Machine-learning-assisted automatic synthesis method for on-chip transformer

Abstract

Disclosed in the present invention is a machine-learning-assisted automatic synthesis method for an on-chip transformer. The method comprises: first, setting electrical parameter objectives and process constraints for an on-chip transformer; using an inverse artificial neural network to rapidly predict a set of structural parameters, which are used as an initial solution; next, performing random sampling on the structural parameters, wherein part of the sample space is obtained by means of perturbation around the initial solution, and performing full-wave electromagnetic simulation to acquire samples; compressing full-wave simulation data of the samples, and then introducing prior knowledge of the on-chip transformer to train a GPR surrogate model, such that the dimensionality of the model is reduced and the prediction accuracy is higher; and then, performing global optimization, constraining an optimization range by means of the initial solution, and further using multi-path optimization enabled by different lower confidence bound constants to enhance the convergence and exploration of an algorithm. The present invention uses an algorithm combining an inverse network and forward optimization, thereby achieving a lower computational cost and a higher optimization speed while ensuring accuracy, and promoting the automation and intellectualization of the design of on-chip passive devices.

Description

Machine learning assisted automatic synthesis method of on-chip transformer TECHNICAL FIELD

[0001] The present application belongs to the field of radio frequency passive devices, in particular relates to a machine learning assisted automatic synthesis method of on-chip transformer. BACKGROUND

[0002] On-chip passive transformer is an important component in radio frequency design, which can realize impedance matching, signal coupling, power transmission, DC isolation and other functions in radio frequency integrated circuits, and it plays an important role in reducing the number of components and chip area. On-chip spiral transformer is coupled by two spiral inductors, which is usually defined by multiple geometric parameters, and only a small part of thousands of geometric parameter combinations can meet the design specifications and design indicators at the same time, so it is difficult to derive the geometric parameters of the transformer according to the indicators. This process, which is contrary to modeling, is called synthesis problem. The synthesis of on-chip spiral transformer has always been a major challenge in passive device design, and the automatic synthesis of on-chip transformer is also a key to the automatic design of radio frequency integrated circuits.

[0003] With the continuous development of machine learning technology, the agent model assisted method seems to be the most suitable solution to the above difficulties. The main idea is to replace the expensive full-wave electromagnetic simulation model with a fast proxy model, which allows the use of population-based metaheuristic algorithms to identify the Pareto set at a lower computational cost. It is a complex and urgent problem to achieve the optimization goal under the premise of high dimension and small sample, and it is very important to find a fast and effective intelligent synthesis algorithm for the design and performance of transformer and even radio frequency integrated circuit. SUMMARY

[0004] The present application provides a machine learning assisted automatic synthesis method of on-chip transformer, which reduces the computational cost of transformer synthesis and improves the optimization speed.

[0005] Technical scheme: In order to achieve the above purpose, the technical scheme adopted by the present application is: a machine learning assisted automatic synthesis method of on-chip transformer, comprising the following steps:

[0006] Step 1: set the on-chip transformer electrical performance parameter target and transformer structure parameter constraint, and set the related process constraint;

[0007] Step 2: use reverse artificial neural network (Artificial Neural Network, ANN) to predict a group of structure parameters as initial solution according to the transformer point performance parameter target;

[0008] Step 3: Randomly sample the structure parameters, and sample the structure parameters around the initial solution, and collect the initial samples through full-wave electromagnetic simulation;

[0009] Step 4: Compress and transform the sample full-wave simulation data to reduce the sample dimension;

[0010] Step 5: Train a Gaussian Process Regression (GPR) surrogate model to learn the mapping relationship between the transformer structure parameters and the performance parameters in the sample set, instead of the traditional equivalent circuit model and the costly full-wave electromagnetic simulation;

[0011] Step 6: Use a population-based meta-heuristic algorithm for global optimization, and reverse the initial solution constraint optimization range of the artificial neural network, to obtain a set of transformer structure parameters;

[0012] Step 7: Full-wave simulation verification, full-wave simulation verification is performed on the structure parameters obtained in step 6, if the design target is met, the optimal structure parameters and / or layout are output, if not, the simulation results are added to the sample set, and steps 5 to 7 are repeated until the iteration stopping condition is met.

[0013] Further, in step 1, the transformer electrical performance parameters include some combinations of primary and secondary coil inductance, quality factor, coupling coefficient, maximum available gain, insertion loss, operating frequency, minimum self-resonant frequency, input impedance, and optimization bandwidth. The structure parameters include maximum line width, minimum line width, maximum line spacing, minimum line spacing, maximum inner diameter, minimum inner diameter, maximum number of turns, minimum number of turns, and maximum area.

[0014] Further, in step 2, an artificial neural network is used to establish a reverse model to learn the mapping relationship between the performance parameters and the structure parameters, and then a set of structure parameters is predicted as the initial solution according to the performance parameter target of the transformer synthesis. The reverse model introduces prior knowledge of on-chip transformers to reduce model dimension, and for wideband optimization inductance, important frequency points in the frequency band are used to replace all frequency points, or the model training time interval is selected; for completely symmetrical transformer structure, only the performance parameters of one of the primary or secondary coils are observed, and the input impedance only focuses on the part within the operating bandwidth; the minimum self-resonant frequency is determined by the positive and negative of the quality factor; the reverse model solves the one-to-many problem by increasing the input parameter features, and introduces auxiliary constraints in the input to limit the solution range of the output value.

[0015] Further, in step 2, the input performance parameters of the reverse model include inductance of the primary and secondary coils, quality factor, coupling coefficient, and maximum available gain, and the output structure parameters are line width, line spacing, inner diameter, and number of turns of the primary and secondary coils.

[0016] Further, in step 3, first, Latin Hypercube Sampling (LHS) is used to randomly sample transformer structure parameters, then a certain proportion of values are floated up and down in the space around the initial solution to obtain transformer structure parameters, and then prior knowledge is used for pre-screening to clean up a part of parameters that do not meet the requirements of transformer modeling, and random sampling also makes the sample cover the design space as much as possible.

[0017] Further, in step 4, in the face of the wideband transformer design problem, since there are many frequency points, directly establishing a proxy model will make the data calculation amount increase exponentially, greatly increasing the optimization time and even being unacceptable, so discrete cosine transform is used for data compression to reduce the calculation amount of the proxy model and ensure its accuracy.

[0018] Further, in step 5, the learning effect and prediction accuracy of the GPR proxy model are improved by selecting the matern3 / 2 kernel function and the fitness function; for the case that there are multiple input parameters with high sensitivity in the on-chip transformer synthesis problem, the single-objective GPR proxy model has higher accuracy and reduces the calculation amount, so each performance parameter is modeled respectively.

[0019] Further, in step 6, genetic algorithm is used for global optimization, and different Lower Confidence Bound (LCB) constants are used to bring multi-path optimization to enhance the convergence and explorability of the algorithm.

[0020] Further, in step 6, data pre-screening technology is used to modify the fitness function, so that the optimization algorithm increases sample points in the area where the optimal solution may appear, avoiding falling into local optimum. At the same time, multiple randomly generated LCB values are used for multi-path optimization to enhance the explorability of the genetic algorithm without increasing the complexity of the algorithm. The standard for LCB to judge the goodness of the new individual is:

[0021] y LCB =y(x)-ws(x)

[0022] Wherein, y(x) is the function value corresponding to the structure parameter x calculated by the GPR proxy model, s(x) is the prediction standard deviation obtained by the GPR proxy model, and w is the LCB value for measuring the exploration ability of the algorithm.

[0023] The application also provides a computer program product comprising computer programs / instructions, which, when executed by a processor, implement the steps of the machine learning assisted on-chip transformer automatic synthesis method.

[0024] Beneficial effects: compared with the prior art, the machine learning assisted on-chip transformer automatic synthesis method provided by the application has the following advantages:

[0025] (1) The application only needs the electrical performance parameter target and structure parameter constraint of the transformer to realize the automatic synthesis design of the on-chip transformer, greatly improving the work efficiency of the radio frequency circuit designer;

[0026] (2) The application establishes a reverse artificial neural network, which can quickly predict a set of initial solutions, greatly improving the design efficiency and sample reuse rate, and saving the computing resources;

[0027] (3) The application establishes a GPR proxy model instead of the traditional equivalent circuit model and the expensive full-wave electromagnetic simulation, reducing the computing cost and time;

[0028] (4) The application combines prior knowledge and data compression to train the GPR model, so that the prediction accuracy and speed of the model are improved, and the time cost is saved;

[0029] (5) The application uses a plurality of randomly generated LCB values for multi-path optimization, compared with the traditional fixed LCB value method, each iteration is randomly generated, which greatly improves the exploratory of the algorithm to avoid falling into local optimum;

[0030] (6) The application adopts the algorithm combining reverse network and forward optimization, which has lower computing cost and higher optimization speed while ensuring accuracy, promoting the automation and intelligentization of on-chip passive device design. BRIEF DESCRIPTION OF DRAWINGS

[0031] Fig. 1 is a general flowchart of the embodiment of the application;

[0032] Fig. 2 is a one-to-many problem schematic diagram of the reverse network of the application;

[0033] Fig. 3 is a schematic diagram of reverse modeling of the artificial neural network of the application;

[0034] Fig. 4 is a training process diagram of the neural network of the application;

[0035] Fig. 5 is an optimal sample coupling coefficient diagram of the optimization iteration of the embodiment of the application;

[0036] Fig. 6 is a coupling coefficient and prediction error diagram of each round of the embodiment of the application;

[0037] Fig. 7 is a primary secondary coil inductance value, quality factor and coupling coefficient diagram of the optimal sample of the embodiment of the application. DETAILED DESCRIPTION

[0038] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0039] Before the embodiments of the present application are described in detail, the basic knowledge of GPR proxy models and inverse artificial neural networks involved in the embodiments of the present application and related improvement work involved in the present application are described.

[0040] The on-chip transformer optimization problem has many parameters and high sensitivity, and a proxy model under a small sample is needed to save full-wave simulation time. GPR proxy model has good performance in machine learning aided circuit design, and the output of GPR has a confidence interval in addition to the predicted value, which is a Gaussian distribution probability function, which is of great benefit to the introduction of multi-path optimization technology. Therefore, in the machine learning aided on-chip transformer design, the online updated Gaussian process regression model is used as the proxy model, which has stable advantages in both prediction accuracy and optimization algorithm iteration number.

[0041] Gaussian process regression predicts the value of unknown points by learning the relationship between points, so the function for calculating the similarity between two points in the input space is crucial, called kernel function, also called covariance function. Many types of kernel functions can be used for GPR, such as linear, polynomial, radial basis function (RBF), Matern, periodic and rational quadratic function, each kernel function corresponds to different parameters to control the shape and scale, which maps the sample to a higher dimensional space.

[0042] Since the parameters of the kernel function directly affect the complexity of the model, different kernel functions have very different effects. However, there is no uniform conclusion on the selection of kernel functions at present, and the kernel function needs to be selected according to different problems or different application scenarios. Some scholars have conducted research in this regard, called kernel engineering. Generally, it is tried to use, and the kernel function can be combined with many optimization algorithms, and the permutation and combination often shows many unexpected performances. At the same time, the kernel function does not need to know the feature space and the conversion function, and only through the point product method for calculation, it is possible to find the data relationship in a high-dimensional space with very low calculation cost.

[0043] The most common kernel function in Gaussian process is RBF kernel, also known as square exponential kernel or Gaussian kernel, defined as follows:

[0044]

[0045] Σ is the covariance of each feature in the observation, if Σ is a diagonal matrix, the formula can be simplified as:

[0046]

[0047] where σ j is the importance of the jth dimension of the defined space, D is the total number of dimensions. If σ j →∞, the jth dimension can be almost ignored. If Σ is isotropic, the isotropic squared exponential kernel is:

[0048]

[0049] where σ 2 is called the bandwidth, the kernel function has infinite spatial dimensions.

[0050] The Matern kernel function is another common kernel for GPR, which links the distance and similarity between training data, and its specific form is as follows:

[0051]

[0052] where r = ||x-x' ||, v>0, l>0, K v is the modified Bessel function, Γ(v) is the gamma function. When v→∞, the Matern kernel function approaches the squared exponential kernel. When , the formula can be simplified as:

[0053]

[0054] Compared with other kernel functions, it has an additional parameter smoothing factor, which is used to control the smoothness of the result function. According to the difference of the smoothing factor, the commonly used are matern5 / 2 and matern3 / 2, and when each predictor has a separate fixed-length parameter, there are ardmatern5 / 2 and ardmatern3 / 2 kernels.

[0055] In order to select the appropriate kernel function, a set of transformer synthesis experiments are carried out below. The transformer structure adopts the staggered interwinding quadrilateral, and the synthesis target is the primary and secondary coil inductance L p = L s = 0.15nH, the coupling coefficient k = 0.7, the self-resonant frequency f sr = 80GHz, and the maximum available gain G max maximization at the working frequency point. The input parameters are the line width W, the line spacing S, the outer diameter D out , the primary and secondary coil turns N p , N s , since the primary and secondary coils of the staggered interwinding structure are completely symmetrical, this experiment can only focus on the inductance L p and the quality factor Q pThe secondary coil is the same, in order to ensure the n:n turns ratio, the primary and secondary coils select the same number of turns.

[0056] Table 1

[0057]

[0058] In the present application, experiments are all carried out by using 40nm CMOS process, and it can be understood that the present application is not limited to a specific process, and is applicable to different processes. Considering the requirements of chip design, in addition to the structural constraints of various parameters, the input parameters should also meet the constraints in Table 1. The lower limit of the numerical value is closely related to the chip design rules, and the upper limit is as large as possible. In addition to experienced engineers, most beginners and people with no relevant design experience cannot know the approximate range of geometric parameters in advance, and this is one of the design intentions of the present application. Two layers of metal are used, and the main metal is distributed in the MTT2 layer, and the TM1 layer is used to lead out the inner circle port.

[0059] The algorithm runs on a server with two AMD EPYC 7F52 processors and 512GB of memory, without parallelism. In order to make the simulation of EMX more accurate, the sweep range is set to [0, f sr +10]GHz, with a step size of 100MHz.

[0060] Table 2

[0061]

[0062] The experimental results are shown in Table 2, and various kernel functions can solve the on-chip transformer synthesis problem, and meet the requirements of various indicators at the working frequency point. However, compared with RBF kernel, the optimization effect is the worst, and each predictor has a separate fixed-length parameter ardmatern5 / 2 and ardmatern3 / 2 and matern5 / 2 result is similar, and matern3 / 2 shows obvious advantages. Therefore, the on-chip transformer synthesis method described below uses matern3 / 2 as the kernel function of the Gaussian regression process.

[0063] For a multi-objective synthesis problem, it is essentially an inverse problem, which solves the structure parameters from the performance target. Traditional forward optimization often needs multiple iterations to converge, and the calculation cost is high. In addition, the inverse model is another very potential alternative solution for the optimization task required to solve the inverse problem. Unlike the common optimization method based on forward mapping, the inverse model learns the relationship between performance parameters and structure parameters, rather than the mapping of structure parameters to performance parameters, which can directly predict the corresponding value of the structure parameter according to the performance index, without any iteration algorithm based on the forward model. Therefore, the inverse modeling method has an advantage in calculation time.

[0064] The present application uses an artificial neural network (ANN) to establish an inverse model. Artificial neural networks are known for their predictive capabilities, and they can provide a high degree of fitting to new data after training. By introducing prior knowledge into the ANN model, the accuracy and requirements of large data samples can be improved.

[0065] Although the inverse model has the advantage of fast prediction rather than iterative optimization, there are also some extremely challenging problems. First, the construction of the inverse model of the microwave design problem is quite complex. Depending on the different desired targets, especially when it involves wideband problems, the dimensionality of the input space is very high. Therefore, the training of the inverse model requires a large amount of full-wave simulation data as samples. Compared with the optimization of the forward proxy model with a small sample, the large amount of "preparation work" cannot be ignored. This is less efficient for single prediction optimization, but if the inverse model is used multiple times, the training time is spread out to each use, which has a great advantage.

[0066] Another urgent problem to be solved is that the mapping relationship of the inverse model is often non-single. For the forward model, a set of structure parameters corresponds to only one set of performance parameters, and there is only one solution. However, for the inverse model, a set of performance parameters can be obtained from multiple sets of structure parameters, as shown in FIG. 2, which means that the inverse model usually has a one-to-many problem. Generally, it is not possible to train a model to fit multiple output values for the same input value, so the model training error will be high and the model accuracy will be poor. Therefore, in many actual inverse problems, it is difficult to train the model.

[0067] To solve the above two problems and the actual transformer synthesis application scenario, the present application proposes the following solutions. First, introduce prior knowledge to reduce model dimensionality. For a wideband optimization problem, the traditional and classic method is still to increase the frequency feature dimension on the original input parameters, but this will increase the training difficulty of the inverse model. This paper combines the prior knowledge of the on-chip transformer with the inverse model. Prior knowledge helps to avoid getting stuck in local minimum and improves the speed of finding the global minimum. At the same time, with less training data, the model can work well. It is known that the inductance value is relatively stable when far away from the self-resonant frequency and mutual inductance resonance frequency. For wideband optimization, the inductance value can be replaced by several important frequency points in the frequency band, such as the maximum value, the minimum value, the value at the center frequency of the bandwidth, etc. Alternatively, the training model can select frequency points at intervals, and sparse sampling without affecting the overall trend. Of course, this method requires modeling of the resonance frequency, and both are indispensable. In addition, for a completely symmetrical transformer structure, only the performance parameters of the primary or secondary coil are observed, the input impedance is only concerned with the part within the working bandwidth, the minimum self-resonant frequency is determined by the positive and negative quality factors, etc. In short, combining the characteristics of the transformer itself will not excessively reduce the accuracy of the inverse model, and can greatly reduce the dimensionality to reduce the computational complexity of the model.

[0068] Second, the one-to-many problem is solved by adding input parameter characteristics. In practical applications, a set of performance parameter optimization objectives can correspond to multiple structure parameters, but the performance parameters corresponding to these structure parameters will inevitably differ in the entire frequency range. A specific auxiliary constraint is introduced in the input to limit the solution range of the output value. Taking on-chip transformer design as an example, there can be multiple comprehensive results when the primary and secondary coil inductance, quality factor, and coupling coefficient are the same at a certain operating frequency. However, if the frequency characteristic dimension is added at this time, appropriately increasing the dimension space will filter out a part of the samples, or if the input impedance target requirement is added at this time, the one-to-many problem can also be converted into a one-to-one problem for solving.

[0069] Next, the specific steps of using ANN to realize on-chip transformer synthesis are introduced through an example. The neural network used is a two-layer network with 10 hidden layers, and the model schematic diagram is shown in FIG. 3. The input is the performance parameters of the primary and secondary coil inductance L p , L s , quality factor Q p , Q s , coupling coefficient k, and maximum available gain G max , and the output is the structure parameters of the line width W, line spacing S, inner diameter D in , and coil turns N p , N s .

[0070] The transformer synthesis target is that the primary and secondary coil inductance L p = L s = 0.15 nH, the upper and lower fluctuations in the bandwidth are not more than 5%, the coupling coefficient k = 0.7, the upper and lower fluctuations in the bandwidth are not more than 5%, the minimum self-resonant frequency f sr = 80 GHz, and the maximum available gain G max in the bandwidth is maximized.

[0071] Before prediction, the network is first trained. 1000 transformer samples are randomly generated, the structure is single-end staggered mutual winding quadrilateral, the target frequency is 45-55 GHz, and the maximum and minimum values in the 45-55 GHz operating band are selected for each input parameter to reduce the dimension. Among them, 700 samples are used for training the model, 150 samples are used for verification, and 150 samples are used for testing. The ANN training process is shown in FIG. 4.

[0072] The prediction error of the model is measured using the mean square error, and the calculation formula is:

[0073]

[0074] where n is the number of samples, y i is the simulation value of the i th sample, is the prediction value of the i-th sample. The prediction error of the neural network is shown in Table 3, and the MSE is stabilized at about 0.6.

[0075] Table 3

[0076]

[0077] After the neural network training is completed, the prediction is performed according to the comprehensive target input: L p = [0.1545 0.1455], L s = [0.1545 0.1455], k = [0.721 0.679], G max = [0.89 0.86]. At this time, the target value of G max is an index that needs to be reached through multiple iterations of manual design.

[0078] After multiple experiments, the average prediction time is 0.008 seconds, and the output solution is as follows:

[0079] Table 4

[0080]

[0081] The structure parameters obtained by prediction are verified through full-wave simulation, and the performance parameters are as follows:

[0082] Table 5

[0083]

[0084] It should be noted that the data listed in Table 5 are the minimum coupling coefficients, the minimum G max in the frequency band, and the inductance value with the largest deviation from the target value. Although the maximum deviation error of the inductance value in the wide frequency band is 5.3%, other targets meet the requirements, and the inductance value is relatively close to the target value. It is worth noting that, in addition to the training time, the single prediction time of the inverse network is much smaller than the iteration optimization time of the forward model, and it still has important reference significance for automatic synthesis design of on-chip transformers.

[0085] Next, taking the maximum coupling coefficient in the wide frequency band as an example, a machine learning assisted automatic synthesis method for on-chip transformers disclosed in the embodiment of the application is described in detail.

[0086] The comprehensive target is that the primary and secondary coil inductance L p = 0.5 nH, L s = 0.5 nH, with a floating range of no more than 10% in the bandwidth, the upper limit of the bandwidth f = 32 GHz, the primary and secondary coil quality factors Q p > 5, Q s > 5, and the minimum self-resonant frequency fsr = 45GHz, seeking to maximize the bandwidth of the coupling coefficient k, so the problem needs to be optimized

[0087] max(k 24-32GHz )

[0088]

[0089] where L p , L s is the inductance value of 24-32GHz obtained by full-wave simulation, Q pf , Q sf is the quality factor at 32GHz, Q pfsr , Q sfsr is the quality factor at 45GHz, and the quality factor at the minimum self-resonant frequency is greater than 0, which is considered not to be resonated.

[0090] As shown in FIG. 1, the machine learning assisted on-chip transformer automatic synthesis method disclosed in the embodiment of the application comprises the following steps:

[0091] Step 1: Set the on-chip transformer electrical parameter target and transformer structure parameter constraint, and set the related process constraint.

[0092] In the embodiment, the related process constraint includes the name of the process to be declared, the name of the metal layer to be used, the constraint group name, etc. The electrical parameter target is slightly different according to the difference of the optimization target, and in general, it includes the primary and secondary coil inductance L p , L s , the quality factor Q p , Q s , the coupling coefficient k, the maximum available gain G max , the insertion loss IL, the working frequency f, the minimum self-resonant frequency f sr , the input impedance Z in and the optimization bandwidth BW. The structure parameters include the maximum line width W max , the minimum line width W min , the maximum line spacing S max , the minimum line spacing S min , the maximum inner diameter D in,max , the minimum inner diameter D in,min , the maximum number of turns N max , the minimum number of turns N min and the maximum area A max , and at the same time, the user can also select the desired transformer structure. The symmetric multilayer transformer geometric parameter constraint is shown in Table 6.

[0093] Table 6

[0094]

[0095] Step 2: A set of structure parameters is predicted as the initial solution using the inverse artificial neural network.

[0096] In this embodiment, the relationship between the performance parameters and the structure parameters of the on-chip transformer is learned using an artificial neural network. After the network training is completed, only a set of performance parameter targets need to be input to quickly solve a set of structure parameter solutions. The subsequent sampling and optimization range are related to the initial solution.

[0097] Step 3: A part of the structure parameters is randomly sampled, and the structure parameters are sampled around the initial solution, and the initial samples are collected through full-wave electromagnetic simulation.

[0098] In this embodiment, the classic sampling method of Latin hypercube sampling is used to randomly sample the structure parameters, a part of which is obtained within a 30% space around the initial solution, and the sampling number is 80. The initial samples are collected through full-wave electromagnetic simulation. On the one hand, since full-wave simulation is very time-consuming, the number of samples of the initial surrogate model needs to be as small as possible, otherwise the algorithm efficiency will decrease. On the other hand, the design space needs to be covered as much as possible, and too sparse sampling will result in very little information in some areas, and the reliability of the surrogate model is poor.

[0099] Step 4: The sample full-wave simulation data is compressed to reduce the sample dimension, which is convenient for the training of the surrogate model.

[0100] In this embodiment, for the synthesis of wide-band on-chip transformers, since there are many frequency points, directly establishing a surrogate model will cause the data calculation to increase exponentially, greatly increasing the optimization time and even being unacceptable. Therefore, discrete cosine transform is used for data compression to reduce the calculation amount of the surrogate model and ensure its accuracy.

[0101] Step 5: Train the GPR surrogate model to learn the mapping relationship between the transformer structure parameters and the performance parameters in the sample set, instead of the traditional equivalent circuit model and the costly full-wave electromagnetic simulation.

[0102] In this embodiment, prior knowledge is combined when training the surrogate model, which can make the model have higher prediction accuracy, reduce the training and prediction time of the surrogate model, and improve the efficiency and performance. For example, for a completely symmetrical interleaved mutual winding structure transformer, this embodiment only models and predicts the inductance and quality factor of the primary coil, reducing the number of models by half. When optimizing a wide-band transformer, the quality factor is relatively smooth overall, and modeling the upper limit quality factor of the bandwidth can also reduce the model calculation amount. The GPR surrogate model uses the matern3 / 2 kernel function. For the case that there are multiple input parameters with high sensitivity in the on-chip transformer synthesis problem, the single-target GPR surrogate model has higher accuracy and reduces the calculation amount, so each performance parameter is modeled separately.

[0103] Step 6: Global optimization is performed by using population-based meta-heuristic algorithm, and the initial solution of the reverse network is constrained to optimize the range, and a set of transformer structure parameters is obtained by optimization.

[0104] In this embodiment, genetic algorithm is used for global optimization, the initial solution of the reverse network is constrained to optimize the range (floating up and down 30% around the initial solution), data pre-screening technology is used to modify the fitness function, and the multi-path optimization brought by different lower limit constants of confidence enhances the convergence and exploration of the algorithm, the results of multi-path optimization are compared and the optimal sample is output, and a set of transformer structure parameters is obtained by optimization.

[0105] Specifically, the fitness function is modified by using data pre-screening technology, so that the optimization algorithm increases sample points in the area where the optimal solution may appear, and avoids falling into local optimum. At the same time, multiple LCB values randomly generated in [-2, 2] are used for multi-path optimization, which enhances the exploration of genetic algorithm without increasing the complexity of the algorithm. The standard for LCB to judge the goodness of the new individual is:

[0106] y LCB = y(x) - ws(x)

[0107] Where y(x) is the function value corresponding to x calculated by the GPR surrogate model, s(x) is the prediction standard deviation obtained by the GPR surrogate model, and w is the LCB value that measures the exploration ability of the algorithm.

[0108] Step 7: Full-wave simulation verification is performed on the structure parameters obtained in step 6, and if the design target is met, the optimal layout is output, otherwise the simulation results are added to the sample set, and steps 5 to 7 are repeated until the iteration stopping condition is met. The optimization termination condition is set to optimize to the maximum number of iterations or each parameter reaches the expected optimization target. When the optimization target contains maximization or minimization, the maximum or minimum value can be set to remain unchanged for a certain number of generations and within the maximum number of iterations, so that the transformer is fully optimized.

[0109] The optimal sample of the optimization iteration described in the embodiment of the application is shown in Figure 5. It is worth noting that the minimum coupling coefficient within the optimal sample bandwidth is shown in Figure 5, that is, all k within the bandwidth are greater than this value. In the initial stage of optimization, the initial solution of the inverse network prediction is the optimal sample, which also shows the powerful prediction ability of the artificial neural network. However, the optimal sample at this time is only the sample closest to the optimization target in the sample set, and does not necessarily fully meet the conditions. Therefore, after several rounds of optimization, the optimal sample changes and the coupling coefficient decreases, which shows that in the optimization process, samples that generally meet the target are found, because in this example, there are other constraints when maximizing the coupling coefficient. After 16 rounds of exploration, the algorithm finds a sample that meets all other requirements and has the maximum coupling coefficient. Continue to optimize, while meeting the constraints, the coupling coefficient gradually increases, until the 22nd round of optimization, the algorithm converges, the optimal sample does not change for 10 consecutive rounds, the optimization process ends, and the optimal solution is output.

[0110] In this embodiment, the root mean square error is used to measure the prediction accuracy of the surrogate model, and the calculation formula is

[0111]

[0112] Wherein is the output prediction matrix, n v is the number of samples for verification, n i is the number of frequency points before the self-resonant frequency of the sample, is the prediction value of the jth frequency point of the ith sample, x ij is the simulation value of the jth frequency point of the ith sample.

[0113] The minimum coupling coefficient within the bandwidth obtained in each round of optimization and its prediction error are shown in Figure 6. It can be seen that as the number of optimization increases, the surrogate model is updated online, the prediction accuracy of the model is higher and higher, and the prediction error within the entire bandwidth gradually decreases, which also shows the effectiveness of the GPR surrogate model.

[0114] The final optimized transformer structure parameters are shown in Table 7.

[0115] Table 7

[0116]

[0117] The performance parameters of this symmetric laminated 2:2 structure transformer are shown in Figure 7. The primary and secondary inductance values are around 0.5nH in the 24-32GHz frequency band, with a deviation of no more than 10%, which is relatively stable. The quality factors at 32GHz are Q p = 7.87 and Q s= 12.02, broadband coupling coefficient k > 0.832. It should be noted that the winding mode of the primary and secondary coils of the symmetric 2:2 structure makes them distributed in different metal layers, and the metal thickness and distance from the substrate are different, and the loss is also different, so their quality factors are different, but the overall trend is the same.

[0118] The embodiments of the present application also disclose a computer program product comprising computer programs / instructions which, when executed by a processor, implement the steps of the machine learning assisted automatic synthesis method of on-chip transformer disclosed by the foregoing embodiments. The program / instruction code for implementing the method of the present application can be written in any combination of one or more programming languages. These program / instruction codes can be provided to the processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, so that the program / instruction codes, when executed by the processor or controller, cause the steps of the method of the present application to be implemented. The program / instruction codes can be executed entirely on a machine, partially on a machine, partially on a machine as a separate software package, and partially on a remote machine, or entirely on a remote machine or server.

[0119] The details of the present application not described are known to those skilled in the art.

[0120] To sum up, only the specific embodiments of the present application, but the protection scope of the present application is not limited to this, any skilled in the art of the present application disclosed in the technical range, can easily think of changes or replacement, should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A machine learning assisted on-chip transformer automatic synthesis method, characterized in that, The method comprises the following steps: Step 1: setting the on-chip transformer electrical performance parameter target and transformer structure parameter constraint, setting the related process constraint; Step 2: using the inverse artificial neural network to predict a group of structure parameters as the initial solution according to the transformer electrical performance parameter target; wherein the inverse model is established by using the artificial neural network, the mapping relationship between the performance parameters and the structure parameters is learned, the inverse model introduces the prior knowledge of the on-chip transformer to reduce the model dimension, for the bandwidth optimization inductance, the more important frequency points in the frequency band are used to replace all the frequency points, or the model training time interval is selected; for the completely symmetrical transformer structure, only the performance parameters of one coil of the primary or secondary coil are observed, and the input impedance only focuses on the part in the working bandwidth; the minimum self-resonant frequency is determined by the positive and negative of the quality factor; the inverse model solves the one-to-many problem by increasing the input parameter characteristics, and the auxiliary constraint is introduced in the input to limit the solution range of the output value; Step 3: a part of the structure parameters are obtained by random sampling, and a part of the structure parameters are obtained by sampling around the initial solution, and the initial sample is collected by full-wave electromagnetic simulation; Step 4: the sample full-wave simulation data is compressed and transformed to reduce the sample dimension; Step 5: the GPR proxy model is trained to learn the mapping relationship between the transformer structure parameters and the performance parameters, instead of the traditional equivalent circuit model and full-wave electromagnetic simulation; Step 6: a population-based meta-heuristic algorithm is used for global optimization, the inverse artificial neural network initial solution is constrained to optimize the range, and a group of transformer structure parameters are obtained by optimization; Step 7: full-wave simulation verification is performed on the structure parameters obtained in step 6, if the design target is met, the optimal structure parameters and / or layout are output, if not, the simulation results are added to the sample set, and steps 5 to 7 are repeated until the iteration stopping condition is met.

2. The machine learning assisted on-chip transformer auto- synthesis method of claim 1, wherein, In step 1, the transformer electrical performance parameters include part of the combination of the primary and secondary coil inductance, quality factor, coupling coefficient, maximum available gain, insertion loss, working frequency, minimum self-resonant frequency, input impedance and optimized bandwidth; the structure parameters include maximum line width, minimum line width, maximum line spacing, minimum line spacing, maximum inner diameter, minimum inner diameter, maximum coil turns, minimum coil turns and maximum area.

3. The machine learning assisted on-chip transformer auto- synthesis method of claim 1, wherein, In step 2, the input performance parameters of the inverse model include the primary and secondary coil inductance, quality factor, coupling coefficient and maximum available gain, and the output structure parameters are line width, line spacing, inner diameter and primary and secondary coil turns.

4. The machine learning assisted on-chip transformer auto- synthesis method of claim 1, wherein, In step 3, Latin hypercube random sampling is first used to obtain transformer structure parameters, then random sampling is performed within a certain proportion of space around the initial solution to obtain transformer structure parameters, and then prior knowledge is used for pre-screening.

5. The machine learning assisted on-chip transformer automatic synthesis method of claim 1, wherein, In step 4, the data is compressed by using discrete cosine transform.

6. The machine learning assisted on-chip transformer automatic synthesis method of claim 1, wherein, In step 5, the GPR proxy model selects the matern3 / 2 kernel function, and a single-target GPR proxy model is established for each performance parameter.

7. The machine learning assisted on-chip transformer automatic synthesis method of claim 1, wherein, In step 6, the genetic algorithm is used for global optimization, and the multi-path optimization enhancement algorithm is used to improve the convergence and exploration of the algorithm with different lower confidence limit LCB constants.

8. The machine learning assisted on-chip transformer automatic synthesis method of claim 1, wherein, In step 6, the fitness function is modified using data pre-screening technology, so that the optimization algorithm increases sample points in the area where the optimal solution may appear, avoiding falling into local optimum; at the same time, multiple randomly generated LCB values are used for multi-path optimization, and the standard for judging the goodness of the new individual is: y LCB =y(x)-ws(x); wherein y(x) is the function value corresponding to the structure parameter x calculated by the GPR surrogate model, s(x) is the prediction standard deviation obtained by the GPR surrogate model, and w is the LCB value measuring the exploration ability of the algorithm.

9. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions, when executed by the processor, implement the steps of the machine learning assisted on-chip transformer automatic synthesis method according to any one of claims 1-8.

Images