Reversible design method and system for broadband active mixer based on improved reversible neural network

Abstract

The invention relates to a broadband active mixer reversible design method and system based on an improved reversible neural network, and the method comprises the steps: obtaining the response of a broadband active mixer, and the response comprises a conversion gain G and a noise index NF; the response is input into the R-INN model, design parameters of the broadband active mixer are output, the active mixer is designed according to the design parameters, the design parameters comprise resistance and capacitance, the R-INN model is constructed through combination of real-value non-volume preserving transformation and a reversible neural network, and the R-INN model is obtained through training of a training set. The training set comprises design parameters in a preset range and corresponding responses. According to the method, the improved reversible neural network model combined with the real-value non-volume preserving transformation technology is introduced, so that the precision of forward prediction is optimized, and the posterior distribution of the design parameters is effectively reversely deduced, and therefore, the efficiency and accuracy of the optimization design of the broadband active mixer are improved.

Description

Technical Field

[0001] The invention relates to the technical field of electronic design automation, and in particular to a reversible design method and system for a broadband active mixer for improving a reversible neural network. Background Art

[0002] With the continuous advancement of integrated circuit technology and the continuous miniaturization of process nodes, the complexity and integration of broadband active mixers are growing exponentially. This trend has led to a sharp increase in the complexity of broadband active mixer design, making the design process very time-consuming and costly.

[0003] To address the above challenges, researchers have proposed some optimization methods, such as heuristic algorithms and proxy models to predict and evaluate the performance of designs. These methods are all aimed at shortening the design cycle and reducing costs by reducing the number of physical verifications. Although they have improved the efficiency of design to some extent, they still have limitations: First, the optimization algorithm may fall into the local optimal solution and miss the global optimal solution, especially when the design space is large and complex. Second, although the proxy model can make rapid predictions without full circuit simulation, the accuracy of its predictions depends heavily on the quality of the training data and the rationality of the model structure. Therefore, inaccurate models and data of insufficient quality are likely to lead to deviations in design decisions, thereby ignoring some potential excellent design solutions. With the deepening of the application of Invertible Neural Networks (INN) in the field of integrated circuit design automation (EDA), INN has shown unique potential in circuit reverse design. The INN method can predict circuit performance through forward inference and effectively predict the posterior distribution through reverse inference, providing a new solution for the reversibility of circuit design. However, existing studies have shown that the use of INN for circuit reverse design is often accompanied by a sacrifice in forward prediction accuracy, which to a certain extent restricts the application scope and efficiency of this method. In addition, in some literature, due to the limitation of the test sample size, the generalizability and reliability of the INN method are still questionable. Summary of the invention

[0004] The purpose of the present invention is to provide a reversible design method and system for a broadband active mixer using an improved reversible neural network. By introducing an improved reversible neural network model combined with real-valued non-volume preserving transformation technology, the accuracy of forward prediction is optimized and the posterior distribution of design parameters is effectively inferred, thereby improving the efficiency and accuracy of the optimization design of the broadband active mixer.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] On the one hand, a reversible design method of a broadband active mixer for improving a reversible neural network is provided, comprising:

[0007] Obtaining a response of a broadband active mixer, wherein the response includes a conversion gain G and a noise figure NF;

[0008] The response is input into an R-INN model, and design parameters of the broadband active mixer are output. The active mixer is designed according to the design parameters, wherein the design parameters include resistance and capacitance. The R-INN model is constructed by combining a real-valued non-volume preserving transformation with a reversible neural network. The R-INN model is obtained by training with a training set, and the training set includes design parameters and corresponding responses within a preset range.

[0009] Optionally, training the R-INN model using a training set includes: training the R-INN model by taking the design parameters within the preset range as input and the responses corresponding to the design parameters within the preset range as output.

[0010] Optionally, the R-INN model includes several groups of identical hierarchical combinations and feature adjustment layers, each hierarchical combination including: a batch normalization layer, a real NVP nonlinear transformation layer and a Jacobi linear layer;

[0011] The batch normalization layer is used to perform batch normalization on the input design parameters within the preset range to obtain standardized data;

[0012] The real NVP nonlinear transformation layer is used to perform nonlinear data transformation on the standardized data using the real-valued non-volume preserving transformation;

[0013] The Jacobian linear layer is used to process the output data of the real NVP nonlinear transformation layer through linear transformation and calculate the Jacobian matrix of the transformation;

[0014] The feature adjustment layer is used to perform data adaptation based on the Jacobian matrix and output a predicted response.

[0015] Optionally, the real-valued non-volume preserving transformation is used to perform a nonlinear data transformation on the standardized data:

[0016] h1=x1

[0017]

[0018] Among them, h1 and h2 are the two vectors output by each step of Real NVP flow, x1 and x2 are the two vectors obtained by separating the input x, s is the scale scaling factor, and t is the transformation coefficient.

[0019] Optionally, the Jacobian matrix is ​​represented by a block matrix:

[0020]

[0021] Among them, h is a vector formed by Real NVP concatenating the two vectors h1 and h2 output by each step of the flow, x is the input vector, k is the kth element, m is an arbitrary function, and I d is the identity matrix.

[0022] Optionally, outputting the design parameters of the broadband active mixer includes: acquiring a corresponding inverse circuit design parameter solution by back-propagating a label of each sample;

[0023] The method of back propagation is:

[0024]

[0025] Among them, x1 and x2 are two vectors obtained by separating the input x, y1 and y2 are the two outputs of the module, s is the scale scaling factor, and t is the transformation coefficient.

[0026] On the other hand, a broadband active mixer reversible design system based on an improved reversible neural network is provided, including a data acquisition module and a design parameter calculation module;

[0027] The data acquisition module is used to obtain the response of the broadband active mixer, wherein the response includes a conversion gain G and a noise figure NF;

[0028] The design parameter calculation module is used to input the response into the R-INN model, output the design parameters of the broadband active mixer, and design the active mixer according to the design parameters, wherein the design parameters include resistance and capacitance, the R-INN model is constructed by combining a real-valued non-volume preserving transformation with a reversible neural network, and the R-INN model is obtained by training with a training set, and the training set includes design parameters and corresponding responses within a preset range.

[0029] Optionally, the R-INN model includes several groups of identical hierarchical combinations and feature adjustment layers, each hierarchical combination including: a batch normalization layer, a real NVP nonlinear transformation layer and a Jacobi linear layer;

[0030] The batch normalization layer is used to perform batch normalization on the input design parameters within the preset range to obtain standardized data;

[0031] The real NVP nonlinear transformation layer is used to perform nonlinear data transformation on the standardized data using the real-valued non-volume preserving transformation;

[0032] The Jacobian linear layer is used to process the output data of the real NVP nonlinear transformation layer through linear transformation and calculate the Jacobian matrix of the transformation;

[0033] The feature adjustment layer is used to perform data adaptation based on the Jacobian matrix and output a predicted response.

[0034] The beneficial effects of the present invention are as follows: the present invention proposes an R-INN framework, an INN model that combines the real-valued Non-Volume Preserving transformation (Real NVP) technology. The model significantly improves the accuracy of forward prediction of broadband active mixers while retaining the advantages of the original INN reverse inference sample posterior distribution, and can realize reverse design of broadband active mixers in the same network. In the task of reverse design, the output dimension is as high as 200 dimensions, and the proposed model can still accurately reversely derive design parameters that match the given circuit performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 It is a framework diagram of the real-valued non-volume preserving transformation improved reversible neural network model R-INN according to an embodiment of the present invention;

[0037] Figure 2 This is an INN network structure diagram of an embodiment of the present invention;

[0038] Figure 3 Schematic diagram of the forward and reverse calculation process of the INN model according to an embodiment of the present invention;

[0039] Figure 4 Schematic diagram of random shuffling between RealNVP layers according to an embodiment of the present invention;

[0040] Figure 5 Schematic diagram of a multi-scale structure diagram in RealNVP according to an embodiment of the present invention;

[0041] Figure 6 A flowchart of a reversible design method for a broadband active mixer of an improved reversible neural network according to an embodiment of the present invention;

[0042] Figure 7 A schematic diagram of a low-power single-transistor active mixer according to an embodiment of the present invention;

[0043] Figure 8 A comparison diagram of forward evaluation values ​​and PySpice simulation values ​​of an embodiment of the present invention;

[0044] Fig. 9 A comparison diagram of the NMSE loss of the reverse prediction of each test sample and the baseline NMSE loss in an embodiment of the present invention;

[0045] Fig.10 Schematic diagram of conditional posterior distribution of design parameters of a low-power single-transistor active mixer predicted in an embodiment of the present invention, wherein (a) is a schematic diagram of the posterior distribution of load impedance, (b) is a schematic diagram of the posterior distribution of LO capacitance, and (c) is a schematic diagram of the posterior distribution of RF capacitance. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] This embodiment provides a reversible design method for a broadband active mixer of an improved reversible neural network, comprising:

[0049] Obtaining a response of a broadband active mixer, wherein the response includes a conversion gain G and a noise figure NF;

[0050] The response is input into the R-INN model, and the design parameters of the broadband active mixer are output. The active mixer is designed according to the design parameters, where the design parameters include resistance and capacitance. The R-INN model is constructed by combining real-valued non-volume preserving transformation with a reversible neural network. The R-INN model is obtained by training with a training set, which includes design parameters and corresponding responses within a preset range.

[0051] Furthermore, training the R-INN model through the training set includes: taking the design parameters within a preset range as input and the responses corresponding to the design parameters within the preset range as output to train the R-INN model.

[0052] Specifically, Figure 6As shown, first, the data containing the design parameters load resistance RL, load capacitance CLO, feedback resistance CRF and mixer response G, NF are divided into a training set and a test set. The data of the training set is used to train R-INN, where the design parameters are used as input and the mixer response is used as output. After the network training is completed, R-INN can be used for reversible prediction, that is, the response can be predicted by inputting the design parameters, and conversely, the design parameters can be predicted by inputting the mixer response.

[0053] Furthermore, the R-INN model includes several groups of identical hierarchical combinations and feature adjustment layers, each hierarchical combination includes: batch normalization layer, real NVP nonlinear transformation layer and Jacobi linear layer;

[0054] Batch normalization layer, used to batch normalize the design parameters within the preset range of the input to obtain standardized data;

[0055] The real NVP nonlinear transformation layer is used to perform nonlinear data transformation on the normalized data using a real-valued non-volume preserving transformation;

[0056] The Jacobian linear layer is used to process the output data of the real NVP nonlinear transformation layer through linear transformation and calculate the Jacobian matrix of the transformation;

[0057] The feature adjustment layer is used to adapt the data based on the Jacobian matrix and output the predicted response.

[0058] Specifically, the R-INN model combines the reversible transformation of Real NVP with the architecture of INN, such as Figure 1 As shown. The model uses the dimension of the response signal as the dimension of the model input and output, the design parameters as input, and fills the values ​​beyond the design parameter dimension with zero, and uses a series of reversible layers to simulate complex data distribution. Among them, the design parameters are randomly generated within a certain range, and some design parameters are randomly generated. The response curve is obtained through simulation, and then the parameters-curve correspond to form a data set, which is used to train R-INN.

[0059] In order to achieve reversible prediction of input and output, the designed network contains five groups of identical hierarchical combinations, each of which includes: batch normalization layer, real NVP nonlinear transformation layer and Jacobi linear layer in sequence.

[0060] Batch Normalization Layer (BatchNorm1d): used to normalize the input design parameters, standardize the distribution of data, and reduce internal covariate shift during training to speed up training and improve model stability. This layer uses 9-dimensional input.

[0061] Real NVP nonlinear transformation layer (Nonlinear): The Real NVP nonlinear transformation layer is responsible for implementing reversible complex transformations, performing nonlinear data transformations, and retaining the reversibility of the data. This layer also uses 9-dimensional input, enabling the model to learn and reproduce the complex distribution of input data.

[0062] Jacobian Linear: This layer processes the data through linear transformation and calculates the Jacobian matrix of the transformation for subsequent back-propagation optimization. It ensures the reversibility of the entire network while maintaining efficiency throughout the conversion process. This layer also uses 9-dimensional input.

[0063] The last layer is the feature adjustment layer, which performs data adaptation based on the Jacobian matrix and outputs the predicted response. The entire model architecture is reversible, that is, the input can be accurately reconstructed from the output of the model, which is particularly important for applications such as density estimation or generative modeling. The model is applied to classification tasks to study its potential and effectiveness in handling complex classification problems.

[0064] Furthermore, a real-valued non-volume preserving transformation is used to perform nonlinear data transformation on the standardized data:

[0065] h1=x1

[0066]

[0067] Among them, h1 and h2 are two vectors output by each flow step of Real NVP, x1 and x2 are two vectors separated from the input x, s and t are both vector coefficients of x1, s is the scale coefficient, and t is the transformation coefficient. The calculation composed of the above s and t operations is called affine calculation. Formally, the second formula corresponds to an affine transformation of x2, so it is called "affine coupling layer".

[0068] Furthermore, the Jacobian matrix is ​​expressed as a block matrix:

[0069]

[0070] Among them, h is a vector formed by Real NVP concatenating the two vectors h1 and h2 output by each step of the flow, x is the input vector, k is the kth element, m is an arbitrary function, and I d is the identity matrix.

[0071] Furthermore, outputting the design parameters of the broadband active mixer includes: obtaining a corresponding inverse circuit design parameter solution by back-propagating the label of each sample;

[0072] The method of back propagation is:

[0073]

[0074] Among them, x1 and x2 are two vectors obtained by separating the input vector x, y1 and y2 are the outputs of the module, s is the scale scaling coefficient, and t is the transformation coefficient.

[0075] Specifically, INN is a flow-based model. Suppose a sample x is drawn from the design space X, and its probability density is P x (x), and the corresponding sample y drawn from the response space Y, whose probability density is unknown P y (y), then by transforming Y = f(X), the variable transformation method can be applied to establish the probability density relationship between them:

[0076]

[0077] All layers of the INN architecture are integrated into a single function f θ In , θ represents the network parameter set. Figure 2 The basic unit of the INN network is a reversible block consisting of two complementary affine coupled layers, which can be trained in two directions. The latent variable z (usually assumed to be Gaussian) is used to learn the nonlinear transformation between the known distribution in the latent space and the original data distribution.

[0078] In the affine coupling block, the input vector is split into two parts [x1, x2], after which they are coupled by a coefficient e s Transformed by the affine function of t:

[0079]

[0080] And for a given block output [y1, y2], it can be reversed by:

[0081]

[0082] Formula (2) represents the forward mapping, while formula (3) represents the reverse mapping. Figure 2 As shown in , these transformations only involve element-wise addition and multiplication operations, thus avoiding the need to inverse the scale s(·) and translation t(·) networks when computing the inverse mapping. The bijective nature of the INN model allows for bidirectional operation and training, so that both the forward and reverse processes can be fully learned, as Figure 3 .

[0083] Real NVP (Real-valued Non-Volume Preserving transformations) is a flow model, and its algorithm flow is manifested in the following aspects:

[0084] (1) Figure 4 As shown in the figure, Real NVP concatenates the two vectors h1 and h2 output by each flow step into a vector h, and then re-arranges the vector randomly. By randomly shuffling the entire vector output by each step, the information can be mixed more evenly.

[0085] (2) Figure 3 As shown in Figure 1, Real NVP introduces an affine coupling layer, which is a general coupling layer that allows more complex transformations to be performed on part of the input:

[0086] h1=x1 (4)

[0087]

[0088] This transformation is not just a simple addition, but includes scale and translation operations, which allows the model to capture more complex distributions. In this way, Real NVP provides stronger representation and flexibility; because the affine coupling layer can scale and translate the data, it is usually more expressive when learning data distribution. The Jacobian matrix of the affine coupling is still a triangular matrix, but the diagonal is not all 1, and it is represented by a block matrix:

[0089]

[0090] The determinant is the product of k elements. To ensure reversibility, each element is constrained to be greater than zero. Due to the special structural design of RealNVP, the affine coupling layer allows efficient inverse calculation.

[0091] (3) Figure 5 As shown in the figure, RealNVP introduces a multi-scale structure. This is a strategy that reduces the complexity of the model and improves the results. After the original input goes through the first flow operation (compound of multiple affine coupling layers), the output is the same size as the input. At this time, the input is split into two halves z1 and z2 (along the channel axis), where z1 is directly output and only z2 is sent to the next flow operation, and so on. For example, in the special case in the figure, the final output consists of z1, z3, and z5, and the total size is the same as the input.

[0092] As multi-scale outputs at different positions, the status of z1, z3, and z5 are unequal. The conditional probability formula of z1, z3, and z5 is:

[0093] p(z1,z3,z5)=p(z1|z3,z5)p(z3|z5)p(z5) (7)

[0094] Since z3 and z5 are completely determined by z2, and z5 is completely determined by z4, the above formula can be rewritten as:

[0095] p(z1,z3,z5)=p(z1|z2)p(z3|z4)p(z5) (8)

[0096] Assume that the three probability distributions on the right are all normal distributions, where the mean variance of p(z1|z2) is calculated by Z2, the mean variance of p(z3|z4) is calculated by Z4, and the mean variance of p(z5) is directly learned. Therefore, the above prior assumptions are equivalent to the following variable substitutions:

[0097]

[0098] in, is the standard normal distribution, μ is the mean, and σ is the variance.

[0099] All three transformations will result in a non-1 Jacobian determinant, which means adding a loss of the form This one.

[0100] In this embodiment, Real NVP (Real Normalizing Flows) is integrated into the network structure of INN (Inverse Neural Network). The reversible network consists of many layers, such as Figure 1 As shown in the figure, each layer contains RealNVP, aiming to explore the potential utility of Real NVP in classification tasks. A significant advantage of reversibility in classification tasks is the ability to recover the original features through back propagation while determining the classification label during the forward propagation process, which provides new possibilities in terms of information preservation and model interpretability. The Real NVP structure enables the model to learn complex feature distributions, which is particularly important when dealing with highly nonlinear and multimodal data. The proposed model constructs a powerful classifier by alternating batch normalization, RealNVP nonlinear transformation layers, and Jacobi linear layers. The classifier can not only perform forward classification tasks efficiently, but also accurately recover input features through inverse operations when necessary, providing a verifiable basis for classification decisions. In this way, Real NVP brings new solutions to traditional classification problems and opens up new avenues for the interpretability and transparency of deep learning models.

[0101] The entire model is deployed on the specified computing device. By calling `to(device)`, the model parameters and calculations are migrated to the specified device (such as GPU or CPU) to improve computing efficiency and processing power.

[0102] The model combines batch normalization layers, RealNVP nonlinear layers, and Jacobi linear layers for multiple repetitions, providing a neural network structure that is efficient and accurate in high-dimensional design space. This structure can significantly outperform traditional methods in terms of forward prediction accuracy and reverse inference capabilities, providing a new solution for integrated circuit design space exploration.

[0103] In order to verify the effectiveness of the R-INN method, this embodiment compares the traditional INN algorithm, the generative adversarial network (GAN) and the variational autoencoder (VAE). The active mixer used in this embodiment is a nonlinear device that realizes frequency summation and difference frequency, and its performance is mainly measured by the conversion gain or loss and the amount of noise introduced into the circuit. The mixer in this embodiment is a downconverter with a radio frequency (RF) operating range of 0.1 to 10 GHz, a local oscillator frequency (LO) of 855 MHz, and operates at a 1 V DC power supply with a current of 600 microamperes (μA). The objectives include: (1) obtaining mixer design parameters that meet the established gain and noise index specifications; (2) verifying the accuracy of the design parameters through forward evaluation.

[0104] The design parameters of the mixer are as follows: Figure 7 As shown, the design space parameters of the low-power single-transistor mixer are shown in Table 1. It should be noted that the selection of parameters is limited by the limits of the transistor working area. In this example, three parameters are optimized, namely RL, CLO and CRF. The optimization targets are conversion gain G and noise figure NF, which are defined as follows:

[0105] G(dB)=P IF -P RF (10)

[0106] as well as

[0107]

[0108] Among them, P IF and P RF are the power at the intermediate frequency (IF) and radio frequency (RF) ports, respectively, and N i G and N 0(mixer) are the input noise and the noise added by the mixer (both directed to the IF port). The RF is scanned from 0.1 GHz to 10 GHz with a step size of 100 MHz. Therefore, each tuple in the design space corresponds to a gain and noise figure with 100 frequency points. In this embodiment, 270 sample data are generated using PySpice, which are divided into a training set and a test set according to 0.8 and 0.2.

[0109] Table 1

[0110]

[0111] In order to verify the effectiveness of the proposed improved inverse neural network (R-INN) model, this embodiment adopts an experimental design similar to the original experiment, and performs two types of reasoning evaluations: forward design evaluation and inverse design evaluation. In the forward evaluation, each parameter tuple is forward input to the improved INN model to obtain the corresponding circuit performance indicators-gain and noise index. In the reverse evaluation, the label of each sample is back-propagated through the model to obtain the corresponding inverse circuit design parameter solution. This evaluation process is not limited to a single or a few samples, but is extended to the entire test set to ensure the breadth and reliability of the results.

[0112] This embodiment tests samples extracted from the test set: Figure 8 Compared with the traditional INN, R-INN can obtain the gain G and noise figure NF of the low-power single-transistor active mixer in the forward evaluation. The design tuple {2.5kΩ, 0.5pF, 1.8pF} generated by INN is verified using the trained R-INN model and PySpice circuit simulator. The results predicted by the R-INN model are almost completely consistent with the responses of the circuit simulator; Fig.10 (a)-(c), The R-INN model can predict the conditional posterior distribution p(x|y_target) of the design parameters of a low-power single-transistor active mixer. The black vertical line indicates the point with the highest probability. When sampling the design parameter point with the highest probability, the tuple obtained is {2.5kΩ, 0.5pF, 1.8pF}.

[0113] The results of the comparison of the NMSE loss results of the inverse design NMSE are shown in Table 2. This embodiment selects 54 samples (accounting for 20% of the total number of samples) as the test set, and calculates the overall average normalized mean square error (NMSE). The average NMSE of G and NF obtained by R-INN is 0.6577 and 0.5313, which is much lower than the test results of GAN, VAE and INN. The inverse design accuracy reaches 95%, which is greatly improved compared with other methods. R-INN surpasses traditional INN, VAE and GAN in inverse design.

[0114] In addition, this embodiment tests each of the 54 samples in the test set separately. The NMSE result line chart measured by the improved model experiment is as follows: Fig. 9 As shown in the figure, it can be observed that the NMSE of all test samples is lower than the test result of the best sample in the original paper experiment. This not only proves the superiority of the method in this embodiment, but also shows its robustness on the entire test set.

[0115] Table 2

[0116]

[0117] On the other hand, the present embodiment also provides a broadband active mixer reversible design system based on an improved reversible neural network, including a data acquisition module and a design parameter calculation module;

[0118] A data acquisition module, used for acquiring a response of the broadband active mixer, wherein the response includes a conversion gain G and a noise figure NF;

[0119] The design parameter calculation module is used to input the response into the R-INN model, output the design parameters of the broadband active mixer, and design the active mixer according to the design parameters, wherein the design parameters include resistance and capacitance. The R-INN model is constructed by combining a real-valued non-volume preserving transformation with a reversible neural network. The R-INN model is obtained by training with a training set, and the training set includes the design parameters and corresponding responses within a preset range.

[0120] Furthermore, the R-INN model includes several groups of identical hierarchical combinations and feature adjustment layers, each hierarchical combination includes: batch normalization layer, real NVP nonlinear transformation layer and Jacobi linear layer;

[0121] The batch normalization layer is used to batch normalize the design parameters within the preset range of the input to obtain standardized data;

[0122] The real NVP nonlinear transformation layer is used to perform nonlinear data transformation on the normalized data using a real-valued non-volume preserving transformation;

[0123] The Jacobian linear layer is used to process the output data of the real NVP nonlinear transformation layer through linear transformation and calculate the Jacobian matrix of the transformation;

[0124] Feature adjustment layer for data adaptation based on Jacobian matrix.

[0125] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A reversible design method for a broadband active mixer based on an improved reversible neural network, characterized in that: include: Obtaining a response of a broadband active mixer, wherein the response includes a conversion gain G and a noise figure NF; The response is input into an R-INN model, and design parameters of the broadband active mixer are output. The active mixer is designed according to the design parameters, wherein the design parameters include resistance and capacitance. The R-INN model is constructed by combining a real-valued non-volume preserving transformation with a reversible neural network. The R-INN model is obtained by training with a training set, and the training set includes design parameters and corresponding responses within a preset range.

2. The reversible design method of a broadband active mixer of an improved reversible neural network according to claim 1, characterized in that: Training the R-INN model through the training set includes: taking the design parameters within the preset range as input and the responses corresponding to the design parameters within the preset range as output to train the R-INN model.

3. The reversible design method of broadband active mixer of improved reversible neural network according to claim 1, characterized in that: The R-INN model includes several groups of identical hierarchical combinations and feature adjustment layers, each hierarchical combination includes: a batch normalization layer, a real NVP nonlinear transformation layer and a Jacobi linear layer; The batch normalization layer is used to perform batch normalization on the input design parameters within the preset range to obtain standardized data; The real NVP nonlinear transformation layer is used to perform nonlinear data transformation on the standardized data using the real-valued non-volume preserving transformation; The Jacobian linear layer is used to process the output data of the real NVP nonlinear transformation layer through linear transformation and calculate the Jacobian matrix of the transformation; The feature adjustment layer is used to perform data adaptation based on the Jacobian matrix and output a predicted response.

4. The reversible design method of a broadband active mixer of an improved reversible neural network according to claim 3, characterized in that: The real-valued non-volume preserving transformation is used to perform a nonlinear data transformation on the standardized data: h1=x1 Among them, h1 and h2 are the two vectors output by each step of Real NVP flow, x1 and x2 are the two vectors obtained by separating the input x, s is the scale scaling factor, and t is the transformation coefficient.

5. The reversible design method of broadband active mixer of improved reversible neural network according to claim 3, characterized in that: The Jacobian matrix is ​​expressed as a block matrix: Among them, h is a vector formed by Real NVP concatenating the two vectors h1 and h2 output by each step of the flow, x is the input vector, k is the kth element, m is an arbitrary function, and I d is the identity matrix.

6. The reversible design method of broadband active mixer of improved reversible neural network according to claim 1, characterized in that: Outputting the design parameters of the broadband active mixer includes: acquiring a corresponding inverse circuit design parameter solution by back-propagating the label of each sample; The method of back propagation is: Among them, x1 and x2 are two vectors obtained by separating the input x, y1 and y2 are the two outputs of the module, s is the scale scaling factor, and t is the transformation coefficient.

7. Improved reversible neural network broadband active mixer reversible design system, characterized in that, Including data acquisition module and design parameter calculation module; The data acquisition module is used to obtain the response of the broadband active mixer, wherein the response includes a conversion gain G and a noise figure NF; The design parameter calculation module is used to input the response into the R-INN model, output the design parameters of the broadband active mixer, and design the active mixer according to the design parameters, wherein the design parameters include resistance and capacitance, the R-INN model is constructed by combining a real-valued non-volume preserving transformation with a reversible neural network, and the R-INN model is obtained by training with a training set, and the training set includes design parameters and corresponding responses within a preset range.

8. The broadband active mixer reversible design system of improved reversible neural network according to claim 7, characterized in that: The R-INN model includes several groups of identical hierarchical combinations and feature adjustment layers, each hierarchical combination includes: a batch normalization layer, a real NVP nonlinear transformation layer and a Jacobi linear layer; The batch normalization layer is used to perform batch normalization on the input design parameters within the preset range to obtain standardized data; The real NVP nonlinear transformation layer is used to perform nonlinear data transformation on the standardized data using the real-valued non-volume preserving transformation; The Jacobian linear layer is used to process the output data of the real NVP nonlinear transformation layer through linear transformation and calculate the Jacobian matrix of the transformation; The feature adjustment layer is used to perform data adaptation based on the Jacobian matrix and output a predicted response.

Images