A reverse modeling design method of a passive device and a computer readable medium

Abstract

The application provides a passive device reverse modeling design method and a computer readable medium, including generating samples of different layout sizes of the passive device, establishing a performance data set; establishing and pre-training a forward model with geometric parameters and working frequencies as input parameters and performance parameters as output parameters; constructing a reverse neural network model; dividing the parameter space of the entire performance data set into multiple sub-training domains, and respectively establishing and training a reverse sub-model in each sub-training domain; after the training is completed, integrating all reverse sub-models to form a final reverse neural network model, and the reverse neural network model is used for outputting geometric parameter solutions meeting the conditions according to the specified working frequencies and performance parameters. The method of the application utilizes the forward model to construct a high-confidence data set, and then divides the data set based on the uniqueness of the output parameters and respectively obtains reverse sub-models, so that fast and accurate mapping from target performance parameters to geometric parameters is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of integrated circuit design, and particularly relates to a reverse modeling design method of a passive device and a computer readable medium. BACKGROUND

[0002] Radio frequency passive devices are indispensable key components in radio frequency integrated circuits, mainly including inductors, capacitors, resistors and transmission lines. They undertake important functions such as impedance matching, filtering, resonance, biasing and signal coupling in various radio frequency unit circuits. The optimization of performance indicators and chip area needs to be considered in the circuit design process. The traditional design method usually relies on repeated adjustment of the layout parameters of the passive device, combined with multiple rounds of electromagnetic (EM) simulation, and gradually approaches the design target through trial and error. However, due to the complexity of electromagnetic simulation itself, the time consumption of a single simulation is often several hours, which leads to the prolongation of the overall design cycle and the significant increase of development cost.

[0003] Therefore, how to quickly and accurately obtain the corresponding physical parameters of the device according to the design purpose is a problem to be solved. In order to improve the design efficiency, various modeling methods have been proposed, especially the intelligent modeling method combined with artificial neural network has been widely applied.

[0004] The existing models mainly include the following types. 1) Equivalent circuit modeling: the device is equivalent to a π-type, double π-type or T-type network, and the element values in the equivalent circuit are extracted to obtain the performance of the device, but the parameter extraction process often involves a lot of repeated iteration work and has large high-frequency errors. 2) Proxy model and target optimization: based on EM simulation data, an accurate proxy model is established using artificial neural network, Kriging interpolation, Gaussian process, vector fitting and other methods, and then coupled with PSO, NSGA-II and other target optimization methods, a reasonable compromise between key performance parameters and area is made according to the given target, and automatic device synthesis is realized. However, for different electrical targets, this optimization process needs to be repeated, which is very time-consuming and prone to local optimal values. 3) Reverse modeling method: reverse modeling defines performance parameters as input and physical parameters as output, and directly establishes a nonlinear mapping relationship from performance parameters to physical parameters. It can directly provide the required geometric size according to the design target without repeated iteration. However, due to the non-uniqueness of the input-output relationship, the difficulty and accuracy of reverse modeling are challenging. SUMMARY

[0005] The purpose of the present application is to provide a reverse modeling design method of a passive device and a computer readable medium to realize fast and accurate mapping from target performance parameters to geometric parameters.

[0006] In order to achieve the above purpose, the present application provides a reverse modeling design method of a passive device, comprising:

[0007] S1: generating samples of different layout sizes of the passive device within the process tolerance range through actual testing combined with simulation, and constructing a performance dataset, the performance dataset including geometric parameters, frequency and performance parameters;

[0008] S2: establishing a forward neural network model with geometric parameters and working frequency as input parameters, the forward neural network model being used for outputting performance parameters; pre-training the forward neural network model by using the performance dataset to obtain a passive device forward model; the passive device forward model being used for result verification of an inverse neural network model;

[0009] S3: constructing an inverse neural network model with working frequency and performance parameters at the working frequency as input parameters and geometric parameters of the layout as output parameters; dividing the parameter space of the entire performance dataset into multiple sub-training domains, and respectively establishing and training an inverse sub-model in each sub-training domain; after the training, integrating all the inverse sub-models to form a final inverse neural network model, the inverse neural network model being used for outputting geometric parameter solutions meeting the conditions according to specified working frequency and performance parameters.

[0010] The passive device includes resistors, capacitors and inductors of different layout sizes, the layout sizes including shape types and geometric sizes of the layout; when the passive device is an inductor, the performance parameters include inductance and quality factor; when the passive device is a resistor, the performance parameter is resistance value; when the passive device is a capacitor, the performance parameters of the capacitor include capacitance value and quality factor of the capacitor.

[0011] The forward neural network model is a forward ANN model, and the inverse sub-model is an inverse ANN sub-model.

[0012] The step S3 adopts a dynamic binning training method, specifically including:

[0013] S31: taking an initial bin as a current bin; and dividing the performance dataset into the current bin as training data; wherein, based on the uniqueness of the output parameters, the performance dataset is divided into different current bins according to at least one of the geometric parameters of the layout;

[0014] S32: independently preprocessing the data of each sub-training domain;

[0015] S33: for each sub-training domain, establishing an inverse sub-model of the same structure with working frequency and performance parameters at the working frequency as input parameters and geometric parameters of the layout as output parameters, and respectively training the inverse sub-model;

[0016] S34: evaluate the accuracy of each inverse sub-model using the forward model of the passive device, and determine whether each current bin meets the iteration termination condition; if not, further subdivide the training data of the corresponding current bin into new current bins and train the next level of inverse sub-models, otherwise terminate;

[0017] S35: load all trained inverse sub-models, integrate them into a complete passive device inverse model, which is used to output the geometric parameters of the layout according to the performance parameters at the working frequency.

[0018] Based on the uniqueness of the output parameters, the performance data set is divided into different current bins according to at least one of the geometric parameters of the layout, which means that the number of one-to-many and many-to-many parameter mapping relationships in each current bin is minimized.

[0019] In steps S31 and S35, when the training data is divided into current bins, each group of adjacent current bins retains partially overlapping training data.

[0020] The passive device reverse modeling design method further comprises step S4: inputting the geometric parameter solution output by the inverse neural network model into the forward neural network model to predict the performance parameters, calculating the deviation from the original design requirements, and outputting the geometric parameter solution of the device corresponding to the deviation less than the preset threshold as the candidate geometric parameter solution.

[0021] Step S4 further comprises: further constraining the candidate geometric parameter solution according to the device area and the resonant frequency, finally outputting the geometric parameters of the layout that meets the electromagnetic performance and automatically generating the physical layout; and / or directly outputting the S parameter file that can be directly used for circuit simulation according to the geometric parameter solution in the forward model, to realize automatic synthesis design of the device.

[0022] Step S1 comprises:

[0023] S11: prepare a wafer of passive devices based on the same device process and containing multiple different layout sizes, perform device flow test on it, and analyze the performance parameters of the passive devices based on the test results, the analysis results including the actual process fluctuation range and the actual average value of the performance parameters;

[0024] S12: determine the actual fluctuation range of the process parameters based on Monte Carlo simulation of electromagnetic simulation using the analysis results; the process parameters include the thickness of each layer of material, the relative dielectric constant and the metal conductivity;

[0025] S13: take the average value of the actual fluctuation range of the process parameters as the process parameters required for electromagnetic simulation, establish a parameter space according to the geometric parameters of the layout of the passive device; obtain the performance parameters by parameter space sampling and formula operation of electromagnetic simulation to establish a performance data set;

[0026] Alternatively, step S1 may include:

[0027] S11': Wafers providing passive devices with multiple different layout sizes under the same device process;

[0028] S12': Establish an equivalent circuit model for the passive device and obtain the parameters of each element in the equivalent circuit model; take the geometric parameters of the layout of each passive device and the parameters of each element in the corresponding equivalent circuit model as a set of data to form a component dataset;

[0029] S13': The neural network model is trained using the component dataset to obtain the network model of the equivalent circuit components. The network model of the equivalent circuit components is used to output the parameters of each component in the equivalent circuit model based on the geometric parameters of the layout of the passive device.

[0030] S14': Establish a parameter space based on the geometric parameters of the passive device layout, obtain the geometric parameters of the layout and the parameters of each component in the corresponding equivalent circuit model through parameter space sampling, and obtain the performance parameters of the corresponding passive devices through electromagnetic simulation to establish a performance dataset.

[0031] On the other hand, the present invention provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0032] This invention provides a reverse modeling design method for passive devices. It utilizes a forward model of the passive device to construct a high-confidence second performance dataset. Based on the uniqueness of the output parameters, the second performance dataset is divided into multiple sub-training domains, and reverse sub-models are trained in each sub-model. This achieves a unique mapping from performance parameters to geometric parameters in each reverse sub-model, enabling a fast and accurate mapping from target performance parameters to geometric parameters. Furthermore, this invention inputs the geometric parameter solution of the layout output by the reverse sub-model into a forward neural network model for verification, achieving a fast and accurate mapping from target performance parameters to layout geometric parameters. Attached Figure Description

[0033] Figure 1 This is a flowchart of a reverse modeling design method for a passive device according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the geometric parameters of spiral inductors of different shapes, showing the line width W, spacing S, inner diameter D, and number of turns T;

[0035] Figure 3 This is a flowchart for generating the performance dataset;

[0036] Figures 4A-4CThis is a graph showing the distribution of the quality factor Q of planar spiral inductors of the same layout size from three different batches of wafers at a frequency of 3.6 GHz. Figure 4A It is a wafer distribution diagram. Figure 4B It is a box diagram. Figure 4C This is a graph of the CDF cumulative distribution function;

[0037] Figure 5 It is a circuit diagram of the π-type equivalent circuit model;

[0038] Figures 6A-6D This is a comparison chart of the positive ANN model prediction results and EM simulation results for inductors with different combinations of geometric parameters;

[0039] Figure 7 This is a flowchart of the dynamic binning training method used in the inverse neural network model. Detailed Implementation

[0040] The invention will be further described below with reference to specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0041] First embodiment: A reverse modeling design method for passive devices

[0042] like Figure 1 As shown, taking the device fabrication process using 0.25μm GaAs RF technology as an example, the reverse modeling design method for passive devices of the present invention specifically includes the following steps:

[0043] Step S1: By combining actual testing and simulation, samples of passive devices with different layout sizes are generated within the process tolerance range to construct a high-confidence performance dataset, which includes geometric parameters, frequency, and performance parameters.

[0044] Simulation refers to automatically generating the layout of passive devices based on geometric parameters and parameterized units, and performing batch electromagnetic simulations to generate samples of passive devices.

[0045] In this embodiment, as Figure 3 As shown, step S1 includes:

[0046] Step S11: Prepare a wafer containing passive devices with multiple different layout sizes based on the same device process, perform device fabrication testing on it, and analyze the performance parameters of the passive devices based on the test results. The analysis results include the actual process fluctuation range and the actual average value of the performance parameters.

[0047] The passive device includes resistors, capacitors and inductors with different layout sizes, including shape types and geometric sizes of the layout. That is, in the case of the same shape type and geometric size of the layout, it is considered as the same layout size.

[0048] As shown in Figure 2 In this embodiment, the passive device is an inductor, and the shape type of the layout of the inductor includes square, circular and octagonal of planar spiral; the geometric parameters include line width W, pitch S, inner diameter D and number of turns T, and the process constraint range is R≥20μm, W≥5μm, S≥4μm and T≥1.25. The design of each process layer of the planar spiral inductor conforms to the Design Rule Check (DRC).

[0049] A wafer containing passive devices with different layout sizes based on the same device process is prepared, specifically including: based on a specified device process (such as 0.25μm GaAs radio frequency process), integrating passive devices with different layout sizes in the layout of the wafer, and obtaining sufficient process fluctuation data samples by multiple batches of flow of the device process.

[0050] Wherein, the wafer is tested, and the performance parameters of the passive device are analyzed based on the test results, at least including: obtaining the S parameters (i.e. scattering parameters) of the passive device, extracting the performance parameters of the passive device from the S parameters, and statistically obtaining the analysis results of the performance parameters. For a two-port network, the S parameters of the passive device include four S parameters , , , The analysis results of the performance parameters include the actual process fluctuation range and the actual average value of the performance parameters.

[0051] In order to verify the accuracy of subsequent electromagnetic simulation and the fluctuation of device process, in this embodiment, wafer-level testing is performed on the wafer of the passive device obtained by flow to obtain the S parameters of the passive device.

[0052] In this embodiment, wafer-level testing is performed to obtain the S parameters of the passive device, specifically including:

[0053] Step S111: confirming the wafer information of the passive device to be tested, i.e. product number (Product ID), batch number (Lot ID) and wafer number (Wafer ID); it should be noted that the flow layout on the wafers from different batches is the same layout.

[0054] Step S112: visually inspecting the wafer of the passive device;

[0055] Step S113: match the needle card with the probe, and load the wafer of the to-be-tested passive device into the probe station after visual inspection;

[0056] Step S114: perform whole-wafer testing on the wafer of the passive device to obtain S parameters of the whole-wafer device.

[0057] The whole-wafer testing is an automatic side view. A test machine is set to automatically test a program, and a small number of wafers of the passive device are tested. After confirming that the wafers of the passive device are correct, whole-wafer testing is performed on all wafers of the passive device. The test frequency band is 0.1-40 GHz. Finally, S parameters of the whole-wafer device are obtained.

[0058] In the embodiment, the passive device is an inductor, and the performance parameters of the inductor include an inductance L, a quality factor Q, and a self-resonant frequency SRF. The inductance L and the quality factor Q are obtained by extracting S parameters, and the calculation formulas are as follows. The results of the inductance L and the quality factor Q are curves varying with a frequency f.

[0059] The calculation formulas of the inductance L and the quality factor Q are as follows.

[0060] ,

[0061] wherein, Y represents an admittance parameter, which can be obtained by converting S parameters; Y represents an imaginary part, Y represents a real part, and f is a frequency.

[0062] The self-resonant frequency SRF is a frequency point corresponding to the quality factor Q = 0.

[0063] In other embodiments, when the passive device is a resistor, the performance parameter of the resistor is a resistance value. When the passive device is a capacitor, the performance parameters of the capacitor include a capacitance value, a quality factor of the capacitor, and a resonant frequency of the capacitor.

[0064] The analysis results of the performance parameters are obtained by statistics, specifically including: generating a wafer distribution map (WaferMap), a CDF cumulative distribution function map (CDFplot), and a box plot (Boxplot) to realize data visualization, analyze the actual process fluctuation range of each performance parameter, and calculate the actual average value and actual standard deviation of each performance parameter.

[0065] As shown in FIG. 6, Figures 4A-4C is a distribution diagram of the quality factor Q of a planar spiral inductor with the same layout size on wafers from three batches at a frequency point of 3.6 GHz, Figure 4A is a wafer distribution map, Figure 4B is a box plot, Figure 4C is a CDF cumulative distribution function map.

[0066] Step S12: determining the actual fluctuation range of the process parameters by using the analysis result of the performance parameters and the Monte Carlo simulation based on the electromagnetic simulation; the process parameters include the thickness of each layer of material, the relative dielectric constant and the metal conductivity.

[0067] Therefore, the application verifies the process fluctuation range of the electromagnetic process parameters by the Monte Carlo simulation, and corrects the process parameters in combination with the actual test data, which helps to improve the stability of the process parameters of the device process.

[0068] The step S12 specifically includes:

[0069] Step S121: writing a parameterized unit of the passive device, which automatically generates a passive device model of a corresponding layout size by driving a defined geometric parameter, for subsequent acquisition of simulation performance parameters.

[0070] In the embodiment, the parameterized unit of the passive device is written by using the AEL language. As described above, the shape of the inductor includes the square, circular and octagonal shapes of the planar spiral; and the geometric parameters of the layout of the inductor include the line width W, the pitch S, the inner diameter D and the number of turns T.

[0071] Wherein, the AEL is a built-in script language of the ADS (Advanced Design System) of the Keysight Company, which is mainly used for automatic layout size design, parameterized device modeling and batch data processing. The size parameterization adjustment can be realized by the AEL, and the design efficiency is improved.

[0072] Step S122: performing fluctuation sampling within the range of ±20% of the standard value of the process parameters, acquiring simulation performance parameters by multiple Monte Carlo simulations, adjusting the process parameters by comparing the simulation performance parameters with the actual process fluctuation range and the actual average value, and determining the current process parameters as the final result when the distribution and average value of the simulation performance parameters most conform to the actual process fluctuation range and the actual average value.

[0073] Therefore, the accuracy of the simulation result is verified by comparing the simulation performance parameters with the performance parameters obtained by the wafer-level test, and the optimal set of process parameters is determined.

[0074] Wherein, the simulation performance parameters are a set of parameters with process fluctuation range; when the process fluctuation range of the simulation performance parameters all falls within the actual process fluctuation range of the performance parameters and the average value of the simulation performance parameters is closest to the actual average value of the performance parameters, it is determined that the combined simulation process parameters are closest to the actual process.

[0075] During the Monte Carlo simulation process, the process parameters are sampled within the range of ±20% of their standard values, obtaining multiple sets of sampling values. Combined with these sampling values and the parameterized unit of the passive device, the corresponding layout size of the passive device model is automatically generated, and then electromagnetic simulation is performed to obtain S parameters, from which the simulation performance parameters are extracted, including inductance value (L), quality factor (Q) and self-resonant frequency (SRF).

[0076] In this embodiment, electromagnetic simulation is performed by ADS Momentμm. ADS Momentμm is a three-dimensional plane electromagnetic field simulation module in the Advanced Design System (ADS) software of Keysight Company, mainly used for electromagnetic characteristic analysis of high-frequency circuits and antennas.

[0077] It should be noted that for different layout sizes (i.e. different line width W, spacing S, inner diameter D, and number of turns T) of the passive device structure, the process parameters are usually consistent. As long as the same device process is used, the process parameters of the substrate for simulation setting are consistent. Therefore, the process parameters obtained in step S22 can be directly used for electromagnetic simulation.

[0078] Step S13: The average value of the actual fluctuation range of the process parameters is taken as the process parameters required for electromagnetic simulation, and a parameter space is established according to the geometric parameters of the layout of the passive device; the Latin hypercube sampling method is used to obtain performance parameters through parameter space sampling and formula operation of electromagnetic simulation to establish a performance data set.

[0079] Since the passive device occupies a large area and has many geometric parameters, it would consume a large amount of layout area if actual test results are used to establish a performance data set. Therefore, we prepare a small number of passive devices to obtain wafer-level test results, and compare them with electromagnetic simulation results to correct the actual fluctuation range of the process parameters and verify the accuracy of the electromagnetic simulation. Subsequently, according to the process parameters required for electromagnetic simulation, a large number of accurate performance parameters are generated to establish a performance data set.

[0080] In this embodiment, the Latin hypercube sampling method is used to perform parameter space sampling and formula operation of electromagnetic simulation within the parameter space of the geometric parameters (i.e. line width W, spacing S, inner diameter D, and number of turns T) of the layout to obtain a large number of electromagnetic simulation results as the performance data set.

[0081] In this embodiment, 1296 sets of data are sampled within the design space of the geometric parameters of the layout. In other embodiments, 1000-8000 sets of data are sampled within the design space of the geometric parameters of the layout.

[0082] In the embodiment, the parameter space of the geometric parameters (line width W, spacing S, inner diameter D, number of turns T) of the layout satisfies the process constraint range and the area cannot be too large. The process constraint range satisfied by the parameter space of the geometric parameters of the layout is: W=5~30μm; S=4~25μm; R=20~150μm; T=1.25~5.25, and the number of turns is set with a fixed step of 0.25.

[0083] In the embodiment, the passive device is an inductor, and the performance data set includes the geometric parameters (line width W, spacing S, inner diameter D, number of turns T) of the layout, the working frequency Freq, and the performance parameters (S parameters, inductance value L, and quality factor Q) corresponding to each working frequency point.

[0084] In another embodiment, the step S1 specifically includes:

[0085] Step S11': providing a wafer of passive devices of multiple different layout sizes under the same device process;

[0086] Step S12': establishing an equivalent circuit model for the passive device, obtaining the parameters of each element in the equivalent circuit model, taking the geometric parameters of the layout of each passive device and the parameters of each element in the corresponding equivalent circuit model as a group of data, and forming an element data set;

[0087] In the embodiment, the passive device is an inductor, and the equivalent circuit model is a π-type equivalent circuit model. Figure 5 The π-type equivalent circuit model is shown, which includes the following elements: a π-type network inductor Lsp, a π-type network series parasitic resistor Rsp, an RL network device (i.e., a resistor Rsh and an inductor Lsh), a substrate parasitic capacitor Csub1 and a substrate parasitic capacitor Csub2, a substrate parasitic resistor Rsub1 and a substrate parasitic resistor Rsub2, a microstrip connection line parasitic inductor Lt1 and a microstrip connection line parasitic inductor Lt2, and a coil overlap and edge parasitic capacitor Cp.

[0088] Step S13': training the neural network model using the element data set to obtain a network model of the equivalent circuit elements, and the network model of the equivalent circuit elements is used to output the parameters of each element in the equivalent circuit model according to the geometric parameters of the layout of the passive device.

[0089] The neural network model is an artificial neural network model, the number of which is the same as and corresponds to the number of elements in the π-type equivalent circuit model, that is, 11 neural network models are constructed in the embodiment, the input of the 11 neural network models is the neural network model of the inductor, and the output is the parameters of the 11 elements in the π-type equivalent circuit model, respectively. The number of hidden layers, the number of neurons and the number of iterations in each neural network model are adjusted according to the parameters of the elements in the corresponding π-type equivalent circuit model. When training, the geometric parameters in the element data set are used as the input of the neural network model, and the parameters of each element in the π-type equivalent circuit model in the element data set are used as the label for training, and finally the relationship between the geometric parameters of the inductor layout and the parameters of the elements in the π-type equivalent circuit model is obtained.

[0090] Step S14': establishing a parameter space according to the geometric parameters of the layout of the passive device, obtaining the geometric parameters of the layout and the parameters of each element in the corresponding equivalent circuit model by sampling the parameter space, and obtaining the corresponding performance parameters of the passive device by electromagnetic simulation to establish a performance data set.

[0091] In the embodiment, 1296 groups of data are sampled in the design space of the geometric parameters of the layout. In other embodiments, the data sampled in the design space of the geometric parameters of the layout is 1000-8000 groups.

[0092] In the embodiment, the parameter space of the geometric parameters (line width W, spacing S, inner diameter D, number of turns T) of the layout satisfies the process constraint range and the area cannot be too large. The process constraint range satisfied by the parameter space of the geometric parameters of the layout is: W=5-30μm; S=4-25μm; R=20-150μm; T=1.25-5.25, the number of turns is set with a fixed step of 0.25.

[0093] In the embodiment, the passive device is an inductor, and the performance data set includes the geometric parameters (line width W, spacing S, inner diameter D, number of turns T) of the layout, the working frequency Freq and the corresponding performance parameters (inductance L, quality factor Q). That is, the performance parameters in the performance data set include the inductance L and the quality factor Q under each working frequency Freq.

[0094] Step S2: establishing a forward ANN model with the geometric parameters (including line width W, spacing S, inner diameter D, number of turns T) and working frequency Freq as input, the forward ANN model being used for outputting performance parameters; and pre-training the forward ANN model by using the performance data set to obtain a passive device forward model. The passive device forward model is used for verifying the result of the reverse neural network model and providing S parameters for circuit simulation.

[0095] In the embodiment, the established forward ANN model includes a first and a second forward ANN model: the first forward ANN model is used to output the inductance value L and the quality factor Q at a specified operating frequency Freq, and the second forward ANN model is used to output the S parameter at the specified operating frequency Freq.

[0096] Generally, the inductance value L and the quality factor Q can be calculated from the S parameter, but this way will introduce error accumulation, resulting in a decrease in accuracy, so two independent ANNs are used in the embodiment to improve the prediction accuracy of the key performance parameters. In the embodiment, both of the two forward ANN models adopt a multi-layer perceptron (MLP) structure, the input is 5 neurons (W, S, D, T, Freq), the hidden layer is 3 layers, and the activation function is ReLU. The first forward ANN model outputs 2 neurons (the inductance value L and the quality factor Q), and the second forward ANN model outputs the S parameter.

[0097] The training settings include: Adam optimizer, learning rate 0.0001, 200 iterations, loss function is mean square error (MSE), and evaluation indicators include MSE and mean absolute error (MAE). Figures 6A-6D The curves of the inductance L and the quality factor Q predicted by the first forward ANN model are compared with the electromagnetic simulation results in the 20 GHz frequency band under different combinations of geometric parameters. In the figure, Q ANN and L ANN represent the model prediction values, and Q EM and L EM represent the electromagnetic simulation values.

[0098] Step S3: constructing a reverse neural network model, the reverse neural network model taking the operating frequency Freq and the performance parameters (inductance L, quality factor Q) at the operating frequency Freq as inputs, and taking the corresponding geometric parameters (line width W, spacing S, inner diameter D, number of turns T) of the layout as outputs; adopting a dynamic binning training method, the parameter space of the entire performance data set is divided into multiple sub-training domains, and a reverse ANN sub-model is established and trained in each sub-training domain to solve the inherent non-unique problem in the mapping of performance parameters to geometric parameters; after training, all reverse ANN sub-models are integrated to form a final reverse neural network model, which is used to output the geometric parameter solution meeting the conditions according to the specified operating frequency and performance parameters.

[0099] The working process of the dynamic binning training method adopted by the reverse neural network model is as follows Figure 7The dynamic binning training method is an iterative process, which specifically comprises: first, taking an initial bin as a current bin; dividing the performance dataset into the current bin as training data and training the inverse ANN sub-model using the data of each current bin; evaluating the accuracy of each inverse ANN sub-model using the passive device forward model, and determining whether the iteration termination condition is met; if not, further subdividing the data of the corresponding current bin into new current bins and training the next level of inverse ANN sub-models, otherwise terminating. The iteration termination condition is that the loss of the inverse ANN sub-model decreases by less than a threshold (i.e., the accuracy of the inverse ANN sub-model reaches a target), or the test error rebounds. The dynamic binning training method avoids excessive division leading to high model complexity by dynamically adjusting the binning granularity, and ensures that the binning result matches the data distribution characteristics.

[0100] Since the same performance parameters (L, Q, Freq) can correspond to multiple sets of different geometric parameters, directly constructing a single inverse model can lead to large errors due to non-unique mapping, and multiple geometric parameter solutions for the same set of performance inputs, making it difficult to meet the precise design requirements. Therefore, the present application can efficiently and accurately obtain physical parameter configurations that meet the design target by combining the forward model with the binning multi-channel inverse sub-model.

[0101] The step S3 adopts a dynamic binning training method, specifically comprising:

[0102] Step S31: taking an initial bin as a current bin; dividing the performance dataset into the current bin as training data.

[0103] In this embodiment, based on the uniqueness of the output parameters, the performance dataset is divided into different current bins according to at least one of the geometric parameters of the layout, which means that the number of many-to-one and many-to-many parameter mapping relationships in each current bin is minimized, i.e., the parameter mapping relationship in each current bin is ensured to be unique as much as possible. For example, in this embodiment, based on the physical value range and data distribution characteristics of the line width W and the pitch S, the performance dataset is divided into different current bins as the intervals of multiple sub-models. In the initial stage, the parameter space is divided into 18 first-level sub-model intervals, and each interval corresponds to an independent inverse sub-model. To reduce the difficulty of nonlinear mapping and avoid many-to-one mapping problems, the size combination distribution of W and S and the value change of the inner diameter D and the number of turns T are fully considered during the division to ensure that the parameter mapping relationship in each current bin is as unique as possible.

[0104] Step S32: independently pre-processing the data of each sub-training domain;

[0105] The step S32 specifically comprises: independently normalizing (mean value of 0, variance of 1) the data of each sub-training domain; and dividing the training set and the verification set according to 8:2 to ensure consistency of distribution.

[0106] The step S33: for each sub-training domain, taking the working frequency Freq and the performance parameters (inductance L, quality factor Q) under the working frequency Freq as input parameters, and taking the geometric parameters (line width W, spacing S, inner diameter D, number of turns T) of the layout as output parameters, an inverse ANN sub-model of the same structure is established and trained respectively.

[0107] Each inverse sub-model maintains independent initialization weight to avoid cross-domain interference; the activation function, the loss function and the optimizer are set. In the present example, the activation function is set to ReLU, the loss function is set to MSE Loss, and the optimizer is set to Adam for training, the learning rate is 0.0001, and a total of 1500 iterations are performed.

[0108] The step S34: the accuracy of each inverse ANN sub-model is evaluated by using the forward model of the passive device, and it is determined whether each current box meets the iteration termination condition; if not, the training data of the corresponding current box is further subdivided into a new current box, and the next level of inverse ANN sub-model is trained, otherwise the training is terminated.

[0109] The iteration termination condition is that the loss of the inverse ANN sub-model decreases by less than a threshold value, or the test error rebounds. That is, if the loss (Loss) of a certain inverse ANN sub-model is significantly higher than the average level, a dynamic binning refinement strategy is started, as shown in FIG. 4, which is further divided into 3-4 secondary sub-model intervals (such as subdivided according to W every 2 μm, S every 3 μm) according to the gradient direction of the line width W and the spacing S, so as to reduce the influence of data sparsity and improve the fitting accuracy. Figure 7

[0110] In the steps S31 and S35, in order to ensure the continuity between the bins, when the training data is divided into the current bins, part of the training data (such as about 5 groups of overlapping data) of each group of adjacent current bins is retained to eliminate boundary mutation, enhance the model generalization ability and reduce the boundary prediction deviation.

[0111] The step S35: all the trained inverse ANN sub-models are loaded and integrated into a complete passive device inverse model, and the passive device inverse model is used to output the geometric parameters of the layout according to the performance parameters under the working frequency Freq.

[0112] ​That is, the working frequency Freq, the inductance L, and the quality factor Q, which are the optimization targets, are input into the complete passive device inverse model, each inverse sub-model is called to output the corresponding geometric parameters (line width W, spacing S, inner diameter D, and number of turns T) as the predicted geometric parameter solution.

[0113] Step S4: input the geometric parameter solution output by the inverse neural network model into the forward ANN model to predict the performance parameters, calculate the deviation from the original design requirements, and output the geometric parameter solution corresponding to the deviation that is less than the preset threshold as the candidate geometric parameter solution.

[0114] The step S4 can further include: further constraining the candidate geometric parameter solution according to the device area and the resonant frequency, and finally outputting the geometric parameters of the layout that meets the electromagnetic performance and automatically generating the physical layout; and / or

[0115] In the forward model, the S parameter file that can be directly used for circuit simulation is directly output according to the geometric parameter solution, and finally the device automatic synthesis design is realized.

[0116] In the step S4, in order to verify the inverse model result, the geometric parameters (line width W, spacing S, inner diameter D, and number of turns T) in the prediction result and the required working frequency Freq are input into the pre-trained forward model to obtain the performance parameters corresponding to each group of predicted geometric parameter solution.

[0117] The deviation from the original design requirement is calculated, and the geometric parameter solution corresponding to the deviation that is less than the preset threshold is output, specifically including: the performance parameters obtained by verification are respectively compared with the target performance parameters (i.e. the initial input Freq, L and Q) of the design requirement, if the error is greater than 5%, it is considered that the physical parameter prediction is inaccurate, and is removed, and the remaining geometric parameter solution is reserved.

[0118] The candidate geometric parameter solution is further constrained according to the device area and the resonant frequency, specifically including: the inductance area of the candidate geometric parameter solution is calculated, and finally two groups of geometric parameter solutions with the smallest area and the highest accuracy are reserved for the designer to weigh and select to quickly perform layout design.

[0119] For example, the size parameters of the inductance L of 3.6nH and the quality factor Q of 21 at the frequency of 3.6GHz are predicted, the error of the geometric parameter solution predicted by the inverse sub-model and the performance parameters corresponding to the geometric parameter solution is shown in Table 1 and Table 2.

[0120] Table 1: predicted geometric parameter solution

[0121] Predicted line width ind_w(1) Predicted spacing ind_s(1) Predicted inner diameter ind_r(1) Predicted turns ind_t(1) Frequency (GHz) Inductance L (nH) Quality factor Q 12.0 15.0 70.0 4.25 3.6 3.6 21.0 16.0 10.0 88.0 4.25 3.6 3.6 21.0 23.0 14.0 52.0 4.5 3.6 3.6 21.0 24.0 15.0 46.0 4.5 3.6 3.6 21.0 11.0 13.0 93.0 4.25 3.6 3.6 21.0 12.0 15.0 85.0 4.25 3.6 3.6 21.0

[0122] Table 2: performance parameters obtained by the forward ANN model and the error thereof

[0123] Predicted inductance predicted_L Absolute error of inductance abs_error_L Relative error of inductance rel_error_L Predicted quality factor predicted_Q Quality factor absolute error abs_error_Q Quality factor relative error rel_error_Q 3.5156 0.0844 2.3454 21.1086 0.1086 0.5171 3.423 0.177 4.9158 21.2378 0.2378 1.1324 3.6463 0.0463 1.2872 20.5935 0.4065 1.9357 3.7661 0.1661 4.6151 20.1465 0.8535 4.0641 3.5156 0.0844 2.3454 21.1086 0.1086 0.5171 3.6377 0.0377 1.0461 21.3749 0.3749 1.7851

[0124] In addition, the application further provides a computer readable medium, which stores a computer program, and the computer program is executed by a processor to realize the reverse modeling design method of the passive device.

[0125] The reverse modeling design method of the passive device provides a new neural network framework, integrates a forward neural network model and a multi-channel reverse sub-model, realizes fast and accurate mapping from target performance parameters to layout geometric parameters, and further provides a high-confidence data set construction method, which is based on actual test data, compares parameter consistency and distribution difference of different batches of devices in multiple dimensions, and provides a basis for process stability analysis.

[0126] The reverse modeling design method of the passive device has the following advantages:

[0127] 1. Significantly shorten the design cycle and reduce the cost: by directly mapping the performance parameters to the physical parameters through the reverse model, the traditional trial-and-error method is avoided, and the time-consuming electromagnetic simulation is avoided, and the design efficiency is significantly improved.

[0128] 2. High precision and reliability: based on the high-confidence performance data set and the sub-domain division strategy of the parameter space sampling, the many-to-one mapping problem is effectively solved, the model prediction error is less than or equal to 5%, and the designer is assisted to quickly weigh between performance and area.

[0129] 3. Improve the design robustness and adaptability: the process fluctuation range of electromagnetic simulation parameters is verified through Monte Carlo simulation, and the model is corrected combined with actual test data, so as to ensure the reliability of the optimization result under process fluctuation, and to support the flexible design demand of various inductance shapes (square, circle, octagon).

[0130] The above is only a preferred embodiment of the application, and is not used to limit the scope of the application, and the above embodiment of the application can be variously changed. That is, any simple, equivalent change and modification made according to the content of the claims and the description of the application fall within the protection scope of the patent claims of the application. The application is not described in detail, which is a conventional technical content.

Claims

1. A method for reverse modeling design of a passive device, the method comprising: The method comprises the following steps: Step S1: generating samples of different layout sizes of the passive device within a process tolerance range by combining actual testing with simulation, and constructing a performance dataset, wherein the performance dataset comprises geometric parameters, frequency and performance parameters; Step S2: establishing a forward neural network model with the geometric parameters and the working frequency as input parameters, and the forward neural network model is used for outputting the performance parameters; pre-training the forward neural network model by using the performance dataset to obtain a passive device forward model; and the passive device forward model is used for verifying the result of a reverse neural network model; Step S3: constructing the reverse neural network model, wherein the working frequency and the performance parameters at the working frequency are used as input parameters, and the geometric parameters of the layout are used as output parameters; dividing the parameter space of the entire performance dataset into a plurality of sub-training domains, and establishing and training a reverse sub-model in each sub-training domain; After the training is completed, all the reverse sub-models are integrated to form a final reverse neural network model, and the reverse neural network model is used for outputting the geometric parameter solution meeting the condition according to the specified working frequency and performance parameters; The step S3 adopts a dynamic binning training method, and specifically comprises the following steps: Step S31: taking an initial bin as a current bin; and dividing the performance dataset into the current bin as training data; wherein, based on the uniqueness of the output parameters, the performance dataset is divided into different current bins according to at least one of the geometric parameters of the layout; Step S32: independently preprocessing the data of each sub-training domain; Step S33: for each sub-training domain, a reverse sub-model of the same structure is established with the working frequency and the performance parameters at the working frequency as input parameters, and the geometric parameters of the layout as output parameters, and the reverse sub-model is trained respectively; Step S34: the accuracy of each reverse sub-model is evaluated by using the passive device forward model, and it is determined whether each current bin meets the iteration termination condition; if not, the training data of the corresponding current bin is further subdivided into a new current bin, and a next-level reverse sub-model is trained, otherwise the training is terminated; Step S35: loading all the trained reverse sub-models, and integrating the reverse sub-models into a complete passive device reverse model, wherein the passive device reverse model is used for outputting the geometric parameters of the layout according to the performance parameters at the working frequency.

2. The method of claim 1, wherein, The passive device comprises resistors, capacitors and inductors of different layout sizes, and the layout sizes comprise shape types and geometric sizes of the layout; When the passive device is an inductor, the performance parameters comprise inductance and quality factor; when the passive device is a resistor, the performance parameter is resistance; and when the passive device is a capacitor, the performance parameters of the capacitor comprise capacitance and quality factor of the capacitor.

3. The method of claim 1, wherein, The forward neural network model is a forward ANN model, and the reverse sub-model is a reverse ANN sub-model.

4. The method of claim 1, wherein, Based on the uniqueness of the output parameters, the performance dataset is divided into different current bins according to at least one of the geometric parameters of the layout, which means that the number of one-to-many and many-to-many parameter mapping relationships in each current bin is minimized.

5. The method of claim 1, wherein, In the step S31 and the step S35, when the training data is divided into the current bins, each group of adjacent current bins retains partially overlapped training data.

6. The method of claim 1, wherein, The step S4 further comprises: further constraining the candidate geometric parameter solution according to the device area and the resonant frequency, finally outputting the geometric parameter of the layout meeting the electromagnetic performance and automatically generating the physical layout; and / or 7. The method of claim 6, wherein, In the forward model, the S parameter file directly used for circuit simulation is directly outputted according to the geometric parameter solution, so as to realize the automatic comprehensive design of the device. The step S1 comprises:

8. The method of claim 1, wherein, The step S11: preparing a wafer of passive devices based on the same device process and containing multiple different layout sizes, testing the passive devices, and analyzing the performance parameters of the passive devices based on the test results; the analysis results include the actual process fluctuation range and the actual average value of the performance parameters; The step S12: determining the actual fluctuation range of the process parameters based on the Monte Carlo simulation of electromagnetic simulation by using the analysis results; the process parameters include the thickness of each layer of material, the relative dielectric constant and the metal conductivity; The step S13: taking the average value of the actual fluctuation range of the process parameters as the process parameters required by the electromagnetic simulation, establishing a parameter space according to the geometric parameters of the layout of the passive device; obtaining the performance parameters by formula operation of parameter space sampling and electromagnetic simulation to establish a performance data set; Alternatively, the step S1 comprises: The step S11': providing a wafer of passive devices of multiple different layout sizes under the same device process; The step S12': establishing an equivalent circuit model for the passive device and obtaining the parameters of each element in the equivalent circuit model; taking the geometric parameters of the layout of each passive device and the parameters of each element in the corresponding equivalent circuit model as a group of data to form an element data set; The step S13': training the neural network model by using the element data set to obtain the network model of the equivalent circuit element, which is used to output the parameters of each element in the equivalent circuit model according to the geometric parameters of the layout of the passive device; The step S14': establishing a parameter space according to the geometric parameters of the layout of the passive device, obtaining the geometric parameters of the layout and the parameters of each element in the corresponding equivalent circuit model by parameter space sampling, and obtaining the performance parameters of the corresponding passive device by electromagnetic simulation to establish a performance data set. The computer program is executed by the processor to realize the method of any one of claims 1-8.

9. A computer readable medium having stored thereon a computer program, characterized in that ​

Images