AI-assisted high-frequency broadband circuit parasitic effect elimination method and device
By combining traditional de-embedding algorithms with neural network models, and utilizing measurement parameters of open-circuit, short-circuit, and through-circuit structures, a parasitic network model is constructed and data correction is performed. This solves the problem of insufficient integration between neural networks and traditional open-short-thru processes, and improves the measurement accuracy and device characteristic extraction accuracy in the high-frequency band.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV
- Filing Date
- 2026-02-11
- Publication Date
- 2026-05-12
AI Technical Summary
In existing technologies, neural networks cannot be effectively combined with traditional open-short-thru de-embedding processes, resulting in decreased measurement accuracy in high-frequency bands. Traditional algorithms fail in high-frequency bands, and neural networks fail to fully utilize the optimized data generated by traditional algorithms, making it difficult to improve training and correction efficiency.
By concatenating traditional de-embedding algorithms with neural network models, and utilizing the measurement parameters of open-circuit, short-circuit, and through-circuit structures, a parasitic network model is constructed for preliminary de-embedding processing. The data is then input into a pre-trained neural network model for correction, and the target device parameters are output.
It improves the modeling accuracy and de-embedding capability of high-frequency parasitic effects, enhances the extraction accuracy of target device parameters, reduces measurement errors and uncertainties, and achieves accurate extraction of high-frequency device characteristics.
Smart Images
Figure CN122017511A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of semiconductor testing and radio frequency microwave measurement technology, and in particular to an AI-assisted method and apparatus for eliminating parasitic effects in high-frequency broadband circuits. Background Technology
[0002] In radio frequency, microwave, and semiconductor testing, extracting the characteristics of high-frequency devices presents challenges. Test fixtures, pads, and other structures introduce parasitic effects, causing measurement results to deviate from the intrinsic characteristics of the device. Accurately separating the true characteristics of the target device from the overall results containing fixture effects has become a key requirement for modeling.
[0003] To address this need, de-embedding algorithms are used to remove parasitic effects. These algorithms can be categorized into one-step, two-step, and three-step types based on the number of test structures used. One-step methods include the "open" method based on open-circuit structures; two-step methods include the "open-short" method combining open and short circuits; and three-step methods include "open-short-thru" and "open-short-load" methods. Furthermore, neural network technology has been introduced into this field to improve accuracy. For example, some methods directly implement de-embedding based on artificial neural networks, eliminating the need for traditional calibration components. Other approaches combine deep neural networks with TRL de-embedding to reduce the requirements for chip area and probe placement.
[0004] However, the aforementioned neural network-based solutions have not yet been effectively integrated with the "open-short-thru" de-embedding process. This results in a decrease in the accuracy of the traditional "open-short-thru" method at high frequencies due to model limitations, while the neural network also fails to fully utilize the optimized data generated by traditional algorithms to improve its training and correction efficiency. Summary of the Invention
[0005] This application provides an AI-assisted method and apparatus for eliminating parasitic effects in high-frequency broadband circuits, in order to solve the problem that neural networks and open-short-thru cannot be effectively combined.
[0006] In a first aspect, this application provides an AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits, comprising: Acquire the first measurement parameter, the second measurement parameter, and the operating condition parameter of the target device, wherein the second measurement parameter is the measurement parameter of the target device in open circuit structure, short circuit structure, and through structure; Based on the second measurement parameters, the first measurement parameters are preliminarily de-embedded using a de-embedding algorithm to obtain preliminary de-embedded parameters; The data set of the first measurement parameter, the second measurement parameter, the operating condition parameter, and the preliminary de-embedding parameter is input into the target neural network model to output the target device parameter through the target neural network model.
[0007] In some feasible embodiments, the target neural network model is a pre-trained neural network model; The method further includes: Obtain a training sample set, which includes first measurement parameters, second measurement parameters, operating condition parameters, preliminary de-embedding parameters, and real network parameters of multiple sample devices; Build a basic neural network model; The training sample set is used as input features, and the real network parameters are used as training targets to train the basic neural network model to obtain the neural network model.
[0008] In some feasible embodiments, the real network parameters are obtained by simulation software under target de-embedding conditions.
[0009] In some feasible embodiments, after outputting the target device parameters through the target neural network model, the method further includes: The parameters of the target device are subjected to jump detection and smoothing to obtain correction parameters; The correction parameters are determined to be the target device parameters.
[0010] In some feasible embodiments, the step of performing jump detection and smoothing processing on the target device parameters to obtain correction parameters includes: Detect the gradient of parameter change at each frequency point in the sequence of parameter changes of the target device with frequency; When the gradient of parameter change at a certain frequency point exceeds the target gradient, the frequency point is marked as an abnormal jump point. The target gradient is a threshold determined by the overall change gradient of the sequence or the change gradient of neighboring frequency points. The parameter values at the abnormal jump points are smoothed to control the frequency response curve of the target device parameters to be continuously smooth, thereby obtaining the correction parameters.
[0011] In some feasible embodiments, the preliminary de-embedding process of the first measurement parameter based on the second measurement parameter using a de-embedding algorithm includes: Based on the measurement parameters of the open-circuit structure, short-circuit structure, and straight-through structure, parasitic network models corresponding to the open-circuit structure, short-circuit structure, and straight-through structure are constructed. The parasitic network model is used to perform de-embedding operations on the first measurement parameters to remove the parasitic effects contained in the first measurement parameters, thereby obtaining the preliminary de-embedding parameters.
[0012] In some feasible embodiments, the target neural network model is a GRNN.
[0013] In some feasible embodiments, the operating condition parameters include the operating frequency and the bias voltage; The step of outputting target device parameters through the target neural network model includes: The operating frequency and the bias voltage are used as input features in the data set, and are input to the target neural network model along with the first measurement parameter, the second measurement parameter, and the preliminary de-embedding parameter. The target neural network model corrects the parameters of the target device according to the operating frequency and the bias voltage, so as to output the target device parameters through the target neural network model.
[0014] Secondly, this application provides an AI-assisted high-frequency broadband circuit parasitic effect elimination device, comprising: The acquisition module is used to acquire the first measurement parameter, the second measurement parameter, and the operating condition parameter of the target device, wherein the second measurement parameter is the measurement parameter of the target device in open circuit structure, short circuit structure, and through structure. The processing module is used to perform preliminary de-embedding processing on the first measurement parameters based on the second measurement parameters using a de-embedding algorithm to obtain preliminary de-embedding parameters; The de-embedding module is used to input the data set of the first measurement parameter, the second measurement parameter, the operating condition parameter, and the preliminary de-embedding parameter into the target neural network model, so as to output the target device parameter through the target neural network model.
[0015] Thirdly, this application provides a computing device including a processor and a memory, the processor being configured to execute instructions stored in the memory to cause the computing device to perform the method as described in the first aspect.
[0016] As can be seen from the above technical solutions, this application provides an AI-assisted method and apparatus for eliminating parasitic effects in high-frequency broadband circuits. The method includes: acquiring a first measurement parameter, a second measurement parameter, and operating condition parameters of a target device, wherein the second measurement parameter is the measurement parameter of the target device in an open-circuit structure, a short-circuit structure, and a through-circuit structure; based on the second measurement parameter, using a de-embedding algorithm to perform preliminary de-embedding processing on the first measurement parameter to obtain preliminary de-embedding parameters; inputting the data set of the first measurement parameter, the second measurement parameter, the operating condition parameters, and the preliminary de-embedding parameters into a target neural network model to output target device parameters through the target neural network model. This method incorporates traditional algorithms as part of data preprocessing and feature generation, enabling the neural network to learn and correct using the partially optimized preliminary de-embedding parameters generated by traditional algorithms. This not only avoids the difficulty of the neural network learning directly from complex raw data and improves training efficiency, but also enhances the comprehensive ability to model and remove high-frequency parasitic effects through two-stage cascaded processing, thereby improving the accuracy of the extracted target device parameters. Attached Figure Description
[0017] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart illustrating the AI-assisted high-frequency broadband circuit parasitic effect elimination method provided in this application embodiment; Figure 2 Error table for RMSE index provided in the embodiments of this application; Figure 3 The present application provides a schematic diagram comparing the real part de-embedding before and after step S11 in the embodiments of this application; Figure 4 The present application provides a schematic diagram comparing the process before and after the removal of the imaginary part in step S11. Figure 5 The present application provides a schematic diagram comparing the real part de-embedding in step S12 before and after the implementation of this application. Figure 6 The present application provides a schematic diagram comparing the process before and after the removal of the imaginary part in step S12. Figure 7 The embodiments of this application provide a comparative schematic diagram of the real part de-embedding before and after S21; Figure 8 The present application provides a schematic diagram comparing the process before and after the illusory part removal in step S21. Figure 9The embodiments of this application provide a comparative schematic diagram of the real part de-embedding before and after S22; Figure 10 The provided embodiment of this application provides a comparative schematic diagram of the illusory part removal process before and after step S22. Detailed Implementation
[0019] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following examples do not represent all embodiments consistent with this application.
[0020] With the rapid development of communication and high-performance computing chips, the operating frequency of electronic devices is constantly increasing to the millimeter-wave and even terahertz range. At this point, any structure such as test fixtures, pads, interconnects, and package transitions will introduce significant parasitic effects, causing the directly measured electrical parameters to deviate from the intrinsic characteristics of the device. Under high-frequency conditions, how to accurately separate the true characteristics of the device under test (DUT) from the overall measurement results including the influence of the test fixture has become a critical problem that urgently needs to be solved in device modeling.
[0021] De-embedding algorithms can be classified according to the number of steps involved in the de-embedding process: one-step de-embedding, two-step de-embedding, three-step de-embedding, and de-embedding with more steps. One-step de-embedding includes open de-embedding based solely on open-circuit test structures; two-step de-embedding includes open-short de-embedding, which adds short-circuit test structures to the open-circuit test structure; and three-step de-embedding includes open-short-through de-embedding, open-short-load de-embedding, and pad-open-short de-embedding. These traditional de-embedding algorithms are effective under specific conditions, but they have the following drawbacks: High-frequency de-embedding effect degradation: At high frequencies, the equivalent circuit assumptions of lumped parameter-based models no longer hold, leading to model failure. For example, the open de-embedding method only considers the coupling effect between the input and output pads. When the frequency is higher than 10 GHz, the series parasitics of the transmission line cannot be ignored, so this de-embedding method is only applicable below 10 GHz. For three-step de-embedding methods such as open-short-thru and pad-open-short, as the signal frequency increases to the millimeter-wave range, the parasitic electromagnetic coupling effect becomes more complex, and using three de-embedding structures becomes less accurate.
[0022] De-embedding effect diminishes as transistor size shrinks: Traditional algorithms rely on accurate and ideal test structure models, but non-ideal parasitic effects are common in actual tests. Some parasitic effects become more pronounced as transistor size shrinks, thus diminishing the de-embedding effect. For example, as the transistor size shrinks further to the tens of nanometers level and the operating frequency increases to hundreds of gigahertz, the open / short-circuit port effects and the substrate coupling path between the input and output pads ignored in the above methods will have a significant impact on the de-embedding results.
[0023] Measurement uncertainty introduces errors: Each step of deembedding typically introduces measurement errors. As the number of deembedding steps increases, the measurement errors eventually lead to a sharp increase in uncertainty in the deembedding results, which poses a huge challenge to accurate measurement.
[0024] Currently, there is considerable research applying neural networks to on-chip testing and modeling of devices, such as ANN-based de-embedding methods, which achieve higher de-embedding accuracy than traditional 1X-AFR and do not require the design and measurement of calibration components. To address the issues of increased chip area due to the need for multiple transmission lines of varying lengths in traditional TRL methods, and the impact of probe placement on calibration results, the TRL-NN method was proposed, combining deep neural networks (DNNs) with traditional RTL de-embedding methods to achieve even higher de-embedding accuracy. However, a de-embedding method that combines neural networks with traditional open-short-thru methods is still lacking.
[0025] This application provides an AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits. By combining traditional algorithms with a neural network model, it achieves high-precision extraction of network parameters for target devices. Figure 1 As shown, it includes the following steps: S100: Obtain the first measurement parameter, the second measurement parameter, and the operating condition parameter of the target device. The second measurement parameter is the measurement parameter of the target device in open circuit structure, short circuit structure, and through structure.
[0026] The target device is a semiconductor device or radio frequency / microwave component from which its intrinsic network parameters are to be extracted. In wafer testing, the target device is fabricated on the test wafer, and its electrodes are connected to the measurement system via pads or probe points. The target device can be an active device, such as a transistor, or a passive device, such as an inductor, capacitor, or transmission line.
[0027] The first measurement parameter is the network parameter obtained by directly measuring the target device in its operating state. It reflects the overall response, including the intrinsic characteristics of the target device and parasitic effects introduced by test fixtures, interconnects, pads, etc. The first measurement parameter is the scattering parameter, or S-parameter, which can be measured by scanning within a set frequency range using a vector network analyzer.
[0028] The second measurement parameter consists of network parameters measured under identical test conditions using standardized test structures designed and manufactured for de-embedding calibration. These standard test structures are fabricated on the same process substrate as the target device and are used to characterize parasitic networks. The second measurement parameter includes measurement data for at least open-circuit, short-circuit, and through-circuit structures.
[0029] The open-circuit structure has the same input / output pads as the target device, but its ends are open; the short-circuit structure is grounded at its ends; and the through structure is a straight-through transmission line connecting the two pads. Measuring these structures allows us to obtain the frequency response characteristics of the parasitic network.
[0030] Operating condition parameters are information about the electrical and environmental conditions under which a target device operates during measurement. For radio frequency (RF) and microwave devices, operating condition parameters include at least the operating frequency of the signal. The operating frequency is typically a frequency range or a series of discrete frequency points. For active devices, operating condition parameters also include the power supply and bias voltages applied to the device, which determine the device's operating point and have a significant impact on its network parameters.
[0031] S200: Based on the second measurement parameters, the first measurement parameters are preliminarily de-embedded using a de-embedding algorithm to obtain preliminary de-embedded parameters.
[0032] De-embedding algorithms are mathematical processing methods used to remove or correct parasitic effects from measurement data containing parasites in order to approximate the intrinsic characteristics of a device. They utilize known standard test structure parameters to model and inversely eliminate the influence of parasitic networks. In this embodiment, the de-embedding algorithm is a traditional de-embedding method that requires the use of measurement data from open-circuit, short-circuit, and through-circuit structures.
[0033] In some embodiments, based on the second measurement parameters, a de-embedding algorithm is used to perform preliminary de-embedding processing on the first measurement parameters, including: constructing parasitic network models corresponding to the open-circuit structure, short-circuit structure, and straight-through structure based on the measurement parameters of the open-circuit structure, short-circuit structure, and straight-through structure; and using the parasitic network models to perform de-embedding operations on the first measurement parameters to remove the parasitic effects contained in the first measurement parameters, thereby obtaining the preliminary de-embedding parameters.
[0034] Specifically, using the measured second measurement parameters of open-circuit, short-circuit, and through-circuit structures, an equivalent circuit model or network model is constructed to describe the common parasitic effects of the test fixture and pads. Subsequently, the parasitic network model is applied to the first measurement parameters of the target device, and through corresponding network parameter operations, such as matrix operations or signal flow graph simplification, the contribution of the parasitic network is subtracted from the total measurement response. The output result is the preliminary de-embedding parameter.
[0035] Parasitic network models are abstract models used to mathematically represent or equivalently represent unwanted electrical responses introduced by non-ideal factors such as test fixtures, probes, pads, and interconnect traces, in order to capture the influence of these parasitic elements on the measured signal. Parasitic network models can take various forms. For example, they can be represented by an equivalent circuit containing lumped parameter elements such as resistors, capacitors, and inductors, whose topology and component values are determined by fitting measurement data for open-circuit, short-circuit, and shoot-through structures.
[0036] De-embedding is a mathematical computational process based on network theory, aiming to mathematically eliminate or compensate for the influence of parasitic network models on the original measurement data of the device under test. This operation is typically performed in the frequency domain and involves manipulation of network parameter matrices, such as S-parameter matrices or ABCD matrices. Essentially, the operation treats the measurement results of the target device as the overall response of its intrinsic and parasitic networks connected in some way (e.g., cascaded or parallel), and solves for the parameters of the intrinsic network using a known parasitic network model.
[0037] The preliminary de-embedding parameters are intermediate results obtained after applying the de-embedding algorithm to the first measurement parameter. This result has preliminarily stripped and corrected the obvious parasitic components in the first measurement parameter that can be characterized by standard test structures. However, due to factors such as model simplification or high-frequency effects, residual errors still exist between the preliminary de-embedding parameters and the true intrinsic parameters of the device.
[0038] S300: Input the data set of the first measurement parameter, the second measurement parameter, the operating condition parameter and the preliminary de-embedding parameter into the target neural network model, so as to output the target device parameter through the target neural network model.
[0039] All the information obtained above, namely the first measurement parameters, the second measurement parameters used for modeling, the operating condition parameters, and the initial de-embedding parameters, are integrated into a structured data set, which is then input into the target neural network model as a whole.
[0040] In some embodiments, the target neural network model is a pre-trained neural network model capable of learning complex mapping relationships between measurement data containing noise and errors and ideal device parameters, receiving datasets as input, and outputting predicted parameters optimized by its internal algorithm.
[0041] The neural network model has learned to master the complex mapping relationship from this composite input feature to the parameters of the ideal device. When processing this dataset, the model can simultaneously consider all the information from the original measurements, the preliminary correction results of traditional algorithms, and the specific operating frequency and bias conditions.
[0042] For example, the model can learn that the residual error of the traditional model follows a certain pattern at a specific high frequency point, or that the device characteristics change under different bias voltages. By comprehensively calculating and judging these complex correlations, the target neural network model finally outputs a set of optimized network parameters, i.e., the target device parameters.
[0043] The target device parameters are network parameters that characterize the intrinsic properties of the target device after final correction and optimization by the high-frequency broadband circuit parasitic effect elimination method assisted by this AI. These parameters reflect the true physical characteristics of the device after stripping away all identifiable parasitic effects as accurately as possible, and are usually represented in the form of an S-parameter matrix.
[0044] To illustrate this more specifically, on-chip S-parameter testing was performed on a transistor operating in the tens of gigahertz frequency band. First, a vector network analyzer measured the transistor's S-parameter matrix at multiple frequency points and bias voltages, serving as the first measurement parameter. Simultaneously, the S-parameters of the open-circuit, short-circuit, and shoot-through calibration structures on the same wafer were measured, serving as the second measurement parameter. The frequency and bias voltage values corresponding to each data point were recorded as operating condition parameters. Then, a traditional open-circuit-short-circuit-shoot-through de-embedding algorithm was executed, using the calibration structure parameters to process the transistor measurement data and obtain preliminary de-embedding S-parameters. Next, a pre-trained generalized regression neural network model was invoked, taking the original transistor S-parameters, calibration structure S-parameters, frequency and bias voltage data, and the preliminary de-embedding S-parameters as input. After internal calculations, the neural network model outputs the final corrected intrinsic transistor S-parameters, i.e., the target device parameters. This process achieves a step-by-step approximation and intelligent correction from the original measurements containing parasitics to the characteristics of the pure device.
[0045] In some embodiments, the implementation architecture of the neural network model can be chosen from several options. One implementation is to use a generalized regressive neural network, which is characterized by its single-learning capability, fast training speed, and good robustness to noise in the input data. Another implementation is to use a feedforward neural network, which constructs complex nonlinear mapping relationships through multiple hidden layers and nonlinear activation functions. This type of network requires iterative training using the backpropagation algorithm to adjust its internal weights.
[0046] In this embodiment, the neural network model is a General Regression Neural Network (GRNN) because it has fast training speed, can model nonlinearity, and is robust to noise. Experiments have shown that GRNN is faster to train and has smaller errors compared to RBF, BP, MLP, LSTM, etc.
[0047] In some embodiments, the training process of a neural network model involves obtaining a training sample set, which includes first measurement parameters, second measurement parameters, operating condition parameters, preliminary de-embedding parameters, and real network parameters of multiple sample devices; constructing a basic neural network model; using the training sample set as input features and the real network parameters as training targets to train the basic neural network model to obtain the neural network model.
[0048] The first measurement parameters, second measurement parameters, operating condition parameters, preliminary de-embedding parameters, and real network parameters obtained in the above steps are used as the training sample set. The real network parameters are obtained using simulation software under the target de-embedding conditions. During training, the real network parameters serve as supervisory signals or target values to guide the neural network model in learning the correct input-output mapping relationship.
[0049] During the training iterations, the algorithm sequentially feeds the training sample set into the basic neural network model to obtain the model's predicted output. Then, it calculates the difference between this predicted output and the sample labels, i.e., the actual network parameters. Based on this difference, optimization methods such as backpropagation are used to adjust the adjustable parameters in the basic neural network model, so that the model's output will more closely approximate the actual network parameters when faced with the same or similar inputs. This process is repeated until the model's predictive performance reaches a preset stable standard on the training or validation set. At this point, the basic neural network model has been transformed into a learned and usable neural network model, i.e., a pre-trained target neural network model.
[0050] The target neural network model is pre-trained, and the target device parameters are output through the target neural network model. The output result of the neural network may be discontinuous in frequency. That is, when the frequency response of the output result of the neural network is larger in a local change at a certain frequency than the overall change or the changes at a few nearby points, it is considered to be a jump. After detection, this change is distributed to a larger frequency range to eliminate the jump, so that the overall frequency response is a continuous and smooth curve.
[0051] To address this issue, in some embodiments, jump detection and smoothing are performed on the target device parameters to obtain correction parameters; the correction parameters are then determined to be the target device parameters.
[0052] Jump detection is an algorithmic process used to analyze and identify abnormal abrupt changes or discontinuities in a data sequence. In the context of de-embedding output, jump detection is specifically performed on the curves of target device parameters changing with frequency. This process determines whether there are abnormal abrupt changes that do not conform to physical laws by calculating and comparing the changes between adjacent data points in the sequence, or by assessing the degree of deviation of local changes from the overall trend.
[0053] Smoothing is a data post-processing method used to correct or mitigate identified aberrations in a data sequence, making the overall trend of the sequence more continuous and smoother. Smoothing is not simply data filtering; rather, it involves targeted, evidence-based adjustments to the data values at and around abrupt changes to eliminate the interference caused by aberrations on the overall data quality and interpretability.
[0054] Specifically, the parameter change gradient at each frequency point in the sequence of target device parameters changing with frequency is detected; when the parameter change gradient at a certain frequency point exceeds the target gradient, the frequency point is marked as an abnormal jump point, where the target gradient is a threshold determined by the overall change gradient of the sequence or the change gradient of neighboring frequency points; the parameter values of the abnormal jump points are smoothed and corrected to control the frequency response curve of the target device parameters to be continuously smooth, thereby obtaining the correction parameters.
[0055] First, the parameters of the target device are read. These parameters are complex S-parameter matrices, which are separated into real and imaginary part sequences for separate processing. Jump detection evaluates whether the parameter change gradient at each point in the sequence is within a reasonable range. The parameter difference between each frequency point in the sequence and its next adjacent point is calculated and normalized to the frequency interval to obtain the parameter change gradient at that point.
[0056] Next, a reasonable target gradient is determined as the threshold for anomaly detection. For example, the processing unit can first calculate the average and standard deviation of the absolute values of all gradients across the entire frequency band, and then set the average plus twice the standard deviation as a global threshold. Simultaneously, the processing unit can also calculate a local threshold for each point, which is determined by averaging the absolute values of the gradients of the points before and after it. The processing unit then compares the absolute value of the gradient at the current point with these dynamically determined target gradients.
[0057] When the absolute value of the gradient change of the parameters at a certain frequency point significantly exceeds the preset target gradient, that frequency point is marked as an anomalous transition point. For example, at a certain high frequency point, the real part gradient of the S-parameter suddenly increases sharply, while the gradient values before and after this frequency band are relatively flat. Global statistics also show that such a large gradient is a low-probability event. Therefore, this point is judged as anomalous, and the frequency positions of all marked anomalous transition points are recorded.
[0058] After detection is complete, the smoothing correction phase begins. For each marked anomalous transition point, a correction algorithm is executed. One approach is to use the two nearest unmarked frequency points before and after the transition point as reliable references, employ linear interpolation to calculate the expected parameter values at the transition point frequency, and then replace the parameter values considered anomalous in the original output of the neural network with these estimated values.
[0059] Another approach is to apply a local moving average filter, but the filter weights avoid the marked outliers, calculating only based on surrounding normal points and assigning the results to the outliers. After correcting all anomalous transition points one by one, the abrupt peaks or dips in the sequence are eliminated, and the transitions in parameter values become natural. The corrected parameter sequence is then reassembled, defined as the correction parameters, and finally determined as the target device parameters output by this de-embedding process.
[0060] To quantitatively evaluate the accuracy of the proposed de-embedding algorithm, this invention uses the root mean square error (RMSE) as an evaluation metric to measure the overall deviation between the S-parameters obtained by the de-embedding algorithm and the actual S-parameters of the device. The smaller the metric, the smaller the overall error of the data set. The calculation formula is as follows: ; in, The device S-parameters are obtained through the de-embedding algorithm. The actual S-parameters of the device are given, and N represents the amount of test data.
[0061] like Figure 2As shown, Before de-embedding represents the RMSE of the S-parameters before de-embedding, open short and open short through represent the RMSE of the S-parameters obtained by de-embedding using these two traditional algorithms, and our measure represents the RMSE of the S-parameters obtained by de-embedding in this invention.
[0062] To more intuitively compare the performance of different de-embedding methods, the S-parameters before de-embedding, the S-parameters obtained by the open short algorithm, the S-parameters obtained by the open short thru algorithm, and the S-parameters obtained in this application are plotted as curves of frequency variation on the same coordinate system and compared with the actual S-parameters of the device.
[0063] Figures 3-10 The S-parameters of the real and imaginary parts of s11, s12, s21, and s22 were plotted. Both the graph and the data show that the error of the traditional de-embedding algorithm gradually increases with frequency, indicating its limited ability to model parasitic effects in the high-frequency range. Introducing a neural network significantly improves the model's fitting performance in the high-frequency range, enabling more accurate recovery of the device's intrinsic characteristics and significantly enhancing de-embedding accuracy.
[0064] Traditional de-embedding algorithms have limited accuracy at high frequencies, and their models are more likely to deviate further due to changes in material parameters at low temperatures close to the physical limits of devices. For modeling and on-chip testing of low-temperature or high-frequency devices, the algorithm designed in this application significantly reduces errors compared to traditional algorithms at high frequencies.
[0065] Based on the above-mentioned AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits, some embodiments of this application also provide an AI-assisted device for eliminating parasitic effects in high-frequency broadband circuits, comprising: The acquisition module is used to acquire the first measurement parameter, the second measurement parameter, and the operating condition parameter of the target device, wherein the second measurement parameter is the measurement parameter of the target device in open circuit structure, short circuit structure, and through structure. The processing module is used to perform preliminary de-embedding processing on the first measurement parameters based on the second measurement parameters using a de-embedding algorithm to obtain preliminary de-embedding parameters; The de-embedding module is used to input the data set of the first measurement parameter, the second measurement parameter, the operating condition parameter, and the preliminary de-embedding parameter into the target neural network model, so as to output the target device parameter through the target neural network model.
[0066] This embodiment also provides a computing device, which includes a bus, a processor, a memory, and a communication interface. The processor, memory, and communication interface communicate with each other via the bus. The computing device can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memory in the computing device.
[0067] A bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. A bus can include a path for transmitting information between various components of a computing device (e.g., memory, processor, communication interfaces).
[0068] The processor may include any one or more of the following: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).
[0069] Memory can include volatile memory, such as random access memory (RAM). Memory can also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).
[0070] The memory stores executable program code, which the processor executes to implement the functions of the aforementioned extension module, simulation module, processing module, and de-embedding module, thereby realizing the data migration method. In other words, the memory stores instructions for executing the de-embedding method.
[0071] The communication interface uses transceiver modules such as, but not limited to, network interface cards and transceivers to enable communication between computing devices and other devices or communication networks.
[0072] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.
Claims
1. An AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits, characterized in that, include: Acquire the first measurement parameter, the second measurement parameter, and the operating condition parameter of the target device, wherein the second measurement parameter is the measurement parameter of the target device in open circuit structure, short circuit structure, and through structure; Based on the second measurement parameter, the first measurement parameter is preliminarily de-embedded using a de-embedding algorithm to obtain the preliminarily de-embedded parameter; The data set of the first measurement parameter, the second measurement parameter, the operating condition parameter, and the preliminary de-embedding parameter is input into the target neural network model to output the target device parameter through the target neural network model.
2. The AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits according to claim 1, characterized in that, The target neural network model is a pre-trained neural network model; The method further includes: Obtain a training sample set, which includes first measurement parameters, second measurement parameters, operating condition parameters, preliminary de-embedding parameters, and real network parameters of multiple sample devices; Build a basic neural network model; The training sample set is used as input features, and the real network parameters are used as training targets to train the basic neural network model to obtain the neural network model.
3. The AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits according to claim 2, characterized in that, The actual network parameters were obtained using simulation software under target de-embedding conditions.
4. The AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits according to claim 1, characterized in that, After outputting the target device parameters through the target neural network model, the method further includes: The parameters of the target device are subjected to jump detection and smoothing to obtain correction parameters; The correction parameters are determined to be the target device parameters.
5. The AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits according to claim 4, characterized in that, The step of performing jump detection and smoothing processing on the target device parameters to obtain correction parameters includes: Detect the gradient of parameter change at each frequency point in the sequence of parameter changes of the target device with frequency; When the gradient of parameter change at a certain frequency point exceeds the target gradient, the frequency point is marked as an abnormal jump point. The target gradient is a threshold determined by the overall change gradient of the sequence or the change gradient of neighboring frequency points. The parameter values at the abnormal jump points are smoothed to control the frequency response curve of the target device parameters to be continuously smooth, thereby obtaining the correction parameters.
6. The AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits according to claim 1, characterized in that, The preliminary de-embedding process for the first measurement parameter based on the second measurement parameter using a de-embedding algorithm includes: Based on the measurement parameters of the open-circuit structure, short-circuit structure, and straight-through structure, parasitic network models corresponding to the open-circuit structure, short-circuit structure, and straight-through structure are constructed. The parasitic network model is used to perform de-embedding operations on the first measurement parameters to remove the parasitic effects contained in the first measurement parameters, thereby obtaining the preliminary de-embedding parameters.
7. The AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits according to claim 1, characterized in that, The target neural network model is GRNN.
8. The AI-assisted method for eliminating parasitic effects in high-frequency broadband circuits according to claim 1, characterized in that, The operating condition parameters include the operating frequency and the bias voltage; The step of outputting target device parameters through the target neural network model includes: The operating frequency and the bias voltage are used as input features in the data set, and are input to the target neural network model along with the first measurement parameter, the second measurement parameter, and the preliminary de-embedding parameter. The target neural network model corrects the parameters of the target device according to the operating frequency and the bias voltage, so as to output the target device parameters through the target neural network model.
9. An AI-assisted parasitic effect elimination device for high-frequency broadband circuits, characterized in that, include: The acquisition module is used to acquire the first measurement parameter, the second measurement parameter, and the operating condition parameter of the target device, wherein the second measurement parameter is the measurement parameter of the target device in open circuit structure, short circuit structure, and through structure. The processing module is used to perform preliminary de-embedding processing on the first measurement parameters based on the second measurement parameters using a de-embedding algorithm to obtain preliminary de-embedding parameters; The de-embedding module is used to input the data set of the first measurement parameter, the second measurement parameter, the operating condition parameter, and the preliminary de-embedding parameter into the target neural network model, so as to output the target device parameter through the target neural network model.
10. A computing device, characterized in that, It includes a processor and a memory, the processor being configured to execute instructions stored in the memory to cause the computing device to perform the method as described in any one of claims 1 to 8.