Reverse design method, system and terminal equipment for on-chip spiral inductors for radio frequency
By constructing a reverse design method based on Transformer and convolutional neural networks, and combining physical constraints and evolutionary algorithms, the problems of low simulation efficiency and spatial conflict resolution in the design of on-chip spiral inductors for radio frequency (RF) chips were solved, achieving efficient and stable geometric structure optimization and improving the performance of RF front-end circuits.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-02
AI Technical Summary
The design of on-chip spiral inductors for radio frequency in the existing technology suffers from problems such as low simulation efficiency, distortion of physical characteristics and conflict in the solution space of reverse design. In particular, the simulation of three-dimensional full-wave electromagnetic field is time-consuming and the results are unstable. The lack of physical constraints in deep learning models leads to the problem of non-uniqueness mapping.
We employ a physical perception forward proxy model based on the Transformer architecture and an inverse generative model based on a one-dimensional convolutional neural network, combined with an evolutionary algorithm. By introducing a physical constraint loss function and a low-frequency weighting strategy, we establish a high-fidelity mapping from geometric parameters to electromagnetic response, and then use a hybrid intelligent optimization strategy to find the optimal geometric structure.
It significantly improves design efficiency, eliminates non-physical prediction phenomena, ensures the accuracy and stability of inductance values, enables rapid positioning of the optimal geometry that combines high inductance accuracy with the maximum quality factor, and enhances the performance of RF front-end circuits.
Smart Images
Figure CN122133492A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radio frequency integrated circuit auxiliary design technology, specifically to a method, system, and terminal equipment for reverse design of on-chip spiral inductors for radio frequency circuits. Background Technology
[0002] On-chip spiral inductors are core passive devices in RF front-end circuits that store energy through magnetic fields. Their inductance and quality factor directly determine the performance of key modules such as low-noise amplifiers and voltage-controlled oscillators. However, due to the influence of high-frequency skin effect, proximity effect and substrate loss, their electromagnetic characteristics exhibit high nonlinearity.
[0003] Currently, the industry mainly relies on three-dimensional full-wave electromagnetic field simulation software (such as HFSS) for design. Although the accuracy is high, each simulation takes a long time, and the consumption of computing resources increases exponentially with the increase of geometric parameters, which seriously restricts the efficiency of design iteration. In recent years, deep learning-based reverse design methods have emerged. Although they have improved the speed to some extent, existing technologies generally face the problem of "non-uniqueness" mapping. That is, multiple sets of geometric structures may correspond to similar S-parameters, which makes it difficult for the reverse model to converge or the output results to be unstable. In addition, conventional neural networks lack physical constraint mechanisms, and the generated S-parameters often exhibit non-physical phenomena such as violation of passivity or numerical oscillation in the low-frequency band, which leads to the failure of DC inductance prediction and makes it difficult to meet the dual requirements of high accuracy and physical consistency in actual engineering.
[0004] To this end, the present invention proposes a reverse design method, system and terminal equipment for on-chip spiral inductors for radio frequency. Summary of the Invention
[0005] The purpose of this invention is to provide a method, system, and terminal device for reverse design of on-chip spiral inductors for radio frequency, so as to solve the problems of low simulation efficiency, distortion of physical characteristics, and conflict of reverse design solution space in the prior art.
[0006] According to a first aspect of the present invention, in order to achieve the above-mentioned objective, the present invention provides the following technical solution: a method for reverse design of an on-chip spiral inductor for radio frequency, comprising the following steps: The geometric parameters of the spiral inductor are received and the parameter space is set. Samples are generated in batches using a three-dimensional full-wave electromagnetic field simulation tool. The full-band S-parameters are extracted and physical validity screening and data preprocessing are performed to generate a physical dataset of the RF spiral inductor. A physical perception forward proxy model is constructed based on the Transformer architecture. The forward proxy model maps geometric parameters to S-parameters. In the training phase of the forward proxy model, a physical constraint loss function based on the basic principles of microwave networks is introduced and fine-tuned using a low-frequency weighting strategy to establish a high-fidelity forward mapping from geometric parameters to electromagnetic response. A reverse generative model based on a one-dimensional convolutional neural network is constructed. A cascaded neural network strategy is adopted to concatenate the reverse generative model with a parameter-frozen forward proxy model. The reverse generative model is trained by minimizing the error between the reconstructed S-parameters and the target S-parameters to solve the non-unique mapping problem in the geometric parameter domain. The ideal S-parameter curve is synthesized based on the user's design indicators. The ideal S-parameter curve is then input into the trained inverse generation model to obtain the initial solution. The initial solution is then injected into the evolutionary algorithm, which uses the physical perception forward surrogate model as the evaluation function to perform global iterative optimization. Finally, the optimal geometric structure parameters that meet the design indicators are output.
[0007] Furthermore, the geometric parameters of the spiral inductor are received and the parameter space is defined. Samples are generated in batches using a three-dimensional full-wave electromagnetic field simulation tool to extract the full-band S-parameters. Physical validity screening and data preprocessing are then performed to generate a physical dataset of the RF spiral inductor, as detailed below: (21) Parametric modeling: Determine the key geometric variables and their value ranges of the octagonal spiral inductor. The geometric variables include the number of turns N, line width W, line spacing S and outer diameter D; set geometric constraints and eliminate geometric combinations with inner diameters smaller than a preset threshold or physical self-intersections. (22) Automated simulation and extraction: Write automated scripts to drive electromagnetic field simulation software to perform Latin hypercube sampling in the preset geometric parameter space; perform frequency scanning simulation for each sample to extract the two-port S-parameter matrix in the 0.1 GHz to 20 GHz frequency band; (23) Physical validity screening: Calculate the inductance value L, quality factor Q and self-resonant frequency SRF of each sample, and remove invalid samples based on physical rules. The removal criteria include: self-resonant frequency is lower than the upper limit of the working frequency band, quality factor is always negative in the entire frequency band or low-frequency inductance value is abnormally divergent. (24) Data preprocessing: Using a single-ended grounding port configuration, the dual-port S-parameters are converted into single-ended input impedance to calculate physical indices; the geometric parameters and S-parameters are normalized to the maximum and minimum values respectively, and mapped to the intervals of [0, 1] and [-1,1] to construct a tensor dataset for neural network training.
[0008] Furthermore, a physical perception forward proxy model is constructed based on the Transformer architecture, which maps geometric parameters to S-parameters. During the training phase of the forward proxy model, a physical constraint loss function based on the fundamental principles of microwave networks is introduced, and a low-frequency weighting strategy is used for fine-tuning to establish a high-fidelity forward mapping from geometric parameters to electromagnetic response, as detailed below: (31) Model architecture construction: Transformer Encoder is used as the backbone network. The input layer contains position encoding to inject frequency sequence information, and multi-head self-attention mechanism is used to capture the long-range dependency of S parameters in the whole frequency band. (32) Physical constraint training: Define the physical perception loss function It includes a data fitting term, a reciprocity constraint term, and a passivity-free constraint term; the reciprocity constraint term is used to penalize... and The difference is that the passive constraint term is used to penalize the part of the S-parameter magnitude greater than 1, so as to ensure that the model output conforms to the basic principles of microwave networks; (33) Low-frequency weighted fine-tuning: After the model pre-training converges, a weighted mean square error strategy is used for fine-tuning; low-frequency data points with frequencies less than 1 GHz are identified and a loss weight of 10 to 50 times is applied to them, forcing the model to prioritize fitting the low-frequency DC characteristics and eliminating non-physical numerical oscillations of the neural network in the low-frequency band.
[0009] Furthermore, we define the physical perception loss function. Specifically: in This represents the data fitting error; This is a reciprocity constraint used to force the transmission coefficients of the network to be symmetric. Used to penalize non-physical predictions with a modulus greater than 1; and These are the weighting coefficients.
[0010] Furthermore, in the low-frequency weighted fine-tuning strategy in step (33), a mean square error penalty with a weight coefficient w=20 is applied to the sampling points with frequencies less than 1.0 GHz, while the weights of the remaining frequency bands remain at 1.0, forcing the model to prioritize fitting the DC characteristics.
[0011] Furthermore, an inverse generative model based on a one-dimensional convolutional neural network is constructed. A cascaded neural network strategy is adopted, concatenating the inverse generative model with a parameter-frozen forward surrogate model. The inverse generative model is trained by minimizing the error between the reconstructed S-parameters and the target S-parameters to address the non-uniqueness mapping problem in the geometric parameter domain, as detailed below: (61) Constructing a generator network: Construct an inverse generator network containing multiple one-dimensional convolutional layers and fully connected layers. Its input is a full-band S-parameter sequence, and its output is a normalized geometric parameter prediction value. (62) Cascaded connection: The inverse generator network is connected in series with the trained and parameter-frozen forward proxy model; wherein, the output of the inverse network is directly used as the input of the forward model, and the output of the forward model is used as the reconstructed S-parameters; (63) Gradient backpropagation: Calculate the mean square error between the reconstructed S-parameters and the true target S-parameters, and use it as the training signal; propagate the gradient through the frozen forward model to the inverse generation network through the chain rule, thereby updating the weights of the inverse network and solving the non-unique mapping problem in the geometric parameter domain.
[0012] Furthermore, an ideal S-parameter curve is synthesized based on the user-designed metrics. This ideal S-parameter curve is then input into a trained inverse generation model to obtain an initial solution. This initial solution is then injected into an evolutionary algorithm, which uses a physical perception forward proxy model as the evaluation function for global iterative optimization. Finally, the optimal geometric structure parameters that satisfy the design metrics are output, as detailed below: (71) Physical target synthesis: Based on the target frequency input by the user Target inductance value and target quality factor Based on the RL series equivalent circuit model Synthesize ideal broadband S-parameter curves; (72) Inverse initial guess: Input the ideal S-parameter curve into the trained inverse generation model and output a set of geometric parameters as the initial guess solution, providing a high-confidence search starting point for the optimization algorithm; (73) Global population initialization: Construct the initial population for the differential evolution algorithm, inject the initial guess solution into the population, and generate the remaining individuals randomly within the design space; (74) Proxy model iterative optimization: In each round of evolutionary iteration, the positive proxy model is called to predict the S parameters of each individual in the population; a fitness function containing a sensitivity error term and a quality factor penalty term is constructed. If the predicted quality factor is lower than the target value, an exponential penalty is applied; the population is updated through mutation, crossover and selection operations until the fitness function converges and the optimal geometric structure is output.
[0013] Furthermore, the fitness function Includes a sensitivity error term and a Q-value penalty term: If the predicted Q value is less than the target Q value, an exponential penalty is applied, forcing the algorithm to evolve towards a high Q value region, and finally outputting the optimal geometric structure parameters that meet the design specifications.
[0014] According to a second aspect of the present invention, the present invention provides a reverse design system for an on-chip RF spiral inductor, for implementing the reverse design method for an on-chip RF spiral inductor described in the first aspect, comprising: The physical dataset construction module is used to receive the geometric parameters of the spiral inductor and set the parameter space, generate samples in batches using a three-dimensional full-wave electromagnetic field simulation tool, extract the full-band S-parameters, and perform physical validity screening and data preprocessing to generate the physical dataset of the radio frequency spiral inductor. The physical perception forward proxy model construction module is used to build a physical perception forward proxy model based on the Transformer architecture. The forward proxy model maps geometric parameters to S-parameters. In the training phase of the forward proxy model, a physical constraint loss function based on the basic principles of microwave networks is introduced and fine-tuned using a low-frequency weighting strategy to establish a high-fidelity forward mapping from geometric parameters to electromagnetic response. The inverse generative model building module is used to build an inverse generative model based on a one-dimensional convolutional neural network. It adopts a cascaded neural network strategy to concatenate the inverse generative model with a parameter-frozen forward proxy model. The inverse generative model is trained by minimizing the error between the reconstructed S-parameters and the target S-parameters to solve the non-unique mapping problem in the geometric parameter domain. The hybrid optimization design module is used to synthesize an ideal S-parameter curve based on user design indicators, input the ideal S-parameter curve into a trained inverse generation model to obtain an initial solution, and inject the initial solution into an evolutionary algorithm. The physical perception forward surrogate model is used as the evaluation function to perform global iterative optimization, and finally output the optimal geometric structure parameters that meet the design indicators.
[0015] According to a third aspect of the present invention, a terminal device is provided, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor, and when the processor loads and executes the computer program, it employs the reverse design method for on-chip spiral inductors of radio frequency described in the first aspect.
[0016] This invention has at least the following beneficial effects: 1. This invention replaces the traditional three-dimensional full-wave electromagnetic field numerical simulation with a data-driven deep neural network, compressing the time for evaluating the electromagnetic performance of a single sample from minutes to milliseconds. This makes it feasible to perform thousands or even tens of thousands of iterations of optimization in the global optimization process. This efficiency breakthrough completely breaks the rigid constraint of traditional simulation methods on the number of design iterations, enabling designers to fully explore the high-dimensional geometric parameter space within a limited time window, significantly shortening the design cycle of radio frequency integrated circuits and reducing the consumption of computing resources and R&D costs.
[0017] 2. This invention introduces reciprocity constraints and passivity constraints during the training process of the forward proxy model, forcing the neural network output to conform to the basic laws of microwave networks. This fundamentally eliminates the non-physical prediction phenomena commonly found in traditional pure data-driven models. At the same time, through a low-frequency weighted fine-tuning strategy, the model is forced to prioritize fitting the electromagnetic characteristics of DC and near-DC frequency bands, effectively eliminating numerical oscillations and phase distortion in the low-frequency band, and ensuring the accuracy and stability of DC inductance prediction. 3. This invention adopts a cascaded neural network architecture, which solves the problem of non-bijective mapping between geometric parameters and electrical performance, and ensures the convergence stability of the inverse model. Combined with the differential evolution hybrid optimization strategy, it overcomes the limitation that a single inverse network is difficult to search for extreme values, and can quickly locate the optimal geometric structure with both high sensitivity accuracy and maximum quality factor, thereby improving the overall performance of the RF front-end circuit.
[0018] 4. This invention organically integrates the high-dimensional mapping capability of deep learning with the global search capability of evolutionary algorithms to form a hybrid intelligent optimization strategy. The initial guess solution provided by the inverse generation model injects high-confidence prior knowledge into the differential evolutionary algorithm, ensuring that the optimization population is in a high-performance region in the early stages of iteration, avoiding blind random search. The forward surrogate model serves as an efficient fitness evaluation function, supporting the evolutionary algorithm to perform large-scale iterations in the global space. The fitness function design, which includes both inductance accuracy and quality factor, combined with an exponential penalty mechanism, ensures that the algorithm always evolves towards the high Q-value region during convergence. This hybrid strategy overcomes the dual limitations of a single neural network's difficulty in searching for extreme values and the low search efficiency of a single evolutionary algorithm. It can actively explore the extreme structure with the highest quality factor in the design space while ensuring that the inductance value meets the design specifications, providing greater design margin for improving the performance of RF front-end circuits.
[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the overall process of the method described in this invention; Figure 2 This is a 3D structural diagram of the on-chip spiral inductor in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the network architecture of the physical perception forward proxy model in Embodiment 1 of the present invention; Figure 4 This is a comparative diagram of the fitting effect of the low-frequency weighted fine-tuning strategy on the S-parameters in Embodiment 1 of the present invention, where (a) is the unweighted fine-tuning and (b) is the weighted fine-tuning. Figure 5 This is a block diagram illustrating the principle of training the inverse generative model using a cascaded neural network architecture in Embodiment 1 of the present invention. Figure 6 This is a flowchart illustrating the algorithm logic of the hybrid intelligent optimization design engine in an embodiment of the present invention. Detailed Implementation
[0021] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.
[0022] Example 1: Please see Figures 1-6 This invention provides a technical solution: a reverse design method for on-chip spiral inductors for radio frequency (RF) chips, comprising the following steps: S1: Constructing a high-precision physical dataset for RF spiral inductors: Define the inductor geometry parameter space, use a three-dimensional full-wave electromagnetic field simulation tool to generate samples in batches, extract S-parameters across the entire frequency band, and perform physical validity screening and data preprocessing. S1.1: Parametric modeling: Determine the key geometric variables and their value ranges for the octagonal spiral inductor. The geometric variables include the number of turns N, line width W, line spacing S, and outer diameter D. Set geometric constraints and eliminate geometric combinations with an inner diameter smaller than a preset threshold or physical self-intersection. S1.2: Automated Simulation and Extraction: Develop automated scripts to drive electromagnetic field simulation software, perform Latin hypercube sampling within a preset geometric parameter space; perform frequency scanning simulation for each sample, and extract the two-port S-parameter matrix in the 0.1GHz to 20GHz frequency band; S1.3: Physical validity screening: Calculate the inductance value L, quality factor Q, and self-resonant frequency SRF for each sample; remove invalid samples based on physical rules, with removal criteria including: self-resonant frequency below the upper limit of the operating frequency band, quality factor that is always negative across the entire frequency band, or abnormal divergence of inductance value at low frequencies; S1.4: Data preprocessing: Using a single-ended grounding port configuration, the dual-port S-parameters are converted into single-ended input impedance to calculate physical parameters; the geometric parameters and S-parameters are respectively subjected to min-max normalization, which maps them to the intervals of [0, 1] and [-1,1] to construct a tensor dataset for neural network training; S2: Construct and train a physical perception positive proxy model: Build a deep neural network based on the Transformer architecture and map geometric parameters to S-parameters; introduce a physical constraint loss function during the training phase and fine-tune it using a low-frequency weighting strategy to obtain a high-precision positive prediction model; S2.1: Model Architecture Construction: Transformer Encoder is used as the backbone network, and the input layer contains positional encoding to inject frequency sequence information; Multi-Head Self-Attention is used to capture the long-range dependencies of S parameters across the entire frequency band. S2.2: Physical Constraint Training: Define the physical perception loss function Ltatol, which includes a data fitting term, a reciprocity constraint term, and a passive constraint term. The reciprocity constraint term is used to penalize the difference between S21 and S12, and the passive constraint term is used to penalize the part of the S-parameter magnitude greater than 1, ensuring that the model output conforms to the basic principles of microwave networks. Define the physical perception loss function Specifically: in This represents the data fitting error; This is a reciprocity constraint used to force the transmission coefficients of the network to be symmetric. Used to penalize non-physical predictions with a modulus greater than 1; and These are the weighting coefficients; S2.3: Low-frequency weighted fine-tuning: After the model pre-training converges, a weighted mean square error (MSE) strategy is used for fine-tuning; low-frequency data points with frequencies less than 1 GHz are identified and subjected to loss weights of 10 to 50 times, forcing the model to prioritize fitting low-frequency DC characteristics and eliminating non-physical numerical oscillations of the neural network in the low-frequency band. S3: Construct and train the inverse generative model with a tandem architecture: Build a generator based on a one-dimensional convolutional neural network, adopt the tandem neural network strategy, cascade the inverse generative model with the parameter-frozen forward proxy model, and train the generator by minimizing the reconstruction error; S3.1: Constructing the generator network: Construct an inverse generator network containing multiple layers of one-dimensional convolutional layers (1D-CNN) and fully connected layers. Its input is a full-band S-parameter sequence, and its output is a normalized geometric parameter prediction value. S3.2: Cascaded Concatenation: The inverse generative network is concatenated with the forward surrogate model trained and with its parameters frozen in S2; the output of the inverse network is directly used as the input of the forward model, and the output of the forward model is used to reconstruct the S parameters; S3.3: Gradient backpropagation: Calculate the mean square error between the reconstructed S-parameters and the true target S-parameters, and use this as the training signal; propagate the gradient through the frozen forward model to the inverse generation network through the chain rule, thereby updating the weights of the inverse network and solving the non-unique mapping problem in the geometric parameter domain. S4: Implement hybrid intelligent optimization design: synthesize ideal physical curves according to design indicators, obtain initial solutions through inverse models, and inject them into differential evolution algorithms. Use the forward surrogate model as the evaluation function to perform global iterative optimization and output the optimal geometric parameters. S4.1: Physical Target Synthesis: Based on the target operating frequency, target inductance value and target quality factor input by the user, the ideal broadband S-parameter curve is analytically calculated based on the RL series equivalent circuit model. S4.2: Inverse Initial Guess: Input the ideal S-parameter curve into the trained inverse generation model, and output a set of geometric parameters as the initial guess solution (Seed), providing a high-confidence search starting point for the optimization algorithm; S4.3: Global population initialization: Construct the initial population for the Differential Evolution algorithm, inject the initial guess solution into the population, and generate the remaining individuals randomly within the design space; S4.4: Proxy Model Iterative Optimization: In each round of evolutionary iteration, the positive proxy model is called to predict the S-parameters of each individual in the population; a fitness function (Cost Function) containing a sensitivity error term and a quality factor penalty term is constructed. If the predicted quality factor is lower than the target value, an exponential penalty is applied; the population is updated through mutation, crossover, and selection operations until the fitness function converges and the optimal geometric structure is output. Furthermore, the fitness function Includes a sensitivity error term and a Q-value penalty term: If the predicted Q value is less than the target Q value, an exponential penalty is applied, forcing the algorithm to evolve towards the high Q value region, and finally outputting geometric parameters that balance sensitivity accuracy and high performance.
[0023] The technical solution of the present invention will be further described below with reference to specific embodiments: Step 1: Constructing a high-precision physical dataset for RF spiral inductors. In this embodiment, an octagonal on-chip spiral inductor is selected as the research object (e.g., Figure 2 As shown, the geometric parameter space is set based on the gallium arsenide IPD process design rules; 1) Parameter space definition: Set the range of values for geometric parameters, where the number of turns N∈[1.5, 6.5] (step size 0.5), line width W∈[5, 35]μm, line spacing S∈[5, 15]μm, and outer diameter D_out∈[100, 650]μm; 2) Automated simulation: Using Python scripts to drive the three-dimensional full-wave electromagnetic field simulation software (Ansys HFSS), discrete frequency sweep simulation was performed in the frequency band from 0.1 GHz to 20.0 GHz, with 401 sampling points, to obtain two-port S-parameter data; 3) Data cleaning and preprocessing: Samples with self-resonant frequency (SRF) below 10GHz were removed; non-passive samples with a consistently negative quality factor (Q) across the entire frequency band were removed; and abnormal samples with low-frequency inductance values less than 0.1nH were removed. Finally, an "elite dataset" containing approximately 8000 valid samples was constructed. The Min-Max normalization method was used to map the geometric parameters to the interval [0, 1], and the real and imaginary parts of the S-parameter matrix were mapped to the interval [-1, 1] to construct tensor data in the format (N, 401, 8). Step 2: Construct and train a physical perception positive agent model, the model architecture of which is as follows: Figure 2 As shown, a Transformer Encoder architecture is adopted. The input layer contains geometric feature embedding and positional encoding, the intermediate layer contains 3 Transformer Blocks (each layer contains 4 self-attention heads, and the hidden layer dimension is 128), and the output layer is mapped back to the S-parameter sequence through a fully connected layer. Define a physical perception loss function, which includes a data fitting term and a reciprocity constraint term. =1.0) and passive constraint terms ( =0.5); Step 3: Implement a low-frequency weighted fine-tuning strategy, targeting... Figure 3 The low-frequency S-parameter prediction divergence problem shown is addressed by applying a mean square error penalty of weight coefficient w=20 to sampling points with frequencies less than 1.0 GHz after model pre-training, while keeping the weights of other frequency bands at 1.0, thus forcing the model to preferentially fit the DC characteristics. Step 4: Construct a reverse generation model of the serial architecture, such as Figure 4 As shown, an inverse generator based on a one-dimensional convolutional neural network (1D-CNN) is constructed. The input is a sequence of 401 S-parameters, and the output is 4-dimensional geometric parameters. This is concatenated with the forward surrogate model trained and with its parameters frozen in step 2. The S-parameters corresponding to the actual geometric parameters are input into the inverse network to obtain the predicted geometric parameters. ;Will Input the frozen forward model to obtain the reconstructed S-parameters. ,calculate Compared with the true S-parameters The error between them is calculated and backpropagated to update the weights of the inverse network. This process avoids the multi-value conflict problem in direct geometric parameter regression and ensures the convergence of the inverse model. Step 5: Implement hybrid intelligent optimization design, the process is as follows: Figure 5 As shown, an ideal S-parameter curve is synthesized based on user-designed metrics and input into the inverse model to obtain an initial guessed solution. This solution is then injected into the initial population of the differential evolution algorithm. Using the forward surrogate model as the evaluation function, the algorithm searches the global space for the geometric structure with a satisfaction value error of less than 5% and the largest quality factor.
[0024] Results Analysis Figure 3 The following comparison of the effects of the low-frequency weighting strategy in this specific embodiment is given. It can be seen that the unweighted model (a) has a huge fitting error in S21 near 0.1-1GHz and cannot be used for DC design; while after adopting the weighting strategy of the present invention (b), the curve is smooth and converges to the physical true value, and the prediction error is reduced to less than 0.5%.
[0025] Table 1 Comparison of Design Results and Target Performance Table 1 shows a comparison between the design results and target performance of the present invention under different design objectives.
[0026] The above results demonstrate that this invention, by replacing traditional finite element simulation with a positive surrogate model, reduces the single evaluation time from minutes to milliseconds (speedup > 1000 times), making it possible to perform thousands of global iterations for optimization in a short time. The hybrid optimization strategy combines the "intuition" of deep learning with the "search" capability of evolutionary algorithms, enabling not only rapid finding of feasible solutions but also uncovering the ultimate design with the highest quality factor within the design space.
[0027] In summary, this invention replaces traditional finite element simulation with a physically-sensory forward proxy model, reducing the time required for a single evaluation from minutes to milliseconds, significantly improving the efficiency of global iterative optimization while maintaining computational accuracy. By utilizing a physical constraint loss function and a low-frequency weighted fine-tuning strategy, it effectively eliminates non-physical numerical oscillations in the low-frequency band of traditional neural networks, ensuring the physical consistency and engineering usability of the design results across the entire frequency band. The adoption of a cascaded neural network architecture solves the problem of non-bijective mapping between geometric parameters and electrical performance, ensuring the convergence stability of the inverse model. Combined with a differential evolution hybrid optimization strategy, it overcomes the limitation of a single inverse network in finding extreme values, enabling rapid location of the optimal geometric structure that combines high sensitivity accuracy and maximum quality factor, thus improving the overall performance of the RF front-end circuit.
[0028] Example 2: This embodiment provides a reverse design system for on-chip RF spiral inductors, used to implement the reverse design method for on-chip RF spiral inductors described in Embodiment 1, including: The physical dataset construction module is used to receive the geometric parameters of the spiral inductor and set the parameter space, generate samples in batches using a three-dimensional full-wave electromagnetic field simulation tool, extract the full-band S-parameters, and perform physical validity screening and data preprocessing to generate the physical dataset of the radio frequency spiral inductor. The physical perception forward proxy model construction module is used to build a physical perception forward proxy model based on the Transformer architecture. The forward proxy model maps geometric parameters to S-parameters. In the training phase of the forward proxy model, a physical constraint loss function based on the basic principles of microwave networks is introduced and fine-tuned using a low-frequency weighting strategy to establish a high-fidelity forward mapping from geometric parameters to electromagnetic response. The inverse generative model building module is used to build an inverse generative model based on a one-dimensional convolutional neural network. It adopts a cascaded neural network strategy to concatenate the inverse generative model with a parameter-frozen forward proxy model. The inverse generative model is trained by minimizing the error between the reconstructed S-parameters and the target S-parameters to solve the non-unique mapping problem in the geometric parameter domain. The hybrid optimization design module is used to synthesize an ideal S-parameter curve based on user design indicators, input the ideal S-parameter curve into a trained inverse generation model to obtain an initial solution, and inject the initial solution into an evolutionary algorithm. The physical perception forward surrogate model is used as the evaluation function to perform global iterative optimization, and finally output the optimal geometric structure parameters that meet the design indicators.
[0029] Example 3: This embodiment provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it adopts the reverse design method of the radio frequency on-chip spiral inductor described in Embodiment 1.
[0030] It should be noted that the terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server, and the terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.
[0031] Furthermore, the processor can be a central processing unit (CPU). Of course, depending on the actual use, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be used. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it in this regard.
[0032] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0033] For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on," "mounted on," "fixed to," or "set on" another element, it may be directly on the other element or there may be an intermediate element present. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be an intermediate element present. The terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible embodiments.
[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
[0035] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
Claims
1. A reverse design method for an on-chip RF spiral inductor, characterized in that, Includes the following steps: The geometric parameters of the spiral inductor are received and the parameter space is set. Samples are generated in batches using a three-dimensional full-wave electromagnetic field simulation tool. The full-band S-parameters are extracted and physical validity screening and data preprocessing are performed to generate a physical dataset of the RF spiral inductor. A physical perception forward proxy model is constructed based on the Transformer architecture. The forward proxy model maps geometric parameters to S-parameters. In the training phase of the forward proxy model, a physical constraint loss function based on the basic principles of microwave networks is introduced and fine-tuned using a low-frequency weighting strategy to establish a high-fidelity forward mapping from geometric parameters to electromagnetic response. A reverse generative model based on a one-dimensional convolutional neural network is constructed. A cascaded neural network strategy is adopted to concatenate the reverse generative model with a parameter-frozen forward proxy model. The reverse generative model is trained by minimizing the error between the reconstructed S-parameters and the target S-parameters to solve the non-unique mapping problem in the geometric parameter domain. The ideal S-parameter curve is synthesized based on the user's design indicators. The ideal S-parameter curve is then input into the trained inverse generation model to obtain the initial solution. The initial solution is then injected into the evolutionary algorithm, which uses the physical perception forward surrogate model as the evaluation function to perform global iterative optimization. Finally, the optimal geometric structure parameters that meet the design indicators are output.
2. The reverse design method for on-chip RF spiral inductors according to claim 1, characterized in that: The geometric parameters of the spiral inductor are received and the parameter space is defined. Samples are generated in batches using a three-dimensional full-wave electromagnetic field simulation tool. The full-band S-parameters are extracted, and physical validity screening and data preprocessing are performed to generate a physical dataset of the RF spiral inductor, as follows: (21) Parametric modeling: Determine the key geometric variables and their value ranges for the octagonal spiral inductor. The geometric variables include the number of turns N, line width W, line spacing S, and outer diameter D. Set geometric constraints to eliminate geometric combinations with inner diameters smaller than a preset threshold or those that physically self-intersect; (22) Automated simulation and extraction: Write automated scripts to drive electromagnetic field simulation software to perform Latin hypercube sampling in a preset geometric parameter space; For each sample, a frequency scanning simulation was performed to extract the two-port S-parameter matrix in the frequency band from 0.1 GHz to 20 GHz; (23) Physical validity screening: Calculate the inductance value L, quality factor Q and self-resonant frequency SRF of each sample, and remove invalid samples based on physical rules. The removal criteria include: self-resonant frequency is lower than the upper limit of the working frequency band, quality factor is always negative in the entire frequency band or low-frequency inductance value is abnormally divergent. (24) Data preprocessing: Using a single-ended grounding port configuration, the dual-port S-parameters are converted into single-ended input impedance to calculate physical indices; the geometric parameters and S-parameters are normalized to the maximum and minimum values respectively, and mapped to the intervals of [0, 1] and [-1,1] to construct a tensor dataset for neural network training.
3. The reverse design method for on-chip spiral inductors according to claim 2, characterized in that: A physical perception forward proxy model is constructed based on the Transformer architecture. This forward proxy model maps geometric parameters to S-parameters. During the training phase of the forward proxy model, a physical constraint loss function based on the fundamental principles of microwave networks is introduced, and a low-frequency weighting strategy is used for fine-tuning to establish a high-fidelity forward mapping from geometric parameters to electromagnetic response. Specifically: (31) Model architecture construction: Transformer Encoder is used as the backbone network. The input layer contains position encoding to inject frequency sequence information, and multi-head self-attention mechanism is used to capture the long-range dependency of S parameters in the whole frequency band. (32) Physical constraint training: Define the physical perception loss function It includes a data fitting term, a reciprocity constraint term, and a passivity-free constraint term; the reciprocity constraint term is used to penalize... and The difference is that the passive constraint term is used to penalize the part of the S-parameter magnitude greater than 1, ensuring that the model output conforms to the basic principles of microwave networks; (33) Low-frequency weighted fine-tuning: After the model pre-training converges, a weighted mean square error strategy is used for fine-tuning; low-frequency data points with frequencies less than 1 GHz are identified and a loss weight of 10 to 50 times is applied to them, forcing the model to prioritize fitting the low-frequency DC characteristics and eliminating non-physical numerical oscillations of the neural network in the low-frequency band.
4. The reverse design method for on-chip spiral inductors according to claim 3, characterized in that: Define the physical perception loss function Specifically: in This represents the data fitting error; This is a reciprocity constraint used to force the transmission coefficients of the network to be symmetric. Used to penalize non-physical predictions with a modulus greater than 1; and These are the weighting coefficients.
5. The reverse design method for on-chip RF spiral inductors according to claim 3, characterized in that: In the low-frequency weighted fine-tuning strategy in step (33), a mean square error penalty with a weight coefficient w=20 is applied to the sampling points with frequencies less than 1.0 GHz, while the weights of the remaining frequency bands are kept at 1.0, forcing the model to prioritize fitting the DC characteristics.
6. The reverse design method for on-chip spiral inductors according to claim 5, characterized in that: A reverse generative model based on a one-dimensional convolutional neural network is constructed. A cascaded neural network strategy is adopted, concatenating the reverse generative model with a parameter-frozen forward surrogate model. The reverse generative model is trained by minimizing the error between the reconstructed S-parameters and the target S-parameters to address the non-uniqueness mapping problem in the geometric parameter domain. Specifically: (61) Constructing a generator network: Construct an inverse generator network containing multiple one-dimensional convolutional layers and fully connected layers. Its input is a full-band S-parameter sequence, and its output is a normalized geometric parameter prediction value. (62) Cascaded connection: The inverse generator network is connected in series with the trained and parameter-frozen forward proxy model; wherein, the output of the inverse network is directly used as the input of the forward model, and the output of the forward model is used as the reconstructed S-parameters; (63) Gradient backpropagation: Calculate the mean square error between the reconstructed S-parameters and the true target S-parameters, and use it as the training signal; propagate the gradient through the frozen forward model to the inverse generation network through the chain rule, thereby updating the weights of the inverse network and solving the non-unique mapping problem in the geometric parameter domain.
7. The reverse design method for on-chip spiral inductors according to claim 6, characterized in that: An ideal S-parameter curve is synthesized based on user-designed metrics. This ideal S-parameter curve is then input into a trained inverse generation model to obtain an initial solution. This initial solution is then injected into an evolutionary algorithm, which uses a physical perception forward surrogate model as the evaluation function for global iterative optimization. Finally, the optimal geometric structure parameters that satisfy the design metrics are output, as detailed below: (71) Physical target synthesis: Based on the target frequency input by the user Target inductance value and target quality factor Based on the RL series equivalent circuit model Synthesize ideal broadband S-parameter curves; (72) Inverse initial guess: Input the ideal S-parameter curve into the trained inverse generation model and output a set of geometric parameters as the initial guess solution, providing a high-confidence search starting point for the optimization algorithm; (73) Global population initialization: Construct the initial population for the differential evolution algorithm, inject the initial guess solution into the population, and generate the remaining individuals randomly within the design space; (74) Proxy model iterative optimization: In each round of evolutionary iteration, the positive proxy model is called to predict the S parameters of each individual in the population; a fitness function containing a sensitivity error term and a quality factor penalty term is constructed. If the predicted quality factor is lower than the target value, an exponential penalty is applied; the population is updated through mutation, crossover and selection operations until the fitness function converges and the optimal geometric structure is output.
8. The reverse design method for on-chip spiral inductors according to claim 7, characterized in that: The fitness function Includes a sensitivity error term and a Q-value penalty term: If the predicted Q value is less than the target Q value, an exponential penalty is applied, forcing the algorithm to evolve towards a high Q value region, and finally outputting the optimal geometric structure parameters that meet the design specifications.
9. A reverse design system for an on-chip RF spiral inductor, used to implement the reverse design method for an on-chip RF spiral inductor as described in any one of claims 1 to 8, characterized in that, include: The physical dataset construction module is used to receive the geometric parameters of the spiral inductor and set the parameter space, generate samples in batches using a three-dimensional full-wave electromagnetic field simulation tool, extract the full-band S-parameters, and perform physical validity screening and data preprocessing to generate the physical dataset of the radio frequency spiral inductor. The physical perception forward proxy model construction module is used to build a physical perception forward proxy model based on the Transformer architecture. The forward proxy model maps geometric parameters to S-parameters. In the training phase of the forward proxy model, a physical constraint loss function based on the basic principles of microwave networks is introduced and fine-tuned using a low-frequency weighting strategy to establish a high-fidelity forward mapping from geometric parameters to electromagnetic response. The inverse generative model building module is used to build an inverse generative model based on a one-dimensional convolutional neural network. It adopts a cascaded neural network strategy to concatenate the inverse generative model with a parameter-frozen forward proxy model. The inverse generative model is trained by minimizing the error between the reconstructed S-parameters and the target S-parameters to solve the non-unique mapping problem in the geometric parameter domain. The hybrid optimization design module is used to synthesize an ideal S-parameter curve based on user design indicators, input the ideal S-parameter curve into a trained inverse generation model to obtain an initial solution, and inject the initial solution into an evolutionary algorithm. The physical perception forward surrogate model is used as the evaluation function to perform global iterative optimization, and finally output the optimal geometric structure parameters that meet the design indicators.
10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs the reverse design method for on-chip spiral inductors of any one of claims 1 to 8.