Improved broadband active mixer reversible design method and system for reversible neural networks

Abstract

The present application relates to a method and system for improving the broadband active mixer reversible design method of reversible neural network, the method comprises: obtaining the response of the broadband active mixer, wherein the response comprises the conversion gain G and the noise figure NF; inputting the response into the R-INN model, and outputting the design parameters of the broadband active mixer, wherein the design parameters comprise resistance and capacitance, the R-INN model is constructed by combining real non-volume preserving transformation with reversible neural network, the R-INN model is obtained by training the training set, and the training set comprises the design parameters in the preset range and the corresponding response. By introducing an improved reversible neural network model combined with real non-volume preserving transformation technology, the accuracy of forward prediction is optimized, and the posterior distribution of the design parameters is effectively back calculated, thereby improving the efficiency and accuracy of the optimization design of the broadband active mixer.

Description

Technical Field

[0001] This invention relates to the field of electronic design automation technology, and in particular to a broadband active mixer reversible design method and system based on an improved reversible neural network. Background Technology

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

[0003] To address these challenges, researchers have proposed several optimization methods, such as heuristic algorithms and surrogate models, to predict and evaluate design performance. These methods shorten design cycles and reduce costs by minimizing the number of physical verifications. While they improve design efficiency to some extent, limitations remain: First, optimization algorithms may get stuck in local optima and miss global optima, especially in cases of large and complex design spaces. Second, although surrogate models can make rapid predictions without full circuit simulation, their accuracy heavily depends on the quality of the training data and the rationality of the model structure. Therefore, inaccurate models and insufficiently high-quality data can lead to biased design decisions, overlooking potentially excellent design solutions. With the increasing application of Invertible Neural Networks (INNs) in integrated circuit design automation (EDA), INNs have demonstrated unique potential in circuit reverse engineering. INN methods can predict circuit performance through forward inference and effectively predict posterior distributions through reverse inference, providing a new solution for the reversibility of circuit design. However, existing research shows that using INNs for circuit reverse engineering often comes at the cost of sacrificing forward prediction accuracy, which limits the application scope and efficiency of this method to some extent. Furthermore, in some literature, the generalizability and reliability of the INN method remain questionable due to the limited sample size. Summary of the Invention

[0004] The purpose of this invention is to provide an improved reversible neural network for the reversible design of broadband active mixers. By introducing an improved reversible neural network model that incorporates 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 broadband active mixer optimization design.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] On the one hand, it provides an improved method for reversible design of broadband active mixers using reversible neural networks, including:

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

[0008] The response is input into the R-INN model, which outputs the design parameters of the broadband active mixer. The active mixer is designed according to the design parameters, which include resistance and capacitance. The R-INN model is constructed by combining real-valued non-volume-preserving transformation with an invertible neural network. The R-INN model is obtained by training a training set, which 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 identical combinations of hierarchical layers and feature adjustment layers. Each combination of hierarchical layers includes: a batch normalization layer, a real NVP nonlinear transformation layer, and a Jacobian linear layer.

[0011] The batch normalization layer is used to batch normalize 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 normalized 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 the predicted response.

[0015] Optionally, the normalized data can be subjected to a nonlinear data transformation using the real-valued non-volume-preserving transformation as follows:

[0016] h1=x1

[0017]

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

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

[0020]

[0021] Where h is a vector formed by concatenating the two vectors h1 and h2 output from each flow step in Real NVP, x is the input vector, k is the k elements, m is an arbitrary function, and I d It is an identity matrix.

[0022] Optionally, outputting the design parameters of the broadband active mixer includes: obtaining the corresponding inverse circuit design parameter solution by backpropagating the label of each sample;

[0023] The method of backpropagation is as follows:

[0024]

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

[0026] On the other hand, it 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;

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

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

[0029] Optionally, the R-INN model includes several identical combinations of hierarchical layers and feature adjustment layers. Each combination of hierarchical layers includes: a batch normalization layer, a real NVP nonlinear transformation layer, and a Jacobian linear layer.

[0030] The batch normalization layer is used to batch normalize 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 normalized 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 the predicted response.

[0034] The beneficial effects of this invention are as follows: This invention proposes an R-INN framework—an INN model that combines Real-valued Non-Volume Preserving transformation (Real NVP) technology. While retaining the advantages of the original INN inverse inference of the posterior distribution of samples, this model significantly improves the accuracy of forward prediction for broadband active mixers and enables the inverse design of broadband active mixers within the same network. In inverse design tasks with output dimensions as high as 200, the proposed model can still accurately derive design parameters that match the performance of a given circuit. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a framework diagram of the R-INN model, an improved invertible neural network model based on real-valued non-volume-preserving transformation, according to an embodiment of the present invention.

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

[0038] Figure 3 This is a schematic diagram illustrating the forward and backward computation processes of the INN model in an embodiment of the present invention;

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

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

[0041] Figure 6 This is a flowchart illustrating the reversible design method of a broadband active mixer based on an improved reversible neural network, according to an embodiment of the present invention.

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

[0043] Figure 8 This is a comparison chart of the forward evaluation values ​​and PySpice simulation values ​​in an embodiment of the present invention;

[0044] Figure 9 This is a comparison chart of the NMSE loss of the reverse prediction for each test sample and the baseline NMSE loss in an embodiment of the present invention.

[0045] Figure 10 This is a schematic diagram of the conditional posterior distribution of the predicted design parameters of a low-power single-transistor active mixer according to an embodiment of the present invention, wherein (a) is a schematic diagram of the posterior distribution of the load impedance, (b) is a schematic diagram of the posterior distribution of the LO capacitor, and (c) is a schematic diagram of the posterior distribution of the RF capacitor. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] This embodiment provides an improved reversible neural network-based method for the reversible design of broadband active mixers, including:

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

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

[0051] Furthermore, training the R-INN model using the training set includes: using 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, such as Figure 6As shown, the data, including design parameters (load resistance RL, load capacitance CLO, feedback resistance CRF) and mixer responses G and NF), is first divided into a training set and a test set. The training set data is used to train the R-INN, with the design parameters as input and the mixer response as output. After the network is trained, the R-INN can be used for reversible prediction; that is, inputting the design parameters can predict the response, and vice versa.

[0053] Furthermore, the R-INN model includes several identical combinations of hierarchical layers and feature adjustment layers. Each combination of hierarchical layers includes: a batch normalization layer, a real NVP nonlinear transformation layer, and a Jacobian linear layer.

[0054] The batch normalization layer is used to batch normalize the design parameters within the preset range of input to obtain standardized data.

[0055] The real NVP nonlinear transformation layer is used to perform nonlinear data transformation on standardized data using 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 to calculate the Jacobian matrix of the transformation.

[0057] The feature adjustment layer is used for data adaptation based on the Jacobian matrix, and outputs 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, this model uses the dimension of the response signal as the dimension of the model input and output, with design parameters as input. Values ​​exceeding the dimension of the design parameters are padded with zeros, and a series of reversible layers are used to simulate complex data distributions. The design parameters are randomly generated within a certain range. Response curves are obtained through simulation by randomly generating some design parameters, and then the parameters and curves are correlated to form a dataset, which is used to train the R-INN.

[0059] To achieve reversible prediction of inputs and outputs, the designed network contains five identical sets of layers, each set consisting of: a batch normalization layer, a real NVP nonlinear transformation layer, and a Jacobian linear layer.

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

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

[0062] Jacobian Linear Layer: This layer processes the data through a linear transformation and simultaneously calculates the Jacobian matrix of the transformation, which is used for subsequent backpropagation optimization. This ensures the invertibility of the entire network while maintaining efficiency throughout the transformation process. This layer also uses 9-dimensional input.

[0063] The final layer is a feature adjustment layer that adapts to the data based on the Jacobian matrix and outputs the predicted response. The entire model architecture is reversible, meaning the input can be accurately reconstructed from the model's output. This property is particularly important for applications such as density estimation or generative modeling. This model is applied to a classification task to investigate its potential and effectiveness in handling complex classification problems.

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

[0065] h1=x1

[0066]

[0067] Here, h1 and h2 are two vectors output by each step of the Real NVP flow, x1 and x2 are two vectors obtained by separating the input x, and s and t are both vector coefficients of x1, where s is the scaling factor and t is the transformation factor. The computation consisting of the above s and t operations is called affine computation. Formally, the second formula corresponds to an affine transformation of x2, hence it is called the "affine coupling layer".

[0068] Furthermore, the Jacobian matrix can be represented using a block matrix as follows:

[0069]

[0070] Where h is a vector formed by concatenating the two vectors h1 and h2 output from each flow step in Real NVP, x is the input vector, k is the k elements, m is an arbitrary function, and I d It is an identity matrix.

[0071] Furthermore, the design parameters for the output broadband active mixer include: obtaining the corresponding inverse circuit design parameter solution by backpropagating the label of each sample;

[0072] The method of backpropagation is as follows:

[0073]

[0074] Where 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 scaling factor, and t is the transformation factor.

[0075] Specifically, INN is a flow-based model that assumes a sample x is drawn from the design space X, with a probability density of P. x (x), and the corresponding sample y drawn from the response space Y, whose probability density is unknown P. y If f(x) = f(y), then by transforming Y = f(x), we can apply variable transformation methods to establish the probability density relationship between them:

[0076]

[0077] In this architecture, all layers are integrated into a single function f. θ In this context, θ represents the network parameter set. For example... Figure 2 The basic unit of an INN network is a reversible block consisting of two complementary affine coupling layers, which can be trained in two directions. The latent variable z (usually assumed to be Gaussian distributed) is used to learn a 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], and then they are coupled through a coefficient e. s Transform with the affine function of t:

[0079]

[0080] For a given block output [y1, y2], it can be reversed in the following way:

[0081]

[0082] Equation (2) represents the forward mapping, while equation (3) represents the reverse mapping. For example... Figure 2 As shown, these transformations involve only element-wise addition and multiplication operations, thus avoiding the inverse operation on 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, thus enabling full learning of both forward and inverse processes, such as... Figure 3 .

[0083] Real NVP (Real-valued Non-Volume Preserving transformations) is a flow model, and its algorithmic process is characterized by the following aspects:

[0084] (1) As Figure 4 As shown, Real NVP concatenates the two vectors h1 and h2 from the flow output at each step into a single vector h, and then randomly reorders this vector. By randomly shuffling the entire vector output at each step, the information can be mixed more thoroughly and evenly.

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

[0086] h1=x1 (4)

[0087]

[0088] This transformation is not merely a simple addition, but includes scaling and translation operations, enabling the model to capture more complex distributions. In this way, Real NVP offers greater representational power and flexibility; because affine coupling layers can scale and translate data, they are generally more expressive when learning data distributions. The Jacobian matrix of affine coupling is still a triangular matrix, but the diagonals are not all 1s, and it is represented as a block matrix as follows:

[0089]

[0090] The determinant is the product of k elements. To ensure invertibility, all elements are constrained to be greater than zero. Due to this special structural design of RealNVP, the affine coupling layer allows for efficient reverse computation.

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

[0092] As multi-scale outputs at different locations, z1, z3, and z5 are not equal in status. The conditional probability formulas for z1, z3, and z5 are as follows:

[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 equation can be rewritten as:

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

[0096] Assuming the three probability distributions on the right-hand side are all normal distributions, where the mean and variance of p(z1|z2) are calculated from Z2, the mean and variance of p(z3|z4) are calculated from Z4, and the mean and variance of p(z5) are directly learned. Therefore, the above prior assumption is equivalent to the following variable substitution:

[0097]

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

[0099] These three transformations all result in a non-1 Jacobian determinant, meaning that a matrix of the form [formula missing] needs to be added to the loss. This item.

[0100] In this embodiment, Real NVP (Real Normalizing Flows) is integrated into the network structure of INN (Inverse Neural Network). The inverse network consists of many layers, such as... Figure 1 As shown, 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 determine the classification label during forward propagation and recover the original features through backward propagation, 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 the use of batch normalization, RealNVP nonlinear transformation layers, and Jacobi linear layers. This classifier not only performs forward classification tasks efficiently but also accurately recovers input features through inverse operations when necessary, providing verifiable evidence for classification decisions. In this way, Real NVP brings a new solution to traditional classification problems and opens new avenues for the interpretability and transparency of deep learning models.

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

[0102] This model provides a neural network structure with high efficiency and accuracy in high-dimensional design spaces by repeatedly applying batch normalization layers, RealNVP nonlinear layers, and Jacobi linear layers. This structure significantly outperforms traditional methods in both forward prediction accuracy and reverse inference capability, offering a new solution for exploring integrated circuit design spaces.

[0103] To verify the effectiveness of the R-INN method, this embodiment compares it with the traditional INN algorithm, generative adversarial networks (GANs), and variational autoencoders (VAEs). 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 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 (LO) frequency of 855 MHz, and operates with a 1V DC power supply and a current of 600 microamps (μA). The objectives include: (1) obtaining mixer design parameters that meet the given gain and noise specifications; and (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 in Table 1, the design space parameters for a low-power single-transistor mixer are as follows. It is important to note that the selection of parameters is limited by the boundaries of the transistor's operating region. In this example, three parameters were optimized: RL, CLO, and CRF. The optimization targets are the conversion gain G and the noise figure NF, defined as follows:

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

[0106] as well as

[0107]

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

[0109] Table 1

[0110]

[0111] To verify the effectiveness of the proposed improved inverse neural network (R-INN) model, this embodiment employs an experimental design similar to the original experiment, performing two types of inference evaluation: forward design evaluation and inverse design evaluation. In the forward evaluation, each parameter tuple is forward-inputted into the improved INN model to obtain the corresponding circuit performance metrics—gain and noise figure. In the inverse evaluation, the label of each sample is backpropagated 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 extends to the entire test set to ensure the breadth and reliability of the results.

[0112] This embodiment tests samples drawn from the test set: such as... Figure 8 Compared to traditional INNs, R-INNs can obtain the gain G and noise figure NF of a low-power single-transistor active mixer in forward evaluation. The design tuple {2.5kΩ, 0.5pF, 1.8pF} generated by the INN was validated using a trained R-INN model and a PySpice circuit simulator. The results predicted by the R-INN model almost perfectly match the response of the circuit simulator; for example... Figure 10 (a)-(c) show that the conditional posterior distribution p(x|y_target) of the design parameters for a low-power single-transistor active mixer can be predicted using the R-INN model. The black vertical line represents the point with the highest probability. When the design parameter point corresponding to the highest probability is sampled, the obtained tuple is {2.5kΩ, 0.5pF, 1.8pF}.

[0113] The NMSE loss results of inverse design are compared in Table 2. In this embodiment, 54 samples (20% of the total samples) were selected as the test set, and the overall mean normalized mean square error (NMSE) was calculated. The average NMSE of G and NF obtained by R-INN were 0.6577 and 0.5313, respectively, which are much lower than the test results of GAN, VAE, and INN. The inverse design accuracy reached 95%, which is significantly improved compared with other methods. R-INN outperformed traditional INN, VAE, and GAN in inverse design.

[0114] Furthermore, in this embodiment, each of the 54 samples in the test set was tested individually. The line graph of the NMSE results obtained through the improved model experiment is shown below. Figure 9 As shown, the NMSE of all test samples is lower than that of the best sample in the original paper's experiment. This not only demonstrates the superiority of the method in this embodiment but also shows its robustness across the entire test set.

[0115] Table 2

[0116]

[0117] On the other hand, this 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] The data acquisition module is used to acquire the response of the broadband active mixer, wherein the response includes the conversion gain G and the noise figure NF;

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

[0120] Furthermore, the R-INN model includes several identical combinations of hierarchical layers and feature adjustment layers. Each combination of hierarchical layers includes: a batch normalization layer, a real NVP nonlinear transformation layer, and a Jacobian linear layer.

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

[0122] The real NVP nonlinear transformation layer is used to perform nonlinear data transformation on standardized data using 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 to calculate the Jacobian matrix of the transformation.

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

[0125] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An improved method of broadband active mixer reversible design of reversible neural networks, characterized by, The method comprises the following steps: acquiring a response of a wideband active mixer, wherein the response comprises a conversion gain G and a noise figure NF; inputting the response into an R-INN model to output design parameters of the wideband active mixer, and designing the active mixer according to the design parameters, wherein the design parameters comprise resistance and capacitance, the R-INN model is constructed by combining a real non-volume preserving transformation with an invertible neural network, and the R-INN model is obtained by training a training set, wherein the training set comprises design parameters in a preset range and corresponding responses; the R-INN model comprises a plurality of groups of same hierarchical combinations and feature adjustment layers, each group of hierarchical combinations comprises a batch normalization layer, a real NVP nonlinear transformation layer and a Jacobian linear layer; the batch normalization layer is configured to perform batch normalization on the input design parameters in the preset range to obtain standardized data; the real NVP nonlinear transformation layer is configured to perform nonlinear data transformation on the standardized data by using the real non-volume preserving transformation; the Jacobian linear layer is configured to process output data of the real NVP nonlinear transformation layer by linear transformation and calculate a Jacobian matrix of the transformation; the feature adjustment layer is configured to perform data self-adaption based on the Jacobian matrix to output a predicted response.

2. The improved method of designing a wideband active mixer reversible neural network of claim 1, wherein, Training the R-INN model by using the training set comprises training the R-INN model by taking the design parameters in the preset range as input and taking the corresponding responses of the design parameters in the preset range as output.

3. The method of claim 1, wherein, The nonlinear data transformation on the standardized data by using the real non-volume preserving transformation is as follows: h1=x1 wherein h1 and h2 are two vectors output by each step of flow of the Real NVP, x1 and x2 are two vectors obtained by splitting the input x, s is a scale coefficient, and t is a transformation coefficient.

4. The method of claim 1, wherein, The Jacobian matrix is expressed by using a block matrix as follows: where h is a vector concatenating the two vectors h1, h2 output by Real NVP at each step, x is the input vector, k is the number of elements, m is an arbitrary function, I d is the identity matrix.

5. The method for improved wideband active mixer reversible design of reversible neural networks of claim 1, wherein, Outputting the design parameters of the wideband active mixer comprises obtaining corresponding inverse circuit design parameter solutions by back propagating labels of each sample. The back propagation method is as follows: wherein x1 and x2 are two vectors obtained by splitting the input x, y1 and y2 are two outputs of the module, s is a scale coefficient, and t is a transformation coefficient.

6. An improved broadband active mixer reversible design system for reversible neural networks, characterized by, The method comprises a data acquisition module and a design parameter calculation module. The data acquisition module is configured to acquire a response of a wideband active mixer, wherein the response comprises a conversion gain G and a noise figure NF. The design parameter calculation module is configured to input the response into an R-INN model to output design parameters of the wideband active mixer, and design the active mixer according to the design parameters, wherein the design parameters comprise resistance and capacitance, the R-INN model is constructed by combining a real non-volume preserving transformation with an invertible neural network, and the R-INN model is obtained by training a training set, wherein the training set comprises design parameters in a preset range and corresponding responses; the R-INN model comprises a plurality of groups of same hierarchical combinations and feature adjustment layers, each group of hierarchical combinations comprises a batch normalization layer, a real NVP nonlinear transformation layer and a Jacobian linear layer; The batch normalization layer is configured to perform batch normalization on the input design parameter in the preset range to obtain standardized data. The real NVP nonlinear transformation layer is configured to perform nonlinear data transformation on the standardized data by using the real value non-volume preserving transformation. The Jacobian linear layer is configured to process output data of the real NVP nonlinear transformation layer by linear transformation and calculate a Jacobian matrix of the transformation. The feature adjustment layer is configured to perform data self-adaptation based on the Jacobian matrix and output a predicted response.

Images