A transmission line s-parameter prediction method and system based on a time series convolution network

Abstract

The application discloses a transmission line S parameter prediction method and system based on a time sequence convolution network, and belongs to the technical field of radio frequency and microwave circuit modeling and intelligent prediction. The method comprises the following steps: obtaining structure description information of a transmission line to be predicted, and parsing the structure description information into a structured sequence; encoding the structured sequence to form a model fixed-length sequence input representation, and inputting the model fixed-length sequence input representation into a time sequence convolution neural network prediction model trained by grouping and frequency band division, and outputting frequency domain orthogonal basis coefficients corresponding to real parts and imaginary parts of S11, S21, S12 and S22 in each sub-frequency band; and performing inverse transformation reconstruction on the frequency domain orthogonal basis coefficients to obtain a transmission line wideband S parameter response in a target frequency range. The application can significantly reduce the calling frequency of full-wave electromagnetic simulation, improve the S parameter prediction efficiency of a complex radio frequency interconnection structure, and has good generalization ability, modular expansion ability and engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency and microwave circuit modeling and intelligent prediction technology, and particularly relates to a method and system for predicting transmission line S-parameters based on temporal convolutional networks. Background Technology

[0002] In RF and microwave system design, transmission lines, as one of the most basic and critical structural units, are widely used in filters, couplers, combiners, and RF front-end modules. The electrical length, impedance distribution, and structural topology of transmission lines directly affect the system's frequency response characteristics and power transmission efficiency. As operating frequencies become wider and higher, the complexity of transmission line structures continues to increase. Small changes in geometric dimensions often cause significant phase shifts and amplitude fluctuations. Therefore, how to quickly and accurately predict the full-band S-parameter response of a transmission line structure has become a fundamental problem in RF interconnect design, parameter optimization, and automated evaluation.

[0003] In existing technologies, a parameterized electromagnetic simulation model of a selected microwave radio frequency transmission structure is established. A pre-trained neural network model is then built based on simulation sample data, and a measured neural network model is built by combining this with physical samples. Finally, the adjoint neural network model of the pre-trained neural network model and the measured neural network model are fused for prediction. This addresses the problems of large sample quantity requirements, high parameter extraction costs, and difficulty in obtaining accurate and sufficient data in process IP simulation modeling. However, its focus is on achieving high-fidelity modeling of radio frequency transmission structures through the fusion of simulation and measurement in process IP simulation scenarios, primarily addressing the problem of fusion and correction between simulation models and measured data under process fluctuation conditions. In contrast, it does not model the structured sequence representation of multi-segment, multi-layer, and variable topology transmission lines, nor does it propose a dedicated sequence modeling mechanism for predicting broadband S-parameters caused by local routing order, interlayer via switching, and different path combinations in transmission lines. A review of existing technologies reveals several shortcomings in current methods for predicting the electrical performance of transmission lines: First, they still heavily rely on full-wave electromagnetic simulation, resulting in high sample acquisition costs and hindering large-scale structural searches and rapid iterations. Second, they lack a unified structural representation suitable for multi-segment traces, interlayer vias, and complex topological changes, making it difficult for models to fully utilize the sequential information and local coupling relationships of transmission lines. Third, in broadband S-parameter prediction, directly regressing high-dimensional complex responses across all frequencies often leads to problems such as high output dimensionality, unstable training, and difficulty in maintaining spectral continuity. Fourth, existing technologies do not adequately connect single-transmission-line modeling with system-level interconnect analysis, making it difficult to directly support rapid cascading evaluation of transmission lines with RF modules such as filters and switches.

[0004] Therefore, it is necessary to propose a new method and system for predicting transmission line S-parameters. Based on fully considering the transmission line structure sequence characteristics, spectral continuity, and system cascading requirements, this method enables rapid and accurate prediction from transmission line structure description to broadband S-parameter response, thereby meeting the needs for efficient modeling and automated design evaluation of complex RF interconnect networks. Summary of the Invention

[0005] The purpose of this invention is to provide a transmission line S-parameter prediction method and system based on temporal convolutional networks, addressing the problems of high dependence on full-wave electromagnetic simulation, high sample acquisition costs, low efficiency of broadband S-parameter prediction, and difficulty in supporting rapid cascaded evaluation of complex RF interconnect systems in existing transmission line electrical performance analysis. By fully considering the sequential characteristics, inter-layer switching features, and spectral response continuity of the transmission line structure, this invention introduces structured coding, temporal convolutional modeling, and frequency domain coefficient reconstruction mechanisms to achieve rapid mapping from transmission line structure description to broadband S-parameter response. Furthermore, it supports system-level cascaded analysis of transmission lines with filters, switches, and other RF devices, thereby significantly improving the modeling efficiency and engineering usability of complex RF interconnect networks while ensuring prediction accuracy.

[0006] To achieve the above objectives, this invention provides a method for predicting transmission line S-parameters based on temporal convolutional networks, comprising the following steps: Step 1: Obtain the structural description information of the transmission line to be predicted, and parse the structural description information into a structured sequence containing the starting routing layer, routing events, geometric parameters, and interlayer via events; Step 2: Perform type encoding, geometric parameter encoding, and position encoding on the structured sequence obtained in Step 1 to form a fixed-length sequence input representation; Step 3: Input the fixed-length sequence obtained in Step 2 into the trained temporal convolutional neural network prediction model, and output the frequency domain orthogonal basis coefficients of the complex responses corresponding to each sub-band S11, S21, S12 and S22 within the target frequency range; Step 4: Perform inverse transform on the frequency domain orthogonal basis coefficients from Step 3 to reconstruct the broadband S-parameter response of the transmission line to be predicted within the target frequency range. Step 5: Output the S-parameter prediction results of the transmission line to be predicted.

[0007] Preferably, the structural description information in step one is expressed in the form of serialized instructions; for right-angle corner transmission lines, the structural description information includes a start layer mark, multiple direction routing instructions, and interlayer via instructions, wherein the direction routing instructions are used to represent planar routing segments in different directions, and the interlayer via instructions are used to represent the switching between metal layers; for arc corner transmission lines, the structural description information includes a straight line segment instruction, a left-turn arc instruction, a right-turn arc instruction, and an interlayer via instruction, wherein the parameter of the arc instruction is used to represent the arc radius.

[0008] Preferably, the process of encoding the structured sequence in step two includes: mapping the event type to a discrete type embedding vector; mapping the trace length or arc radius to a continuous geometric feature vector; introducing position encoding for the sequence position; and fusing the type embedding vector, the continuous geometric feature vector, and the position encoding to form a unified fixed-length sequence input representation.

[0009] Preferably, in step two, the geometric parameter encoding adopts Fourier feature encoding, which expands the trace length or arc radius into multi-frequency sine and cosine features, enhancing the model's ability to characterize changes in electrical length, phase advance, resonance shift, and in-band ripple. At the same time, a sequence representation with a fixed maximum length is adopted, and the input of transmission line samples with different complexities is unified through padding and masking mechanisms.

[0010] Preferably, in step three, training data for the temporal convolutional neural network prediction model is generated through an automated electromagnetic modeling and simulation process. Specifically, this includes: the scheduler automatically drives the full-wave electromagnetic simulation software to perform three-dimensional structure construction, port setting, broadband frequency sweeping, and S-parameter export based on the transmission line structure description information, forming a one-to-one correspondence sample dataset between the transmission line structure description and the corresponding broadband S-parameter response.

[0011] Preferably, the training process in step three includes grouping the samples according to the number of transmission line segments and performing frequency band modeling of the target broadband S-parameters according to the frequency range; wherein, samples with fewer segments are designated as low-complexity groups, samples with more segments are designated as high-complexity groups, and corresponding models are trained for different complexity groups respectively; the frequency range is divided into multiple sub-frequency bands, and corresponding sub-models are trained for each sub-frequency band, reducing the difficulty of cross-frequency band regression tasks and improving the prediction accuracy within the frequency band.

[0012] Preferably, the training process further includes an enhancement strategy based on geometric mirror equivalence, which performs a mirror transformation on the original structural description information by swapping left and right-side events, and reuses the S-parameter response of the original samples as supervision labels, thereby expanding the training sample size and improving the model's generalization performance.

[0013] Preferably, the temporal convolutional neural network prediction model adopts a multi-layer one-dimensional dilated convolutional structure, with the dilation rate of the convolutional layers set in an incremental manner, and performs multi-scale joint modeling of the coupling relationship between local adjacent line segments and the overall path relationship across multiple segments; after the convolutional feature extraction, the model uses an attention pooling mechanism to weighted converge the effective sequence positions.

[0014] Preferably, in step three, the frequency domain orthogonal basis coefficients are represented by low-order coefficients corresponding to the discrete cosine transform. The time-series convolutional neural network prediction model outputs low-dimensional coefficient vectors of the real and imaginary parts of S11, S21, S12, and S22 in each sub-frequency band. After inverse normalization and orthogonal basis reconstruction, a continuous and smooth broadband S-parameter response is obtained. Furthermore, in system-level applications, the predicted transmission line S-parameters are cascaded with the S-parameter models of filters, RF switches, packaged inductors, packaged capacitors, or other RF devices to obtain the overall S-parameter response of the RF interconnect network.

[0015] A transmission line S-parameter prediction system based on temporal convolutional networks, applying the transmission line S-parameter prediction method based on temporal convolutional networks as described in any one of the above descriptions, includes: a structure parsing module, an input encoding module, an S-parameter prediction module, a response reconstruction module, and a system-level evaluation module; The structure parsing module is used to obtain the structural description information of the transmission line to be predicted and parse it into a structured sequence containing the starting trace layer, trace events, geometric parameters, and interlayer via events. Input encoding module: used to perform type encoding, geometric parameter encoding and position encoding on the structured sequence to form a fixed-length sequence input representation; S-parameter prediction module: used to input the fixed-length sequence input representation into the trained temporal convolutional neural network prediction model, and output the frequency domain orthogonal basis coefficients corresponding to each sub-frequency band; Response reconstruction module: used to perform inverse transformation and reconstruction on the frequency domain orthogonal basis coefficients to obtain the transmission line broadband S-parameter response within the target frequency range; System-level evaluation module: used to cascade the predicted transmission line S-parameters with known RF device S-parameter models to obtain the overall response results of the RF network including transmission line interconnections.

[0016] Therefore, the beneficial effects of the transmission line S-parameter prediction method and system based on temporal convolutional networks described above are as follows: (1) This invention constructs an end-to-end S-parameter prediction framework for complex transmission line structures. By uniformly representing multi-layer routing, corner changes and inter-layer via switching as a structured sequence, and combining type encoding, geometric parameter encoding and position encoding, the mapping relationship between transmission line geometric topology and broadband S-parameter response can be stably learned, thereby effectively solving the problem of difficult unified modeling of complex transmission lines in traditional methods. (2) This invention uses a temporal convolutional neural network to extract multi-scale features from transmission line sequences and captures both local structural discontinuity effects and overall path accumulation effects through dilated convolution mechanism. Compared with the traditional analysis process that relies on a large number of full-wave simulations, it can significantly reduce computational costs and improve broadband S-parameter prediction efficiency, and is suitable for large-scale sample inference and rapid design evaluation. (3) The present invention further adopts a frequency domain orthogonal basis coefficient compression and reconstruction mechanism to transform the prediction of high-dimensional broadband S-parameter curves into a low-dimensional coefficient regression problem. While maintaining the continuity of the spectrum and physical smoothness, the output dimension is reduced, and the model training stability and cross-structure generalization ability are improved. (4) This invention can not only achieve fast S-parameter prediction of a single transmission line, but also cascade the predicted transmission line S-parameters with the known S-parameter models of filters, switches and other RF devices, thereby supporting system-level response analysis of complex RF interconnect networks. It has good modular expansion capability and engineering practical value. (5) This invention is applicable to radio frequency transmission line path modeling in multi-layer dielectric and multi-metal layer interconnect scenarios, and can provide an efficient and reliable technical approach for rapid electrical performance prediction, automated design screening and system-level collaborative analysis in high-speed interconnect of radio frequency front-end modules, package interconnect structures and printed circuit boards. The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the transmission line structure encoding form in an embodiment of the present invention, wherein (a) is a serialized encoding form diagram of a right-angle turn transmission line, and (b) is a serialized encoding form diagram of a 45-degree circular arc turn transmission line; Figure 2 This is a schematic diagram of the grouping and frequency band training strategy in an embodiment of the present invention; Figure 3 This is a schematic diagram showing that the S-parameter performance remains unchanged after the transmission line is mirrored in an embodiment of the present invention; Figure 4 This is a schematic diagram of the transmission line input encoding compression method in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the compression of S-parameters through frequency domain orthogonal bases in an embodiment of the present invention; Figure 6 This is a schematic diagram showing different expansion rates when the convolution kernel is fixed in an embodiment of the present invention; Figure 7 This is a schematic diagram of the temporal convolutional neural network architecture in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the system-level S-parameter prediction verification of a right-angle transmission line with added components in an embodiment of the present invention; Figure 9 The above is a comparison of the prediction results of the right-angle turn transmission line cascaded filter and the switch in the embodiment of the present invention, where (a) is the S-parameter amplitude-frequency characteristic curve and (b) is the Smith chart. Figure 10 This is a schematic diagram of a cascaded filter for transmission lines with radian angles in an embodiment of the present invention; Figure 11 This is a comparison chart of prediction results for the cascaded filter of the arc-angle transmission line in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0019] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0020] The following is combined with Figures 1-11 The embodiments of the present invention will be described in detail below.

[0021] Example 1 In one specific embodiment of the present invention, the design object is an interconnect transmission line structure in a radio frequency (RF) and microwave system. The transmission line is widely used for electrical connections between filters, switches, combiners, and RF front-end modules. Its structure can be multiple planar traces on a multilayer dielectric substrate, or it can include interlayer vias, corner transitions, and variations in geometric dimensions. To achieve rapid prediction of the broadband S-parameters of this type of interconnect structure, this embodiment constructs a transmission line S-parameter prediction process based on structured sequence coding, frequency domain coefficient compression, and time-series convolutional neural network modeling.

[0022] like Figure 1 As shown, this embodiment illustrates the unified structural representation of different transmission line topologies in this invention. First, the transmission line structure is represented by a unified encoding. Figure 1 (a) represents the serialized encoding form of a right-angle transmission line, where A, B, C, and D represent traces in different directions, E and F represent vias from the top layer to the inner layer and from the inner layer to the top layer, respectively, M1 represents the first metal layer, and so on. For right-angle transmission lines, a serialized representation consisting of a start layer marker, direction trace events, and interlayer via events is used. The start layer marker indicates that the transmission line starts routing from the first or second metal layer; direction trace events represent planar trace segments in different directions, with parameters appended to the events indicating the trace length; and interlayer via events represent the interlayer interconnection behavior when switching from one metal layer to another. In this way, the overall transmission line topology can be described by an ordered sequence formed by sequentially piecing together multiple local trace operations.

[0023] Figure 1 (b) represents the serialized encoding format for a 45-degree arc-corner transmission line, where S, L, and R represent a straight line, a left turn, and a right turn, respectively, with the corner angle fixed at 45 degrees. For 45-degree arc-corner transmission lines, serialization instructions are also used. Unlike right-angle corner scenarios, its basic event set includes straight line segment events, left arc-corner events, right arc-corner events, and interlayer via events. The parameter following the straight line segment event represents the straight line length, the parameter following the arc-corner event represents the arc radius, and the interlayer via event describes metal layer switching. Since arc-corner configurations are more common in high-speed interconnects and high-frequency routing, they can mitigate local impedance abrupt changes caused by geometric discontinuities. Therefore, this encoding method makes the present invention applicable not only to right-angle corner routing but also to arc-smooth transition scenarios.

[0024] In this embodiment, to ensure the uniformity and high traceability of training samples, transmission line sample data is generated through an automated electromagnetic modeling and simulation process. Specifically, MATLAB can be used as the scheduling endpoint, calling full-wave electromagnetic simulation software via an interface to automatically complete the establishment of the three-dimensional trace structure, setting of the dielectric layer, definition of the metal layer, port loading, broadband frequency sweeping, and S-parameter export according to the aforementioned serialized instructions. This forms a one-to-one correspondence sample dataset between the transmission line structure description and the corresponding broadband S-parameters. This automated sample generation method significantly reduces the cost of manual modeling and provides highly consistent supervisory data for subsequent data-driven training.

[0025] like Figure 2 The diagram illustrates the process of grouping sample complexity according to the number of transmission line segments and training corresponding sub-models by dividing the target frequency range into multiple sub-bands. After obtaining the sample data, this embodiment employs a joint training strategy of grouping and frequency band division. First, the sample complexity is grouped according to the number of effective routing segments in the transmission line instruction sequence. Samples with no more than three segments are assigned to the low complexity group, while samples with more than three segments are assigned to the high complexity group. This grouping method makes the samples within the same group more similar in topological complexity and path scale, thereby reducing fitting conflicts when a single model simultaneously considers both simple and complex topologies.

[0026] On the other hand, this embodiment also performs frequency-band modeling of the broadband S-parameters. Taking the broadband frequency range of 0.1 GHz to 6 GHz as an example, it is divided into multiple sub-bands, and corresponding sub-models are trained for each. The low-frequency band mainly corresponds to quasi-static and weak resonance behavior, the mid-frequency band gradually reflects propagation phase accumulation and impedance perturbation, and the high-frequency band is more prone to dense ripples and local resonances. Through this frequency-band training method, the high-dimensional regression task with a large span across frequency bands can be decomposed into multiple sub-tasks with more consistent spectral shapes, thereby improving the fitting accuracy and training stability within each frequency band.

[0027] like Figure 3 The diagram illustrates a training data augmentation strategy based on planar mirror equivalence. This embodiment further introduces a mirror data augmentation strategy. Since the electromagnetic response of some transmission line structures remains unchanged after in-plane mirror transformation when material parameters, linewidth, thickness, and boundary conditions are kept consistent, the left and right directional events in the original command sequence can be interchanged to obtain a new sample after geometric mirroring, and the S-parameters of the original sample can be reused as supervision labels. In this way, the training sample size can be expanded without increasing the overhead of full-wave simulation, thereby improving the model's learning ability and generalization ability to structural command expressions.

[0028] like Figure 4The diagram illustrates the process of parsing the initial layer, splitting events, padding with fixed length, and masking the original transmission line commands. In the input encoding stage, this embodiment first normalizes the original command string, identifies the initial layer marker, and maps it to an initial layer event, writing it at the beginning of the sequence. Subsequently, the command body is parsed segment by segment, representing each routing segment or inter-layer interconnection event as an event unit composed of event type and parameter value. For right-angle transmission lines, routing events correspond to different directions and their lengths; for arc-shaped transmission lines, routing events correspond to straight lines, left arcs, right arcs, and their corresponding lengths or radii. All events are written into the sequence in the order of appearance to preserve topological order information.

[0029] To ensure consistent input dimensions across different samples, this embodiment employs a fixed-length sequence representation. Samples shorter than the specified length are padded with events, and a mask matrix is ​​constructed to mark the positions of the real events and the padded events. This approach supports batch training and efficient parallel computation while preventing the model from misinterpreting the padded content as the real structure. Subsequently, event types are mapped to discrete embedding vectors, geometric parameters are mapped to continuous feature vectors, and these are then added to and fused with positional encodings to form a unified input representation.

[0030] In terms of numerical parameter representation, this embodiment does not directly input length or radius as a single scalar into the network. Instead, it employs Fourier feature encoding to expand continuous geometric parameters into multi-frequency sine and cosine features. This approach better expresses the impact of length variations on phase advance, resonance shift, and in-band ripple, making it easier for the model to learn the broadband response differences corresponding to variations in different geometric scales. Ultimately, the transmission line input consists of three parts: type encoding, geometric parameter encoding, and position encoding, thus forming a unified sequence representation that includes both discrete topological semantics and continuous geometric scale information.

[0031] like Figure 5 The diagram illustrates the process of compressing a broadband S-parameter curve into low-dimensional frequency domain coefficients and reconstructing it using an inverse transform. At the output, this embodiment does not directly regress the complex S-parameter sequence across all frequencies; instead, it employs a compressed representation based on a frequency domain orthogonal basis. Specifically, the real and imaginary parts of the S11, S21, S12, and S22 curves within each sub-band are projected onto an orthogonal cosine transform basis, retaining only the first few low-order basis coefficients as the dominant representation coefficients for that frequency band. In this way, the originally high-dimensional broadband curve prediction problem is transformed into a low-dimensional coefficient regression problem, thereby reducing the output dimensionality and improving training stability and sample utilization efficiency.

[0032] During the inference phase, the model outputs low-dimensional coefficient vectors within each sub-band. After inverse normalization, these vectors are multiplied by the corresponding orthogonal basis matrices to reconstruct the complete real and imaginary curves of S11, S21, S12, and S22. Since the output curves are linearly reconstructed using fixed orthogonal bases, they naturally satisfy certain smoothness and continuity constraints, which helps improve the physical plausibility of the predicted curves and suppress high-frequency numerical noise.

[0033] like Figure 6 and Figure 7 As shown, Figure 6 This is used to illustrate a multi-scale modeling mechanism that expands the receptive field by increasing the dilation rate while keeping the kernel size constant. Figure 7 This is used to illustrate the overall relationship between input representation, temporal convolutional feature extraction, attention pooling, and frequency domain coefficient output.

[0034] This embodiment employs a temporal convolutional neural network as the core prediction model. The model first receives the structured input representation described above and then extracts features through multiple temporal convolutional modules. Each convolutional module uses a fixed kernel size and achieves rapid expansion of the receptive field through progressively increasing dilation rates. Taking a kernel size of 3 as an example, the dilation rates can be set sequentially to 1, 2, 4, 8, and 16, enabling the model to simultaneously capture local coupling relationships between adjacent path segments and global topological relationships across multiple paths at a relatively shallow number of layers.

[0035] Within each temporal convolutional module, two concatenated one-dimensional convolutional layers can be used, followed by a normalization layer, a non-linear activation layer, and a random deactivation layer in sequence to suppress overfitting and improve training stability. To achieve a stable deep optimization path, the output of the convolutional module can be added to the input residual. After feature extraction via multi-layer dilated convolutions, the model further uses attention pooling to weighted convergence at effective sequence locations, highlighting key structural segments that have a more significant impact on S-parameters, such as corners, abrupt length changes, and local regions near via switching points. The pooled global representation is then output via a feedforward mapping head, yielding the predicted frequency domain orthogonal basis coefficients for each sub-band.

[0036] In this embodiment, the model can adopt the following training configuration: input embedding dimension of 192, geometric parameter Fourier encoding dimension of 48, temporal convolution channel number of 256, number of convolutional layers of 5, convolutional kernel size of 3, dilation rate increasing sequentially, and random deactivation ratio set to 0.15. The optimizer can use AdamW, combined with weight decay, warm-up, and cosine decay learning rate scheduling to achieve a balance between convergence speed and generalization ability. For different sub-frequency bands, different numbers of orthogonal basis retention orders can be set to adapt to the differences in curve complexity across different frequency bands.

[0037] like Figure 8The diagram illustrates a verification method for cascading the predicted transmission line S-parameters with known S-parameter models of filters, packaged inductors, packaged capacitors, and RF switches to obtain the overall response of a complex interconnect network. After predicting the S-parameters of a single transmission line, this embodiment further applies the predicted transmission line S-parameters to system-level cascade analysis. Specifically, several transmission line S2p networks predicted by the model of this invention can be connected with the S3p files of filters, the SNP files of packaged inductors and packaged capacitors, and the equivalent S-parameter models of RF switches in a circuit-level simulation environment to form a more complex RF interconnect system and obtain the overall multi-port S-parameter response of the system. This method avoids the high-cost full-wave modeling required for every interconnect combination of the overall system, thereby significantly improving the evaluation efficiency of complex systems.

[0038] In a specific example, a system consisting of six transmission lines cascaded with filters and capacitors / inductors can be constructed first, with packaged inductors and capacitors introduced for impedance adjustment. The S-parameters of each transmission line are independently predicted using the method of this invention, and then cascaded segment by segment with the device model to obtain the overall three-port response. Furthermore, two RF switches and additional transmission line paths can be added to form a more complex multi-path interconnect system. This extended verification method is used to test the applicability of the model of this invention under conditions of increased system scale, more complex topologies, and deeper device cascading levels.

[0039] like Figure 9 (a) and (b) illustrate the predictive performance of the method of the present invention on the overall frequency response in more complex system-level topologies. In an extended cascaded system containing filters and switches, the overall S-parameter results predicted and cascaded based on the method of the present invention maintain good consistency with the full-wave cascaded reference results in terms of passband position, stopband depth, resonant point position, and impedance trajectory shape. This indicates that the present invention can not only quickly predict the broadband S-parameters of a single transmission line, but also effectively approximate the overall response of more complex RF systems through modular cascading, demonstrating good system-level portability and engineering practical value.

[0040] Example 2 In another specific embodiment of the present invention, the transmission line adopts a 45-degree rounded corner. Compared with the right-angled corner transmission line in Embodiment 1, this type of structure is closer to the common wiring form in actual PCB and package interconnection. The reason is that the rounded transition can, to a certain extent, reduce the local impedance change caused by geometric discontinuity compared with a sharp right angle, and reduce the additional reflection and parasitic radiation of high-frequency signals at the corner. Therefore, it has higher engineering application value in high-speed interconnection and high-frequency routing scenarios.

[0041] In this embodiment, the structural description of the arc-corner transmission line still adopts a serialized event representation. Its input format remains consistent with Embodiment 1, uniformly represented as a structured sequence consisting of the starting layer, routing action, geometric parameters, and interlayer via events. The difference lies in that the event set is adjusted to straight line segment events, left-turn arc events, right-turn arc events, and interlayer via events. The parameter following the straight line segment indicates the trace length, while the parameters following the left-turn and right-turn arc events indicate the arc radius. The interlayer via event represents the interconnect behavior of switching from one metal layer to another. Through this encoding method, the geometric path of the arc-corner transmission line can be uniformly represented as a sequence structure composed of straight line segments, left arc segments, right arc segments, and interlayer vias, thereby introducing a geometric prior that more closely resembles the actual PCB routing form while maintaining the simplicity of the input expression.

[0042] Since the data format of the 45-degree circular arc corner transmission line is consistent with that of the right-angle corner transmission line, there is no need to redesign the overall network framework in the prediction model construction. The structured input encoding method, temporal convolution backbone, multi-scale dilated convolution feature extraction mechanism, and frequency band compression and reconstruction strategy at the output end described in Example 1 are still used. Only some neural network hyperparameters need to be adaptively adjusted based on the geometric complexity and parameter distribution characteristics of the circular arc corner structure. Compared to the right-angle corner transmission line, the 45-degree circular arc corner structure is affected not only by the trace length but also by the continuous geometric degree of freedom of the arc radius. Therefore, its design space is larger, structural variations are more detailed, and response distribution is richer. To ensure that the model can fully learn the influence of arc radius changes on broadband S-parameters, this example further expands the data accumulation scale, constructs a 45-degree circular arc corner transmission line sample dataset, and uses this dataset to complete model training. Training results show that the convergence of training loss and the comparison between predicted S-parameters and full-wave simulation results for this type of structure are close to those in the right-angle transmission line scenario, indicating that the temporal convolution modeling framework adopted in this invention has good adaptability and transferability to different corner forms.

[0043] like Figure 10The diagram illustrates a system-level verification structure formed by cascading a filter with three relatively complex 45-degree arc-angle transmission lines. The three transmission lines connect different ports of the filter to external interfaces and include straight segments, left and right arc transitions, and interlayer metal vias. To further verify the generalization ability of the 45-degree arc-angle transmission line prediction model in system-level scenarios, this embodiment uses a filter cascaded with three relatively complex arc-angle transmission lines to construct a verification structure closer to the actual PCB interconnection configuration. The three transmission lines connect different ports of the filter to external interfaces and include straight segments, left and right arc transitions, and interlayer metal vias, thus enabling a more comprehensive examination of the model's ability to preserve broadband S-parameters under complex path conditions. Similar to the right-angle transmission line system-level verification in Embodiment 1, this embodiment also uses a system-level cascading method to introduce the transmission line response predicted by the surrogate model into the filter network and compare it with the full-wave simulation results to evaluate the model's applicability in actual module interconnection environments.

[0044] It should be particularly pointed out that, Figure 10 The third transmission line in the verification structure shown contains two novel topological configurations that did not explicitly appear during the data accumulation phase. The first is a straight segment directly connected to a via, which continues propagating in a straight line after layer switching. The second is a structure where a via is immediately followed by a circular arc corner. Neither of these connection patterns appeared in the training data, therefore they are unseen combinations for the surrogate model. Predicting these two types of transmission lines is significant because it not only tests the model's ability to fit known geometric patterns but also its ability to extrapolate and generalize to novel topological combinations. If the model can still accurately recover the overall S-parameter response under these conditions, it indicates that the sequence encoding and temporal convolution modeling method constructed in this invention captures not local template memory, but rather the more fundamental mapping between the transmission line structure and the electromagnetic response.

[0045] In the system-level verification process, two paths can be used to obtain the overall response results. The first path uses the full-wave simulation cascade results as a reference benchmark. That is, the multi-port S-parameter files of the filter and interconnect network are obtained from the full-wave simulation, and the network connection is completed in the circuit-level environment to obtain the system-level S-parameter response. The second path uses the S-parameters of the arc-angle transmission line predicted by this invention as the interconnect model. While keeping the filter module unchanged, the transmission line S-parameter results predicted by the surrogate model are connected to the ports to construct a system-level network equivalent to the reference benchmark. By comparing the results obtained from the two paths, the system-level application capability of this invention in the arc-angle transmission line scenario can be verified.

[0046] like Figure 11The diagram illustrates the comparison between system-level S-parameter results obtained from full-wave simulation cascading and those obtained from cascading transmission line predictions based on surrogate models. The comparison shows that the results obtained from full-wave simulation cascading and cascading transmission line predictions based on surrogate models maintain a high degree of consistency across the frequency band. The passband position, stopband rejection depth, and resonant concave frequency of the system-level S-parameter curves are basically aligned, especially in the main variation region around 1.8 GHz to 2.1 GHz, where the prediction results can effectively track the variation trend of the full-wave results. Although some amplitude differences still exist near individual sharp notches, and high-frequency local details are slightly offset, these deviations are generally small and do not change the performance judgment within the main operating frequency band.

[0047] Furthermore, from Figure 11 The impedance trajectory shown in the Smith chart also demonstrates that the predicted results and the full-wave results maintain a good overlap in trajectory direction, rotation trend, and main loop morphology. This indicates that even when the system contains circular arc corners, interlayer vias, and complex path connections, the model of this invention can still stably recover port impedance changes and the overall energy coupling path. Figure 11 The results show that the 45-degree circular arc angle transmission line proxy model constructed in this invention can not only be used for fast S-parameter prediction of a single transmission line, but also participate in system-level cascade prediction together with the filter module, providing support for the rapid evaluation and design verification of complex PCB interconnect networks.

[0048] Overall, this embodiment demonstrates that the present invention still exhibits good generalization ability and engineering applicability in the scenario of a 45-degree circular arc corner transmission line. Even when the verification structure contains local connection methods not explicitly appearing in the training data, such as combinations like straight-line through-hole followed by continued straight-line propagation and through-hole after a corner, the model can still maintain relatively accurate S-parameter prediction results at the overall network level. This indicates that the present invention learns not only the surface correspondence of a fixed topology template, but also a relatively stable mapping law between the geometric changes of the circular arc corner transmission line and its electromagnetic response, thus possessing good modular expansion capabilities and system-level application potential.

[0049] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0050] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0051] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0052] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting transmission line S-parameters based on temporal convolutional networks, characterized in that, Includes the following steps: Step 1: Obtain the structural description information of the transmission line to be predicted, and parse the structural description information into a structured sequence containing the starting routing layer, routing events, geometric parameters, and interlayer via events; Step 2: Perform type encoding, geometric parameter encoding, and position encoding on the structured sequence obtained in Step 1 to form a fixed-length sequence input representation; Step 3: Input the fixed-length sequence obtained in Step 2 into the trained temporal convolutional neural network prediction model, and output the frequency domain orthogonal basis coefficients of the complex responses corresponding to each sub-band S11, S21, S12 and S22 within the target frequency range; Step 4: Perform inverse transform on the frequency domain orthogonal basis coefficients from Step 3 to reconstruct the broadband S-parameter response of the transmission line to be predicted within the target frequency range. Step 5: Output the S-parameter prediction results of the transmission line to be predicted.

2. The method for predicting transmission line S-parameters based on temporal convolutional networks according to claim 1, characterized in that, In step one, the structural description information is expressed in the form of serialized instructions. For right-angle corner transmission lines, the structural description information includes a start layer marker, multiple direction routing instructions, and interlayer via instructions. The direction routing instructions are used to represent planar routing segments in different directions, and the interlayer via instructions are used to represent the switching between metal layers. For arc corner transmission lines, the structural description information includes a straight line segment instruction, a left-turn arc instruction, a right-turn arc instruction, and an interlayer via instruction. The parameter of the arc instruction is used to represent the arc radius.

3. The method for predicting transmission line S-parameters based on temporal convolutional networks according to claim 2, characterized in that, Step two involves encoding the structured sequence by: mapping event types to discrete type embedding vectors; mapping trace lengths or arc radii to continuous geometric feature vectors; introducing position encoding for sequence positions; and fusing the type embedding vectors, continuous geometric feature vectors, and position encoding to form a unified fixed-length sequence input representation.

4. The method for predicting transmission line S-parameters based on temporal convolutional networks according to claim 3, characterized in that, In step two, the geometric parameter encoding adopts Fourier feature encoding, which expands the trace length or arc radius into multi-frequency sine and cosine features to enhance the model's ability to represent changes in electrical length, phase advance, resonance shift, and in-band ripple. At the same time, a sequence representation with a fixed maximum length is adopted, and the input of transmission line samples with different complexities is unified through padding and masking mechanisms.

5. The method for predicting transmission line S-parameters based on temporal convolutional networks according to claim 4, characterized in that, In step three, training data for the temporal convolutional neural network prediction model is generated through an automated electromagnetic modeling and simulation process. Specifically, this includes: the scheduler automatically drives the full-wave electromagnetic simulation software to build the three-dimensional structure, set the ports, perform broadband frequency sweeping, and export the S-parameters based on the transmission line structure description information, forming a one-to-one correspondence sample dataset between the transmission line structure description and the corresponding broadband S-parameter response.

6. The method for predicting transmission line S-parameters based on temporal convolutional networks according to claim 5, characterized in that, The training process in step three includes grouping the samples according to the number of transmission line segments and performing frequency band modeling of the target broadband S-parameters according to the frequency range. Among them, samples with fewer segments are classified as low-complexity groups, and samples with more segments are classified as high-complexity groups. Different complexity groups train corresponding models respectively. The frequency range is divided into multiple sub-bands, and each sub-band trains a corresponding sub-model, which reduces the difficulty of cross-band regression tasks and improves the prediction accuracy within the frequency band.

7. The method for predicting transmission line S-parameters based on temporal convolutional networks according to claim 6, characterized in that, The training process further includes an enhancement strategy based on geometric mirror equivalence, which performs a mirror transformation on the original structural description information by swapping left and right events, and reuses the S-parameter response of the original samples as supervision labels to expand the training sample size and improve the model's generalization performance.

8. The method for predicting transmission line S-parameters based on temporal convolutional networks according to claim 7, characterized in that, The temporal convolutional neural network prediction model adopts a multi-layer one-dimensional dilated convolutional structure, with the dilation rate of the convolutional layers set in an incremental manner, and performs multi-scale joint modeling of the coupling relationship between local adjacent line segments and the overall path relationship across multiple segments; after the convolutional feature extraction, the model uses an attention pooling mechanism to weighted converge the effective sequence positions.

9. The method for predicting transmission line S-parameters based on temporal convolutional networks according to claim 8, characterized in that, In step three, the frequency domain orthogonal basis coefficients are represented by low-order coefficients corresponding to the discrete cosine transform. The time-series convolutional neural network prediction model outputs low-dimensional coefficient vectors of the real and imaginary parts of S11, S21, S12, and S22 in each sub-frequency band. After inverse normalization and orthogonal basis reconstruction, a continuous and smooth broadband S-parameter response is obtained. Furthermore, in system-level applications, the predicted transmission line S-parameters are cascaded with the S-parameter models of filters, RF switches, packaged inductors, packaged capacitors, or other RF devices to obtain the overall S-parameter response of the RF interconnect network.

10. A transmission line S-parameter prediction system based on temporal convolutional networks, characterized in that, The method for predicting transmission line S-parameters based on temporal convolutional networks as described in any one of claims 1-9 includes: a structure parsing module, an input encoding module, an S-parameter prediction module, a response reconstruction module, and a system-level evaluation module; The structure parsing module is used to obtain the structural description information of the transmission line to be predicted and parse it into a structured sequence containing the starting trace layer, trace events, geometric parameters, and interlayer via events. Input encoding module: used to perform type encoding, geometric parameter encoding and position encoding on the structured sequence to form a fixed-length sequence input representation; S-parameter prediction module: used to input the fixed-length sequence input representation into the trained temporal convolutional neural network prediction model, and output the frequency domain orthogonal basis coefficients corresponding to each sub-frequency band; Response reconstruction module: used to perform inverse transformation and reconstruction on the frequency domain orthogonal basis coefficients to obtain the transmission line broadband S-parameter response within the target frequency range; System-level evaluation module: used to cascade the predicted transmission line S-parameters with known RF device S-parameter models to obtain the overall response results of the RF network including transmission line interconnections.

Images