device comprises a processing module, a correction module and an updating module;

[0014] The processing module is configured to, when iteratively optimizing the layout design parameters based on a Bayesian optimization algorithm, in the process of each iteration, when it is judged that the iteration has not terminated based on the simulation electromagnetic response corresponding to the current layout and the preset expected electromagnetic response, determine the rough center frequency, the rough relative bandwidth, the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout according to the S-parameter model of each component in the current layout and the topology of the current layout.

[0015] The processing module is configured to, for any one group of non-adjacent component pairs in the current layout, predict the coupling coefficient of the group of non-adjacent component pairs by using the prediction model corresponding to the group of non-adjacent component pairs according to the size of each component in the group of non-adjacent component pairs and the relative position of the group of non-adjacent component pairs.

[0016] The processing module is configured to obtain a rough coupling matrix corresponding to the current layout according to the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout and the coupling coefficients of all non-adjacent component pairs in the current layout.

[0017] The correction module is configured to correct the rough coupling matrix, the rough center frequency and the rough relative bandwidth by using a current correction model to obtain a precise coupling matrix, a precise center frequency and a precise relative bandwidth.

[0018] The processing module is configured to take the precise coupling matrix, the precise center frequency and the precise relative bandwidth as prediction values of the proxy model, and select the next evaluation point based on the prediction values and their uncertainties to generate the next generation layout.

[0019] The updating module is configured to update the current correction model by using the coarse coupling matrix, the coarse center frequency, the coarse relative bandwidth, and the simulated electromagnetic response.

[0020] The passive chip layout optimization method and device provided in the application first uses different methods to respectively solve self-coupling coefficients, coupling coefficients between adjacent component devices, and coupling coefficients of pairs of non-adjacent component devices, so as to effectively decouple different types of coupling relationships, make the layout model structure clear, and make the modeling process efficient. Further, the correction model is used to correct the coarse coupling matrix, the coarse center frequency, the coarse relative bandwidth, and other coarse electromagnetic responses, so as to ensure the prediction accuracy of the proxy model for the coupling relationship and other electromagnetic responses and improve the reliability of the layout design method. Further, in each iteration, the correction model updates the coarse electromagnetic response and the simulated electromagnetic response of the current layout as samples, so that the accuracy of the correction model is continuously improved as the iteration proceeds, avoiding the limitation that the accuracy of the result obtained by the layout optimization design based on the fixed model depends on the model accuracy, and the next generation of layout selected by the optimization algorithm is more likely to meet the index requirements, and the optimization process is more efficient and accurate. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flowchart of the passive chip layout optimization method provided in the application embodiment one;

[0022] Figure 2 A flowchart of the passive chip layout optimization method provided in the application embodiment two;

[0023] Figure 3 A flowchart of the passive chip layout optimization method provided in the application embodiment three;

[0024] Figure 4 A structural schematic diagram of the passive chip layout optimization device provided in the application embodiment one. DETAILED DESCRIPTION

[0025] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, the same numbers refer to the same or similar elements unless otherwise represented. The embodiments described in the following exemplary embodiments do not represent all the embodiments consistent with the application.

[0026] The terms used in the application are only for the purpose of describing specific embodiments, and are not intended to limit the application. The singular forms "a", "an", and "the" used in the application are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein means and includes any or all possible combinations of one or more associated listed items.

[0027] It should be understood that, although the terms first, second, third, etc. can be used in this application to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information without departing from the scope of the application. Specific embodiments are given below to detail the technical solutions of the application.

[0028] Figure 1 The flowchart of the passive chip layout optimization method embodiment provided in the present application is shown in Figure 1. Figure 1 The passive chip layout optimization method provided in the present embodiment comprises:

[0029] S101, in the process of iterative optimization of the layout design parameters based on the Bayesian optimization algorithm, in each iteration, when it is judged that the iteration has not terminated based on the simulation electromagnetic response corresponding to the current layout and the preset expected electromagnetic response, the rough center frequency, the rough relative bandwidth, the self-coupling coefficient and the coupling coefficient between adjacent components in the current layout are determined according to the S-parameter model of each component in the current layout and the current layout topology.

[0030] The method provided in the present embodiment can optimize the geometric size of the layout components, and does not involve the optimization design of the topology structure. In addition, the simulation electromagnetic response includes the simulation coupling matrix, the simulation center frequency and the simulation relative bandwidth; the expected electromagnetic response includes the expected coupling matrix, the expected center frequency and the expected relative bandwidth.

[0031] It should be noted that before introducing the passive chip layout optimization method provided in the present application, a brief introduction of the Bayesian optimization algorithm is given.

[0032] Specifically, the Bayesian optimization algorithm is a global optimization method based on a proxy model, which is usually used for expensive black box optimization problems. The core idea is to build a proxy model, and in each iteration, the design parameter value that is most likely to improve the objective function value is selected according to the current evaluation result and the predicted uncertainty. The Bayesian optimization algorithm is widely used in many fields, especially in engineering optimization problems with large design parameter space and high computational complexity.

[0033] The main process of the Bayesian optimization algorithm in the current layout design method includes:

[0034] (1) Obtain the simulation electromagnetic response (the simulation electromagnetic response can include the coupling matrix, the simulation center frequency, the simulation relative bandwidth, etc.) of the current layout based on the electromagnetic full-wave simulation, and judge whether the iteration is terminated based on the simulation electromagnetic response and the preset expected electromagnetic response (the expected electromagnetic response includes the expected coupling matrix, the expected center frequency and the expected relative bandwidth, etc.); if yes, output the current layout, if no, execute the following process.

[0035] (2) Based on the proxy model, the predicted value and its uncertainty are obtained, and the next generation of layout is obtained based on the predicted value and its uncertainty by using the acquisition function, and the next round of iteration is entered.

[0036] The present application is based on the rough electromagnetic response (rough coupling matrix, rough center frequency, rough relative bandwidth, etc.) and the correction model to construct the proxy model, to obtain the predicted accurate coupling matrix, accurate center frequency, and accurate relative bandwidth as the predicted value, and then to obtain the next generation of layout according to the predicted value and its uncertainty, and enter the next round of iteration. The passive chip layout optimization method provided in the embodiment will be introduced below.

[0037] Specifically, in the process of each round of iteration, when it is judged that the iteration has not terminated based on the simulation electromagnetic response corresponding to the current layout and the preset expected electromagnetic response, in this step, the self-coupling coefficient and the coupling coefficient between adjacent components in the current layout are determined according to the S-parameter model of each component in the current layout and the current layout topology.

[0038] It should be noted that in specific implementation, the expected S parameter can be constructed according to the index, and then the expected coupling matrix, the expected center frequency and the expected relative bandwidth can be obtained based on the expected S parameter by using the vector fitting method.

[0039] In specific implementation, the overall S parameter model can be obtained according to the layout topology and the S parameter model of each component in the PDK. From the overall S parameter model, the rough center frequency and the rough relative bandwidth can be obtained. Further, the self-coupling coefficient and the coupling coefficient between adjacent components can be obtained by using the vector fitting method in combination with the rough center frequency, the rough relative bandwidth and the current topology of the layout.

[0040] In specific implementation, the process of determining the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout according to the S parameter model of each component in the current layout and the current layout topology can include:

[0041] (1) According to the S parameter model of each component in the current layout and the current layout topology, the overall S parameter model of the current layout is constructed.

[0042] Specifically, the S parameter model is a tool for describing the signal reflection and transmission characteristics in radio frequency and microwave circuits, and is usually used to describe the input and output characteristics of components (such as capacitors, inductors, resistors, filters, power dividers, etc.). In circuit design, S parameter model can help understand and analyze the transmission of signals between different ports. Further, by using the S parameter cascade and series-parallel connection rules, the S parameter models of multiple components are gradually combined according to the layout topology to form the overall S parameter model of the entire layout.

[0043] (2) determining the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout according to the overall S parameter model.

[0044] Firstly, the rough center frequency and the rough relative bandwidth can be read from the overall S parameter model.

[0045] Further, in a possible implementation, the self-coupling coefficient and the coupling coefficient between adjacent components can be determined according to the following vector fitting method:

[0046] Step one, determining the order of the S parameter according to the number of components in the overall S parameter model, and converting the original frequency to the angular frequency domain of the low-pass prototype.

[0047] If the layout contains N S components, the S parameter is N S order, and the S parameter can be expressed in the form of partial fraction as:

[0048] ;

[0049] Wherein, a i , c i and d are the poles, residues and constant terms of the S parameter to be solved; s = jΩ, Ω is the angular frequency converted to the low-pass prototype.

[0050] If the S parameter of the designed component is a band-pass characteristic, the corresponding low-pass prototype angular frequency domain is:

[0051] ;

[0052] Wherein, ω c is the original angular frequency; ω 0c is the rough center frequency; FBW c is the rough relative bandwidth.

[0053] Step two, constructing an auxiliary function and randomly generating its poles as the initial value of iteration.

[0054] The auxiliary function is also expressed in the form of partial fraction as:

[0055] ;

[0056] Wherein, is the residue of the auxiliary function to be solved; is the randomly generated complex pole.

[0057] Step three, solving the equation composed of the auxiliary function and the overall S parameter model by using the least square method to obtain the parameters c i , d and .

[0058] The equation consisting of the auxiliary function and the global S-parameter model is as follows:

[0059] ;

[0060] That is, if the poles of the auxiliary function are consistent with the global S-parameter model, then .

[0061] Frequency sampling is performed on the overall S-parameter model to transform the equations into a system of linear equations, and the least squares method is used to solve for c. i , d and .

[0062] Step four: Solve for the zeros of the auxiliary function, and use them as the poles of the auxiliary function in the next round of iteration, until the poles of the two rounds of iteration do not change significantly.

[0063] Will and The partial fraction expansion in the form of zeros and poles can be expressed as:

[0064] ;

[0065] ;

[0066] Then we have:

[0067] ;

[0068] That is, the zeros of the auxiliary function are the poles of the global S-parameter model. To improve the accuracy of pole extraction, the zeros can be... Replace the initial pole of the auxiliary function, and repeat steps three and four until... No significant changes.

[0069] Step 5, As the poles of the S-parameters, the least squares method is used to solve the system of linear equations consisting of S-parameters in partial fractional form to obtain the residues of the S-parameters; the S-parameters in partial fractional form are pole residues.

[0070] The S-parameters are sampled at different frequencies to obtain the S-parameter residues and the linear equations formed by the S-parameters at different frequency points. The S-parameter residues are then obtained by solving the least squares method.

[0071] Step 6: Obtain the values ​​of each element of the coupling matrix from the relationship between the overall S-parameter model and the coupling matrix.

[0072] The relationship between the overall S-parameter model and the coupling matrix is ​​shown below:

[0073]

[0074] ;

[0075] ;

[0076]

[0077]

[0078] where j is the imaginary unit; M is the coupling matrix; ω c is the angular frequency; ω 0c is the center frequency; FBW c is the relative bandwidth; N S is the number of PDK components in the layout. The equation can be solved by means of eigenvalues and other information to obtain the values of each element of the coupling matrix. However, the obtained coupling structure is generally difficult to implement and does not match the topology of the layout.

[0079] Step seven, using a similarity transformation to eliminate the coupling matrix, so that the coupling matrix corresponds to the topology structure consistent with the layout topology structure, and then obtain the self-coupling coefficient and the coupling coefficient between adjacent components.

[0080] Because the overall S-parameter model obtained by cascading and series-parallel S-parameters only considers the self-coupling coefficient and the coupling between adjacent elements, a similarity transformation is needed to make the coupling coefficient between non-adjacent components zero.

[0081] S102, for any pair of non-adjacent components in the current layout, according to the size of each component in the pair of components and the relative position of the pair of components, the coupling coefficient of the pair of components is predicted by using the prediction model corresponding to the pair of non-adjacent components.

[0082] Specifically, the pair of components refers to a combination of two components in the layout, which has a certain energy coupling relationship, and the strength of the coupling relationship is related to the size and physical position of the components. By obtaining the size of each component in the pair of components and their relative position as input, the prediction model is used to predict the coupling coefficient of the pair of non-adjacent components. In a layout, there are usually multiple pairs of components, and the coupling coefficient of each pair of non-adjacent components is calculated. Through this process, the coupling coefficients of multiple pairs of non-adjacent components can be obtained, which reflect the electromagnetic coupling strength between different components. For convenience of explanation, the coupling coefficient of the pair of non-adjacent components is denoted as the coupling coefficient between non-adjacent components.

[0083] It should be noted that the input of the prediction model is the size and relative position of the component pair, and the output is the coupling coefficient between the non-adjacent components. Specifically, when predicting, the size and relative position of the component pair are taken as input. After receiving the above input, the prediction model outputs the corresponding non-normalized coupling coefficient between the non-adjacent components.

[0084] In the embodiment, different non-adjacent component pairs correspond to different prediction models, and a corresponding prediction model is trained for each type of component pair. In this way, by training a corresponding prediction model for each type of component pair, higher prediction accuracy and reliability can be achieved. Specifically, since different types of components differ in size, position and electromagnetic characteristics, using a general model may not be able to capture the specific effects of these differences on the coupling coefficient between non-adjacent components. By training a prediction model for the characteristics of each type of component pair, the interaction and coupling effect between components can be more accurately reflected, thereby providing more accurate model support for optimizing the layout design.

[0085] S103, obtaining a rough coupling matrix corresponding to the current layout according to the self-coupling coefficient, the coupling coefficient between adjacent components, and the coupling coefficient of all non-adjacent component pairs in the current layout.

[0086] Specifically, the rough coupling matrix is formed by integrating the self-coupling coefficient, the coupling coefficient between adjacent components, and the coupling coefficient between non-adjacent components, and is a key intermediate result in the process of optimizing the layout of a passive chip, used to quickly represent the coupling relationship between all components in the layout. The rough coupling matrix is a symmetric matrix of (N S +2)×(N S +2), N S is the number of components in the layout, the diagonal elements are the self-coupling coefficients, the i-th row and j-th column elements correspond to the i-th and j-th components, if the i-th and j-th components are adjacent, the element is the coupling coefficient between adjacent components, and if the i-th and j-th components are not adjacent, the element is the coupling coefficient between non-adjacent components.

[0087] It should be noted that the coupling coefficient output by the prediction model is the non-normalized coupling coefficient between non-adjacent components, so before this step, the coupling coefficient between non-adjacent components is normalized. Specifically, the normalization process can be represented as: m i,j = M i,j / FBW c , where M i,j is the non-normalized coupling coefficient between non-adjacent components, FBW c is the rough relative bandwidth.

[0088] The method provided by the embodiment decouples the determination process of the coupling matrix, determines the coupling coefficients between adjacent components and the self-coupling coefficients based on the component S-parameter model provided in the PDK, and determines the coupling coefficients between non-adjacent components based on a prediction model. In this way, the self-coupling coefficients, the coupling coefficients between adjacent components, and the coupling coefficients between non-adjacent components are independently calculated, the rough coupling matrix can be quickly predicted, and the layout optimization design is accelerated. First, the component S-parameter model provided in the PDK is used to determine the coupling coefficients between adjacent components and the self-coupling coefficients. Without significantly reducing the model accuracy, only the main influencing factors (electromagnetic coupling between adjacent components through transmission lines) of these parameters are considered to avoid using complex and time-consuming electromagnetic full-wave simulation, thereby saving modeling time. Second, the coupling coefficients between non-adjacent components are determined through a prediction model, so that the prediction model can quickly obtain relatively accurate coupling coefficients between non-adjacent components in different layouts and different component pairs.

[0089] In addition, by decoupling, only the prediction model needs to be trained, and the self-coupling coefficients and the coupling coefficients between adjacent components can be quickly obtained by combining the PDK, without training a global model based on the layout design parameters to output a complete coupling matrix. This can significantly simplify the calculation process and improve efficiency. This method enables the prediction model to be more flexible in adapting to different circuit configurations and component types. Only the coupling coefficients need to be predicted according to the size, position, and other characteristics of each pair of non-adjacent components, thereby avoiding redundant calculations of the global model. At the same time, the scalability and maintainability of the prediction model are enhanced. In the future, when facing new design requirements, only the prediction model needs to be fine-tuned, without retraining the global model, which greatly improves the flexibility and adaptability of the layout optimization design.

[0090] S104, correcting the rough coupling matrix, the rough center frequency, and the rough relative bandwidth by using the current correction model to obtain a precise coupling matrix, a precise center frequency, and a precise relative bandwidth.

[0091] It should be noted that, in order to further improve the prediction accuracy of the coupling matrix, in the embodiment, after the rough coupling matrix is obtained, the rough electromagnetic response (the rough electromagnetic response includes the rough coupling matrix, the rough center frequency, and the rough relative bandwidth) is further corrected by a correction model to obtain a precise electromagnetic response, and the precise electromagnetic response includes a precise coupling matrix, a precise center frequency, and a precise relative bandwidth. Specifically, the input of the correction model is the rough coupling matrix, the rough center frequency, and the rough relative bandwidth, and the output is the precise coupling matrix, the precise center frequency, and the precise relative bandwidth.

[0092] S105, taking the precise coupling matrix, the precise center frequency, and the precise relative bandwidth as the prediction value of the proxy model, selecting the next evaluation point based on the prediction value and its uncertainty, and generating the next generation of layout.

[0093] According to the description of the Bayesian optimization algorithm in S101, it can be understood that the Bayesian optimization algorithm generates the next generation layout based on the evaluation result of the current layout performance and the uncertainty of the evaluation result.

[0094] In this embodiment, the accurate electromagnetic response such as the accurate coupling matrix, the accurate center frequency, and the accurate relative bandwidth is taken as the prediction value of the proxy model, and the next evaluation point is selected based on the prediction value and its uncertainty to generate the next generation layout.

[0095] In a specific implementation, when the next evaluation point is selected based on the prediction value and its uncertainty, an expected improvement function (EI) can be used as a collection function to determine the design parameter (geometric size) x' corresponding to the next generation layout.

[0096] ;

[0097] wherein x is a vector composed of design parameters; y* is the objective function value of the layout that best meets the index in the current sample; min(y*-y,0) is the minimum value between y*-y and 0; is the posterior error of the Bayesian optimization objective function prediction value y corresponding to the vector x composed of design parameters, which can be obtained by the Bayesian principle; x' is the geometric parameter that minimizes the EI, which can be obtained by optimization.

[0098] Optionally, in a possible implementation, the objective function y of the Bayesian optimization algorithm is a function that comprehensively considers the gap between the accurate electromagnetic response (coupling matrix, center frequency, relative bandwidth, etc.) and the expected electromagnetic response.

[0099] For example, in an embodiment, the objective function y can be defined as:

[0100] ;

[0101] wherein is the norm; M(x), ω0(x), and FBW(x) are the accurate coupling matrix, the accurate center frequency, and the accurate relative bandwidth, respectively; M g , ω 0g , and FBW g are the expected coupling matrix, the expected center frequency, and the expected relative bandwidth, respectively; and α, β, and γ are weight coefficients for balancing the contributions of different factors.

[0102] S106, updating the current correction model using the rough coupling matrix, the rough center frequency, the rough relative bandwidth, and the simulated electromagnetic response.

[0103] Specifically, after the end of each round, the current correction model can be updated based on the rough electromagnetic response and the simulated electromagnetic response to further improve the accuracy of the correction model and reevaluate the uncertainty of the predicted value.

[0104] Specifically, the rough electromagnetic response and the simulated electromagnetic response are combined in the form of sample pairs to form a training sample, for example, a vectorized rough coupling matrix as an input feature and a vectorized simulated coupling matrix as a supervised label to form a training sample for training the correction model. The training sample will be added to the sample set. The correction model adopts a supervised learning mechanism to adjust its network weights and thresholds through error back propagation and an optimization algorithm, updates the uncertainty of the predicted value based on a Bayesian optimization algorithm, and realizes the learning of the correction model on the mapping relationship contained in the current layout structure.

[0105] In a specific implementation, in one embodiment, the coupling matrix, the center frequency, and the relative bandwidth can be combined into a vector to train a neural network; in another embodiment, the coupling matrix, the center frequency, and the relative bandwidth can be combined into a vector separately to train a neural network. In this embodiment, it is not limited.

[0106] It should be noted that through the update of each iteration, the correction model can gradually improve the prediction accuracy, and as the training deepens, it can accurately capture the small changes in the layout optimization design, thereby improving the overall prediction accuracy.

[0107] The passive chip layout optimization method and device provided by the application first use different methods to solve the self-coupling coefficient, the coupling coefficient between adjacent component devices, and the coupling coefficient of non-adjacent component pairs, to effectively decouple different types of coupling relationships, so that the layout model structure is clear and the modeling process is efficient. Further, the rough electromagnetic response such as the rough coupling matrix, the rough center frequency, and the rough relative bandwidth is corrected by the correction model to ensure the prediction accuracy of the coupling relationship and other electromagnetic responses of the proxy model and improve the reliability of the layout design method. Further, in each iteration, the correction model updates the rough electromagnetic response and the simulated electromagnetic response of the current layout as samples, so that the accuracy of the correction model is continuously improved as the iteration proceeds, avoiding the limitations of the optimization design based on the fixed model, such as the accuracy of the results depending on the model accuracy. The next generation layout selected by the optimization algorithm is more likely to meet the index requirements, and the optimization process is more efficient and accurate.

[0108] Figure 2 The flowchart of the second embodiment of the passive chip layout optimization method provided by the application is provided. Please refer to Figure 2 On the basis of the above embodiment, the training process of the prediction model corresponding to each group of non-adjacent component pairs can include:

[0109] S201, for any group of components, according to the orthogonal test design method, set the corresponding multiple design parameter combination schemes of the group of components.

[0110] Specifically, for each group of components, first set its adjustable design parameters, for example, in this example, to realize the optimization of the layout component geometric size, the adjustable design parameters include size and phase position. Further, through orthogonal test design, the different values of each design parameter are determined systematically, and combined into multiple design parameter combination schemes.

[0111] Specifically, suppose the currently selected group of components is composed of components and , whose types are , , the vectors composed of size parameters are , , and the relative positions between components are represented by the position vector , then a design parameter combination scheme can be represented as:

[0112] =[ , , ];

[0113] S202, for each design parameter combination scheme, obtain the coupling coefficient between non-adjacent components corresponding to the design parameter combination scheme through electromagnetic full-wave simulation.

[0114] Specifically, two standard components in the PDK are randomly selected (the two standard components can be of the same type or different types) as simulation modeling objects, and the pair of components is placed in the chip layout for arrangement. The arrangement mode can include same-layer arrangement or different-layer arrangement, and is not limited to a specific arrangement mode, so as to enhance the representativeness of the sample and the generalization ability of the model. In specific implementation, for a design parameter combination scheme, simulation modeling is performed based on the size and relative position of the components determined in step S201.

[0115] It is worth noting that, as described above, different types of component pairs correspond to different prediction models, for example, if there are l types of standard components defined in the PDK, theoretically, up to independent prediction models can be constructed, each corresponding to a different type of component pair (including same-type component pairs and different-type component pairs).

[0116] It should be noted that the prediction model is constructed in a good modularization and portability manner, chip manufacturers can pre-construct the prediction model, and further package the prediction model with the PDK to provide to the user, so as to support adaptive size optimization design under different layout topologies, significantly improve the design efficiency and automatic modeling capability, improve the layout optimization efficiency, and shorten the optimization time.

[0117] Further, in a possible implementation, for each group of design parameter combination schemes, after obtaining the coupling coefficient of the corresponding device pair by electromagnetic full-wave simulation (it can be understood that the coupling coefficient between the non-adjacent devices is an unnormalized coefficient), the coupling coefficient can be further standardized and cleaned, and the obviously abnormal data can be deleted.

[0118] In addition, in order to control the complexity of the prediction model and improve the training efficiency, the maximum distance between the non-adjacent device pairs is limited to twice the maximum working wavelength of the layout, and the coupling effect beyond the distance is generally significantly weakened. The precise modeling has limited effect on the accuracy of the overall rough coupling matrix, but it will significantly increase the training time and storage overhead of the prediction model. The selection of the maximum distance reflects the balance strategy between modeling accuracy and resource consumption.

[0119] It should be noted that when the maximum distance of the device pair is set to zero, the rough coupling matrix degenerates into a simplified model that does not include non-adjacent device coupling, and only the self-coupling coefficient and the coupling coefficient between adjacent devices are retained, which is suitable for layout optimization design with weak coupling interference.

[0120] S203, each group of design parameter combination schemes is taken as an input feature of a sample for training a prediction model, and the non-adjacent device coupling coefficient corresponding to the design parameter combination scheme is taken as a label to construct a sample.

[0121] It can be understood that first, the input features of each group of non-adjacent device pairs in the sample set are taken as the input of the neural network, and the input features include the respective size parameters (such as length, width, etc.) of the two devices and the relative position (such as the center distance, the relative rotation angle, etc.) between them; further, the unnormalized non-adjacent device coupling coefficient obtained by electromagnetic full-wave simulation is taken as the output label to construct a group of samples. The sample set is divided into a training set, a validation set and a test set. The training set is used for parameter fitting, the validation set is used to adjust the network structure, the learning rate, the regularization coefficient and other hyperparameters, to avoid overfitting of the prediction model. The generalization ability of the trained model is tested by the test set data, that is, the prediction accuracy and stability of the model on unseen samples are evaluated, to verify whether the prediction model has the ability to quickly predict any new device pair in actual application.

[0122] S204, training the initial neural network pre-constructed by using all samples to obtain a corresponding prediction model of the component pair.

[0123] Specifically, in this step, the prediction model between non-adjacent components can be trained based on the obtained data set to establish the mapping relationship between the physical properties (geometric size and relative position, etc.) of the components and the coupling coefficients between the non-adjacent components.

[0124] Optionally, in a possible implementation, the prediction model adopts an initial neural network structure, which can realize nonlinear function approximation between high-dimensional input parameters and coupling coefficients between non-adjacent components.

[0125] Specifically, an initial neural network model for different component combinations is constructed. For example, in an embodiment, the network structure can include three hidden layers, use a ReLU activation function, and output the predicted coupling coefficients between non-adjacent components.

[0126] Optionally, in a possible implementation, the structures of the initial neural networks corresponding to different adjacent component pairs can be the same or different.

[0127] Further, the initial neural network is trained by minimizing the mean square error (MSE) loss of the predicted value and the label. It can be understood that, in the training process, the back propagation algorithm and the gradient descent method can be used for network training, and finally a prediction model with strong generalization ability is obtained. The prediction model can be used to predict the non-normalized coupling coefficients between the two non-adjacent components under any given size and relative position.

[0128] Optionally, in a possible implementation, the initial neural network can adopt a double-branch network. One branch learns the influence of the relative position between components on the coupling coefficients between non-adjacent components, and the other branch learns the influence of the geometric size of the components on the coupling coefficients between non-adjacent components. The features formed by the two branches are fused and output through a fully connected layer to obtain the coupling coefficients. Further, when training different groups of non-adjacent component pairs, only the position branch needs to be fine-tuned, and the weights and thresholds in the remaining network structure need to be trained, so that the information related to the relative position can be quickly mastered, and the training cost of the coupling coefficients between non-adjacent components is reduced.

[0129] The method provided by the embodiment combines orthogonal test design and electromagnetic full-wave simulation, generates a large number of samples by using a plurality of component parameter combination schemes, and trains a non-adjacent component inter-coupling coefficient prediction model by using an initial neural network. In this way, first, the orthogonal test design can obtain the mapping relationship information between the parameters of the most component pairs and the non-adjacent component inter-coupling coefficients by using as few samples as possible. Second, the electromagnetic full-wave simulation is used to obtain accurate non-adjacent component inter-coupling coefficient data, so as to ensure the authenticity and reliability of the samples. Through machine learning, the model can automatically learn the complex relationship between the parameter combination and the non-adjacent component inter-coupling coefficient, and efficient and accurate prediction is realized. In this way, when facing a new design scheme, the non-adjacent component inter-coupling coefficient can be quickly and accurately predicted, and the time-consuming electromagnetic full-wave simulation is not needed, so that the efficiency of the layout optimization is improved and the time of the layout optimization is shortened.

[0130] Figure 3 A flowchart of the passive chip layout optimization method embodiment four provided by the present application is shown in FIG. 4. Figure 3 On the basis of the above embodiment, the next evaluation can be selected based on the predicted value and the uncertainty thereof, which can include:

[0131] S301, determining a target iteration stage corresponding to a current iteration number according to the current iteration number and a preset correspondence between iteration numbers and iteration stages.

[0132] According to the preset correspondence between iteration numbers and iteration stages, the iteration process can be divided into a plurality of stages. For example, an exploration stage, a balance stage, and a convergence stage. Each stage can represent different focuses, such as exploring the surrounding area of the historical sample and developing the uncertainty area.

[0133] According to the current iteration number, the target iteration stage corresponding to the current iteration can be calculated. For example, if the total iteration number is 100 times, and the system divides it into 5 stages, each stage contains 20 iterations. According to the current iteration number, it is determined which stage it belongs to, and the target iteration stage is determined.

[0134] S302, determining a target acquisition function corresponding to the target iteration stage according to the target iteration stage and a preset correspondence between iteration stages and acquisition functions.

[0135] Specifically, according to the mapping relationship between the target iteration stage and the acquisition function, the target acquisition function used in this round of iteration is selected. For example, the upper confidence bound function UCB is used in the initial stage to enhance the coverage of the unexplored area. The expected improvement function EI is used in the middle stage to balance the performance utilization and space exploration. The probability improvement function PI is used in the later stage to accelerate convergence.

[0136] Specifically, in the Bayesian optimization algorithm, different iteration stages use different acquisition functions, which can significantly improve the efficiency and accuracy of the optimization process. In the early stage, extensive exploration sampling can help obtain a comprehensive understanding of the entire search space, avoiding premature convergence to a local optimal solution. As the Bayesian optimization algorithm progresses, the acquisition function that balances exploration and exploitation can perform more concentrated optimization in known good areas while maintaining moderate exploration to ensure that potential favorable areas are not overlooked. In the later stage, the acquisition function that focuses on local exploitation can target the optimization of areas where the objective function performs well, thereby quickly approaching the global optimal solution. By dynamically adjusting the acquisition function, the optimal balance between exploration and exploitation can be ensured, improving optimization efficiency and accelerating convergence.

[0137] S303, selecting a next evaluation point based on the accurate coupling matrix and the target acquisition function.

[0138] Given the accurate electromagnetic response of the current layout, such as the accurate coupling matrix, the accurate center frequency, the accurate relative bandwidth, and the target acquisition function, the target acquisition function is optimized to obtain the evaluation point that is most likely to meet the target under the current proxy model, and the next round of evaluation is performed.

[0139] The method provided by the embodiment effectively improves the flexibility and efficiency of the Bayesian optimization algorithm by dynamically adjusting the sampling strategy according to the current iteration progress and stage target. First, the target iteration stage is determined according to the current iteration number, which can ensure that the most suitable acquisition function is used in different stages, so that the search space can be widely explored in the early stage, the exploration and exploitation can be balanced in the middle stage, and the fine optimization can be focused on in the later stage. Different acquisition functions are used in different stages, which helps to maximize optimization efficiency while avoiding premature convergence to a local optimal solution or ineffective search area. In addition, the proxy model prediction based on the accurate coupling matrix can efficiently guide the sampling process, reduce unnecessary computational load, and gradually improve the prediction accuracy of the model with each evaluation update. This method of adjusting the sampling strategy in stages not only speeds up the optimization process, but also improves the quality and accuracy of the final solution.

[0140] Optionally, in a possible implementation, the method further includes: calculating the global sensitivity of the coupling matrix based on the accurate electromagnetic response corresponding to the current layout; and dynamically adjusting the sampling strategy in the Bayesian optimization algorithm according to the global sensitivity of the coupling matrix when selecting the next evaluation point based on the predicted value and the uncertainty thereof.

[0141] Specifically, the global sensitivity of the coupling matrix refers to the global influence degree of each design parameter (such as the size, position, etc. of the component) on each matrix element (i.e. coupling coefficient) of the coupling matrix. The sensitivity analysis can be achieved by calculating the influence of the slight change of the design parameter in the design parameter space on the coupling matrix element. For example, in an embodiment, these influences can be estimated based on random sampling using numerical differentiation. In specific implementation, for example, in each iteration, based on the morris sensitivity analysis method, the design parameters are sampled using Latin hypercube sampling, the accurate electromagnetic response of the current layout is calculated using the proxy model, the accurate coupling matrix of each sampling point is obtained, for each design parameter, the design parameter is modified by introducing a small perturbation, and the accurate coupling matrix is recalculated. Then, the difference of the coupling matrix before and after the perturbation of each design parameter of each sampling point is calculated, and the average of each sampling point is taken to obtain the sensitivity index of each design parameter. In addition, the accuracy of the sensitivity index improves with the iteration of the correction model.

[0142] The Bayesian optimization algorithm is a global optimization method based on the proxy model, and a collection function is usually used to select the next evaluation point. The collection function (such as expected improvement, probability improvement, etc.) usually selects the next evaluation point according to the predicted value and uncertainty of the proxy model. Through sensitivity analysis, the design parameters that have a greater influence on the coupling matrix are screened out, and these design parameters are used as optimization parameters to optimize the collection function, and the values of the remaining design parameters are the values of the corresponding design parameters of the current optimal layout, which assists in dynamically adjusting the sampling strategy. By screening the design parameters with high correlation (i.e. high sensitivity) with the coupling matrix, the design parameter space can be compressed, and the efficiency of the Bayesian optimization can be significantly improved.

[0143] Corresponding to the foregoing embodiment of the passive chip layout optimization method, the present application also provides an embodiment of a passive chip layout optimization device.

[0144] Figure 4 A structure diagram of an embodiment of a passive chip layout optimization device provided by the present application is shown in FIG. 4. As shown in FIG. 4, the device provided by the embodiment includes a processing module 410, a correction module 420 and an updating module 430. Figure 4 The device provided by the embodiment includes a processing module 410, a correction module 420 and an updating module 430.

[0145] The processing module 410 is configured to, when iteratively optimizing the layout design parameters based on the Bayesian optimization algorithm, in the process of each iteration, when it is judged that the iteration has not terminated based on the simulation electromagnetic response corresponding to the current layout and the preset expected electromagnetic response, determining the rough center frequency, the rough relative bandwidth, the self-coupling coefficient and the coupling coefficient between adjacent components of the current layout according to the S-parameter model of each component in the current layout and the topology of the current layout.

[0146] The processing module 410 is configured to, for any one group of non-adjacent component pairs in the current layout, predict a coupling coefficient of the group of non-adjacent component pairs according to sizes of each component in the group of component pairs and relative positions of the group of component pairs by using a prediction model corresponding to the group of non-adjacent component pairs.

[0147] The processing module 410 is configured to obtain a coarse coupling matrix corresponding to the current layout according to the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout and the coupling coefficients of all groups of non-adjacent component pairs in the current layout.

[0148] The correction module 420 is configured to correct the coarse coupling matrix, the coarse center frequency and the coarse relative bandwidth by using a current correction model to obtain an accurate coupling matrix, an accurate center frequency and an accurate relative bandwidth.

[0149] The processing module 410 is configured to take the accurate coupling matrix, the accurate center frequency and the accurate relative bandwidth as prediction values of a proxy model, select a next evaluation point based on the prediction values and their uncertainties, and generate a next generation layout.

[0150] The updating module 430 is configured to update the current correction model by using the coarse coupling matrix, the coarse center frequency, the coarse relative bandwidth and the simulated electromagnetic response.

[0151] The application further provides a passive chip layout optimization device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and characterized in that the processor implements the steps of any one of the methods provided in the first aspect of the application when executing the program.

[0152] The device of the embodiment can be used to execute the steps of the method embodiment, and the specific implementation principles and implementation processes are similar, which will not be described here again. Figure 1 The functions and roles of the units in the above device are implemented in detail in the implementation processes of the corresponding steps in the above method, which will not be described here again.

[0153] The functions and roles of the units in the above device are implemented in detail in the implementation processes of the corresponding steps in the above method, which will not be described here again.

[0154] For the device embodiment, since it basically corresponds to the method embodiment, the related parts can be referred to the part of the method embodiment. The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or distributed on multiple network units. According to actual needs, part or all of the modules can be selected to achieve the purpose of the application scheme. Those skilled in the art can understand and implement it without creative labor.

[0155] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of protection of the present application.

Claims

1. A method of passive chip layout optimization, the method comprising: The passive chip layout optimization method comprises: In the process of iteratively optimizing the layout design parameters based on the Bayesian optimization algorithm, when it is judged that the iteration has not terminated based on the simulation electromagnetic response corresponding to the current layout and the preset expected electromagnetic response, the rough center frequency, the rough relative bandwidth, the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout are determined according to the S-parameter model of each component in the current layout and the current layout topology; For any one group of non-adjacent component pairs in the current layout, the coupling coefficient of the group of component pairs is predicted by using the prediction model corresponding to the group of non-adjacent component pairs according to the size of each component in the group of component pairs and the relative position of the group of component pairs; The rough coupling matrix corresponding to the current layout is obtained according to the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout, and the coupling coefficient of all non-adjacent component pairs in the current layout; The precise coupling matrix, the precise center frequency and the precise relative bandwidth are obtained by correcting the rough coupling matrix, the rough center frequency and the rough relative bandwidth by using the current correction model; The precise coupling matrix, the precise center frequency and the precise relative bandwidth are used as the prediction value of the proxy model, and the next evaluation point is selected based on the prediction value and its uncertainty to generate the next generation layout; The current correction model is updated by using the rough coupling matrix, the rough center frequency, the rough relative bandwidth and the simulation electromagnetic response.

2. The method of claim 1, wherein, The training process of the prediction model corresponding to each group of non-adjacent component pairs comprises: For any one group of components, a plurality of design parameter combination schemes corresponding to the group of components are set according to an orthogonal experimental design method; wherein each design parameter combination scheme comprises the size and position of two components in the group of components; For each design parameter combination scheme, the non-adjacent component coupling coefficient corresponding to the design parameter combination scheme is obtained by electromagnetic full-wave simulation; Each group of design parameter combination schemes is used as the input feature of the sample for training the prediction model, and the non-adjacent component coupling coefficient corresponding to the design parameter combination scheme is used as the label to construct a sample; An initial neural network is trained by using all samples to obtain the prediction model corresponding to the group of components.

3. The method of claim 1, wherein, The rough center frequency, the rough relative bandwidth, the self-coupling coefficient and the coupling coefficient between adjacent components corresponding to the current layout are determined according to the S-parameter model of each component in the current layout and the current layout topology, comprising: The overall S-parameter model of the current layout is constructed according to the S-parameter model of each component in the current layout and the current layout topology; The rough center frequency and the rough relative bandwidth are read from the overall S-parameter model; The self-coupling coefficient and the coupling coefficient between adjacent components are determined by the overall S-parameter model based on a vector fitting method.

4. The method of claim 3, wherein, The self-coupling coefficient and the coupling coefficient between adjacent components are determined by the overall S-parameter model based on a vector fitting method, comprising: The order of the S-parameter is determined according to the number of components in the overall S-parameter model, and the original frequency is converted to the angular frequency domain of the low-pass prototype; Constructing an auxiliary function and randomly generating poles of the auxiliary function as initial values of iteration; Solving equations composed of the auxiliary function and the overall S-parameter model by using the least square method to obtain specified parameters; Solving zeros of the auxiliary function as poles of the auxiliary function in the next round of iteration until the poles in the previous and next rounds of iteration have no obvious change; Taking the zeros as poles of S parameters, solving a linear equation group composed of partial fraction form S parameters by using the least square method to obtain residues of the S parameters; the partial fraction form S parameters are pole residue form; Obtaining element values of the coupling matrix according to a relationship between the overall S-parameter model and the coupling matrix; Using a similar transformation to eliminate the coupling matrix to make a topology corresponding to the coupling matrix consistent with a current layout topology, and obtaining self-coupling coefficients and coupling coefficients between adjacent component pairs.

5. The method of claim 1, wherein, The method further comprises: Calculating a global sensitivity of the coupling matrix based on a precise electromagnetic response corresponding to the current layout; When selecting a next evaluation point based on the predicted value and uncertainty thereof, dynamically adjusting a sampling strategy in the Bayesian optimization algorithm according to the global sensitivity of the coupling matrix.

6. The method of claim 1, wherein, The acquisition function of the Bayesian optimization algorithm is a function that comprehensively considers the predicted value and uncertainty.

7. The method of claim 1, wherein, Selecting a next evaluation point based on the predicted value and uncertainty thereof comprises: Determining a target iteration stage corresponding to a current iteration number according to a correspondence between the current iteration number and preset iteration numbers and iteration stages; Determining a target acquisition function corresponding to the target iteration stage according to a correspondence between the target iteration stage and preset iteration stages and acquisition functions; different iteration stages correspond to different acquisition functions; Selecting a next evaluation point based on the precise coupling matrix and the target acquisition function.

8. The method of claim 2, wherein, Different adjacent component pairs correspond to different structures of initial neural networks.

9. A passive chip floorplan optimization apparatus, comprising: The device comprises a processing module, a correction module and an updating module; The processing module is configured to, when iteratively optimizing layout design parameters based on a Bayesian optimization algorithm, in the process of each iteration, when it is determined that iteration has not terminated based on a simulation electromagnetic response corresponding to a current layout and a preset expected electromagnetic response, determining a rough center frequency, a rough relative bandwidth, self-coupling coefficients and coupling coefficients between adjacent components corresponding to the current layout based on S-parameter models of components in the current layout and a topology of the current layout; The processing module is configured to, for any one group of non-adjacent component pairs in the current layout, predicting coupling coefficients of the group of non-adjacent component pairs by using a prediction model corresponding to the group of non-adjacent component pairs according to sizes of components in the group of non-adjacent component pairs and relative positions of the group of non-adjacent component pairs; The processing module is configured to obtaining a rough coupling matrix corresponding to the current layout according to the self-coupling coefficients and the coupling coefficients between adjacent components corresponding to the current layout and coupling coefficients of all groups of non-adjacent component pairs in the current layout; The correction module is configured to correcting the rough coupling matrix, the rough center frequency and the rough relative bandwidth by using a current correction model to obtain a precise coupling matrix, a precise center frequency and a precise relative bandwidth; and The updating module is configured to updating the current correction model based on the precise coupling matrix, the precise center frequency and the precise relative bandwidth. The processing module is configured to take the accurate coupling matrix, the accurate center frequency and the accurate relative bandwidth as predicted values of the proxy model, select a next evaluation point based on the predicted values and their uncertainties, and generate a next generation layout; The updating module is configured to update the current correction model by using the coarse coupling matrix, the coarse center frequency, the coarse relative bandwidth and the simulated electromagnetic response.

10. A passive chip floorplan optimization apparatus, comprising: The passive chip layout optimization device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor implements the steps of the method according to any one of claims 1-8 when executing the program.

Images