A gain-gated printed circuit board routing priori on-demand activation method and system

Abstract

The application discloses a printed circuit board wiring priori on-demand enabling method and system based on gain gating, and relates to the field of automatic circuit board wiring. The method obtains design data and wiring results of multiple printed circuit boards, extracts network-level geometric features and constructs network-level discriminant labels, respectively evaluates the discriminability of a single geometric feature baseline and the discriminability of a multi-feature learning model under a board-level evaluation aperture, calculates the predictability gain of the learning model relative to the baseline, and generates a gating indicator according to the comparison result of the gain and a preset threshold; when the gain meets the standard, the learning priori is enabled, and when the gain is insufficient, the learning priori is blocked and the optimal single-feature baseline is returned to generate a same single-feature priori, and the selected priori is injected into a router to modulate wiring costs and perform wiring. Through the board-level same-aperture evaluation and gain gating mechanism, the application realizes the on-demand enabling of the learning priori, guarantees the wiring guidance effect, and improves the cross-board generalization reliability and resource utilization efficiency.

Description

Technical Field

[0001] This invention belongs to the field of automatic circuit board routing technology, and particularly relates to a method and system for enabling prior routing on demand for printed circuit boards based on gain gating. Background Technology

[0002] In the automated routing process of printed circuit boards (PCBs), the router typically needs to determine network ordering, cost weights, avoidance areas, or routing priorities based on information such as network geometry, pin distribution, local congestion, obstacles, and board resources. Existing solutions often use the results of routerability prediction, congestion prediction, or residual network prediction as routing priors. With the development of machine learning methods, multi-feature learning models are further used to predict network residual probabilities, congestion risks, or routerability scores before directly inputting them into the downstream routing process. However, in actual engineering, not all designs require complex learning models to obtain effective priors. In some designs, single, inexpensive geometric features such as the number of network pins, half-perimeter line length, bounding box area, local pin density, or simplified congestion indicators can already effectively distinguish between difficult-to-route and easy-to-route networks. If complex models are still trained and enabled by default, it will increase the costs of training, inference, data maintenance, and deployment, and may lead to prior instability due to insufficient sample size or differences in board distribution. However, if a fixed single-feature heuristic is relied upon entirely, it is difficult to fully utilize its discriminative power on a design set where multi-feature models have significant gains.

[0003] Furthermore, multiple networks within the same PCB board typically share device distribution, layer constraints, local channel resources, and congestion patterns. If similar networks from the same board are simultaneously included in the training and validation sets when evaluating complex learning models, it's easy to obtain artificially high discriminative power that doesn't represent the true generalization ability of the new board-level design. Existing technologies focus primarily on improving routeability or congestion prediction accuracy and embedding prediction results into the routing process. They typically default to using a certain prediction model or fixed heuristic, lacking a gating mechanism to determine whether the complex learning model has sufficient discriminative gain compared to a relatively inexpensive single-feature baseline before prior generation. They also lack a unified board-level evaluation caliber and fallback path. Therefore, existing processes struggle to automatically select learning priors or single-feature priors across different board-level designs, making it difficult to balance routing guidance effectiveness, cross-board generalization reliability, and resource efficiency. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a gain-gated printed circuit board routing prior-on-demand activation method, comprising: Obtain design data and routing results for multiple printed circuit boards; Based on the design data, extract the network-level geometric features of each network, and construct network-level discrimination labels for each network based on the wiring results; Based on the network-level geometric features and the network-level discriminant labels, the discriminant power of at least one single geometric feature baseline and the discriminant power of the multi-feature learning model are evaluated under the board-level evaluation caliber, respectively, and the baseline discriminant power and the learning discriminant power are obtained accordingly. Calculate the predictability gain based on the learned discriminant power and the baseline discriminant power; Based on the comparison result between the predictability gain and the preset threshold, a gating indicator is generated; Based on the gating indication, a learning prior or a single-feature prior is selected as the prior to be injected; wherein, when the predictability gain does not reach the preset threshold, the learning prior is blocked and the system reverts to the optimal single-feature baseline to generate the single-feature prior that is isomorphic to the learning prior; The prior to be injected is injected into the router. The prior to be injected is used to generate network sorting priority, modulate routing cost, generate congestion avoidance cues, or configure search budget to guide the router to perform printed circuit board routing.

[0005] Optionally, acquiring the design data and wiring results of multiple printed circuit boards includes: Obtain the netlist, pin coordinates, layer information, prohibited routing areas, obstacles, and routing rules of the printed circuit board; Obtain the wiring results of the printed circuit board, including historical wiring results, trial wiring output results, and intermediate wiring results in the electronic design automation pipeline.

[0006] Optionally, network-level geometric features of each network are extracted based on the design data, including: The network, pins, bounding boxes, local neighborhoods, obstacles, and existing resource distribution are analyzed from the design data to extract the inexpensive geometric features of each network and form a network-level feature vector.

[0007] Optionally, the process of obtaining the baseline discriminative power and the learned discriminative power includes: At least one candidate single feature is selected from the network-level geometric features. The forward ranking discriminant power and the reverse ranking discriminant power of each candidate single feature are calculated under the board-level evaluation caliber. The better one is taken as the discriminant power of the candidate single feature, and the optimal value among the discriminant powers of each candidate single feature is taken as the baseline discriminant power. The learned discriminative power is obtained by predicting the network-level discriminative label using the multi-feature learning model under the same board-level evaluation caliber.

[0008] Optionally, the board-level evaluation criteria include leave-one-board cross-validation and dividing the training set and validation set by board, wherein networks within the same board do not appear simultaneously in the training set and validation set.

[0009] Optionally, obtaining the learned discriminative power includes: The leave-one-board cross-validation method is adopted, in which one board is used as the validation set and the remaining boards are used as the training set. On each of the validation sets, the multi-feature learning model trained on the training set is used to make predictions, and the discriminative power on that validation set is calculated. The average of the discriminative power on all validation sets is calculated as the learned discriminative power.

[0010] Optionally, based on the gating indication, a learned prior or a single-feature prior is selected as the prior to be injected, including: When the predictability gain reaches the preset threshold, the network risk score is output by the multi-feature learning model as the learning prior. When the predictability gain does not reach the preset threshold, the learning prior is blocked from entering the router, and the single feature prior, which is isomorphic to the learning prior, is generated by the optimal single feature baseline and its direction.

[0011] Optionally, the prior to be injected into the injector is modulated according to the prior to be injected, including: The prior to be injected is converted into network prior scores; Obtain the base cost of the router; Based on the network prior score and the preset prior injection coefficient, a prior modulation term is generated; The sum of the base cost and the prior modulation term is used as the modulated wiring cost.

[0012] Optionally, performing printed circuit board routing based on the modulated routing cost includes: Perform global routing, detailed routing, or sequential routing based on the modulated routing cost; Record the routing results and feed them back to the label construction stage for subsequent board-level evaluation or model updates.

[0013] This invention also provides a gain-gated printed circuit board routing prior-on-demand activation system, comprising: The network feature extraction module is used to extract network-level geometric features from printed circuit board design data; The tag construction module is used to construct network-level discrimination tags based on the wiring results. The single-feature baseline evaluation module is used to evaluate the discriminative power of at least one single geometric feature baseline under the plate-level evaluation caliber, and obtain the baseline discriminative power; The learning model evaluation module is used to evaluate the discriminative power of the multi-feature learning model under the same board-level evaluation caliber, and obtain the learning discriminative power. The gain calculation and gating module is used to calculate the predictability gain based on the learned discriminative power and the baseline discriminative power, and generate a gating indication based on the comparison result of the predictability gain and a preset threshold. The learning prior generation module is used to generate a learning prior based on the gating indication when the predictability gain reaches the preset threshold. A single-feature prior generation module is used to block the learning prior and generate a single-feature prior that is identical to the learning prior when the predictability gain does not reach the preset threshold, based on the gating indication. The prior injection module is used to inject the learned prior or the single-feature prior into the router through a unified interface, and modulate the routing cost according to the injected prior. A routing execution module is used to perform printed circuit board routing according to the modulated routing cost.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention implements board-level predictability gain gating before routing priors enter the router. It compares the learned model with a low-cost single-feature baseline using a same-diameter board-level evaluation. The learning prior is only activated when the learned model has sufficient cross-board discriminative gain, thus avoiding inflated evaluations caused by same-board sample leakage and unnecessary overhead from default activation of complex models. When the gain is insufficient, the system blocks the learning prior and reverts to a single-feature prior of the same form. This prior continues to act on network ranking, cost modulation, congestion avoidance, or search budget through the same injection interface, allowing the prior selection result to directly change the router behavior. Therefore, this invention forms a complete technical chain of board-level same-diameter evaluation, gain gating, same-form backoff, and unified injection, focusing on the on-demand activation and backoff control mechanism of the prior, without relying on specific learning models or specific router types. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the gain-gated PCB routing prior on-demand activation method according to an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the single-feature baseline and learning model board-level evaluation process in an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of the gain gating and dual prior selection process in an embodiment of the present invention.

[0018] Figure 4This is a schematic diagram illustrating the process of injecting a priori router and generating routing result feedback in an embodiment of the present invention.

[0019] Figure 5 This is a block diagram of the PCB routing prior on-demand activation system architecture based on gain gating, according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 This embodiment provides a gain-gated printed circuit board routing prior-on-demand activation method, including: Acquire design data and routing results for multiple printed circuit boards; design data includes netlist, pin coordinates, layer information, prohibited routing areas, obstacles, and routing rules; routing results include historical routing results, trial routing output results, or intermediate routing results in electronic design automation pipelines; Based on the design data, extract the network-level geometric features of each network, including: parsing the network, pins, bounding boxes, local neighborhoods, obstacles and existing resource distribution from the design data, extracting the low-cost geometric features of each network, and forming network-level feature vectors; Construct network-level discrimination labels for each network based on the wiring results; Based on network-level geometric features and network-level discriminant labels, the discriminant power of at least one single geometric feature baseline and the discriminant power of the multi-feature learning model are evaluated under the board-level evaluation caliber, respectively, to obtain the baseline discriminant power and the learning discriminant power. The board-level evaluation caliber includes leave-one-board cross-validation or dividing the training set and validation set by board, wherein the network in the same board does not appear in the training set and validation set at the same time. Calculate the predictability gain based on the learned discriminant power and the baseline discriminant power; Based on the comparison between the predictability gain and the preset threshold, a gating indicator is generated; Based on the gating instruction, either the learning prior or the single-feature prior is selected as the prior to be injected. When the predictability gain reaches a preset threshold, the network risk score output by the multi-feature learning model is used as the learning prior. When the predictability gain does not reach the preset threshold, the learning prior is blocked from entering the wirebreaker, and a single-feature prior that is identical to the learning prior is generated by the optimal single-feature baseline and its direction. The prior to be injected is injected into the router, and the routing cost is modulated according to the prior to be injected, including: converting the prior to be injected into the network prior score, obtaining the basic cost of the router, generating a prior modulation term according to the network prior score and the preset prior injection coefficient, and using the sum of the basic cost and the prior modulation term as the modulated routing cost. Perform printed circuit board routing based on the modulated routing cost, including: performing global routing, detailed routing, or sequential routing based on the modulated routing cost; record the routing results and feed them back to the label construction stage for subsequent board-level evaluation or model updates; The process of obtaining the baseline discriminative power and the learned discriminative power includes: selecting at least one candidate single feature from the network-level geometric features; calculating the forward and reverse ranking discriminative power of each candidate single feature under the same board-level evaluation caliber; taking the better one as the discriminative power of the candidate single feature; and using the optimal value among the discriminative powers of each candidate single feature as the baseline discriminative power; and using a multi-feature learning model to predict the network-level discriminative label under the same board-level evaluation caliber to obtain the learned discriminative power. The method of obtaining the learned discriminative power by leaving one board as the validation set includes: using one board as the validation set and the remaining boards as the training set; making predictions on each validation set using the multi-feature learning model trained on the training set and calculating the discriminative power on that validation set; and calculating the average of the discriminative power on all validation sets as the learned discriminative power.

[0023] In summary, this method includes: acquiring network-level geometric features and routing result labels of multiple printed circuit board designs; constructing network labels for each network within each board, the labels being used to characterize whether the network is residual, whether it is in a high-risk congestion state, or whether it belongs to a difficult-to-route network; selecting at least one single inexpensive geometric feature as a baseline candidate, and calculating the discriminative power of each single feature in board-level cross-validation or leave-one-board evaluation; simultaneously, using a multi-feature learning model to calculate the discriminative power of the learning model under the same board-level partitioning caliber; using the difference between the discriminative power of the learning model and the discriminative power of the optimal single-feature baseline as the predictability gain; comparing this gain with a preset threshold, if the threshold is reached, generating a learning prior and injecting it into the router; if the threshold is not reached, generating a single-feature prior and injecting it into the router.

[0024] For the The first on the block Let there be a network, and let its network-level feature vector be... , tag as Candidate single feature is The strength of the plate level judgment is The discriminative power of a single feature baseline can be expressed as: ; in, Indicates the first The board-level discriminative power is obtained by ranking or scoring the network using individual features. This is used to cover both positive and negative correlation scenarios for the single feature. The board-level discriminative strength is not limited to AUC; it can also be ranking accuracy, average precision, recall-precision area ratio, correlation coefficient, or other equivalent metrics used to evaluate the network's risk ranking ability. Different metrics are used to compare the relative gain between the single-feature baseline and the learning model under the same board-level division caliber. Let... Indicates that except for the first The discriminative power of the learning model trained on data other than the board can be expressed as: ; in, For the number of board-level designs, The model represents the first time. The discriminative power on the block is left on the plate. The predictability gain is defined as: ; Further set the activation threshold as The gate control indication quantity is Then the gating decision can be expressed as: ; when When the system uses the learning priors output by the learning model, it will activate the learning model. When this condition is met, the system reverts to a single-feature prior. The square brackets indicate that a value of 1 is taken if the condition is true and 0 is taken otherwise.

[0025] In one implementation, the learned prior score is represented as The prior score for a single feature is expressed as: The final network prior score injected into the router is: ; The network prior score can be further transformed into network ranking priority, routing cost weight, congestion avoidance hints, or candidate path penalties. Let the basic cost of the router be... ,network In wiring unit The prior modulation term on is The prior injection coefficient is Then the routing cost after injection can be expressed as: ; Furthermore, the network prior scores can also directly form the network ranking and search budget. Let... Represents network The routing priority is determined by prior scores, where higher scores indicate higher priority. Therefore: ; in, This indicates that sorting numbers are generated based on the values ​​within the parentheses, from smallest to largest. Let... Indicates assignment to the network Cabling search budget, Based on search budget, Let be the budget modulation coefficient, then we have: ; Therefore, the present invention uses gating indication quantity The decision of "whether to enable complex learning priors" is transformed into an explicit technical step in the wiring prior generation chain. This ensures that the learning model is only enabled when a relatively inexpensive baseline has sufficient gain, and when the gain is insufficient, the learning prior is blocked from entering the router, and a homomorphic prior is generated instead from the optimal single-feature baseline. This homomorphic prior is not a temporary replacement after failure, but a formal technical branch of the gating results, which can influence network ranking, wiring costs, congestion avoidance cues, or search budgets through the same interface as the learning prior.

[0026] The following is a detailed explanation with reference to the accompanying drawings: like Figure 1 As shown, the method of the present invention specifically includes the following steps: Step 1: Obtain PCB design data and corresponding routing results. The PCB design data includes a netlist, pin coordinates, layer information, prohibited routing areas, obstacles, routing rules, and board-level design identifiers. The routing results can come from historical routing projects, trial routing outputs, or intermediate results from the current EDA pipeline.

[0027] Step two, extract inexpensive geometric features for each network. For the first... The first on the block A network is formed to generate feature vectors. This feature vector can contain either a single geometric indices or multiple geometric indices. Candidate Single Feature Selected from the components of the feature vector.

[0028] Step 3: Construct network-level discriminant labels. (Labels) It can indicate whether the network is a residual network, whether it is located in a high-congestion area, whether there has been a cabling failure, or whether there are more than a preset number of redeployment records. For continuous congestion levels, thresholding or sorting transformation can also be performed first to adapt to the discrimination strength.

[0029] Step four, perform single-feature baseline assessment. For example... Figure 2As shown, the system calculates the board-level discriminative power for each candidate single feature and simultaneously compares the forward and reverse ranking effects of that feature, ultimately obtaining the optimal single feature baseline discriminative power. This step is used to provide the "minimum interpretable performance benchmark that can be achieved without enabling complex learning models".

[0030] Step 5: Perform board-level evaluation of the learning model. See also... Figure 2 The system employs a multi-feature learning model to predict network labels and obtains results through cross-validation with either one board left or board-based partitioning. The learning model can be a tree model, a linear model, a graph model, a neural network, or other model that can output a network risk score; this invention does not limit the specific model structure.

[0031] Step six: Calculate the predictability gain and perform gating decision. For example... Figure 3 As shown, the system calculates and according to Output the gating result. If... Enter the prior learning path; if Enter the single-feature prior backtracking path.

[0032] Step 7: Generate and inject the selected routing priors. When entering a learned prior path, the system uses the learned model to output network risk scores, priorities, or congestion avoidance hints; when entering a single-feature prior path, the system uses the optimal single feature and its direction to generate a prior of the same form. The single-feature prior path is the formal output branch of the gating decision, not a temporary alternative after the learned path fails. Both types of priors are converted into network rankings, cost modulations, path penalty terms, or search budgets that the router can consume through a unified prior injection module, allowing downstream routers to execute the corresponding routing strategies without being aware of the prior source.

[0033] Step 8: Perform routing and record the results. For example... Figure 4 As shown, the router performs global routing, detailed routing, or sequential routing based on the injected priors. The routing results can be fed back into the label construction module for subsequent board-level evaluation, threshold correction, or model updates. This feedback does not require online closed-loop real-time training; it can also be used to generate the next round of prior evaluation data in an offline batch processing manner.

[0034] In this embodiment, the system collects several PCB designs, each containing a certain number of nets. The system calculates the number of pins, half-circuit trace length, bounding box area, local pin density, local obstacle density, and local routing requirement density for each net, and uses whether the net is not fully routed after routing as a label. Discrimination strength is then determined. Ranking-based AUC can be used, or ranking accuracy, average precision, recall-precision area ratio, correlation coefficient, or equivalent network risk ranking metrics can be employed. Regardless of the metric used, the single-feature baseline and the learning model are compared using the same board-level partitioning caliber. Threshold A value between 0.03 and 0.05 is acceptable. If the leave-one-board AUC of the learning model is 0.82, and the optimal single-feature baseline AUC is 0.79, then... When the threshold is 0.03, the gating is successful, and the system enables the learning prior. If the AUC of the learned model in another batch of designs is 0.78, and the AUC of the optimal single-feature baseline is 0.77, then... When the threshold is 0.03, the system reverts to a single-feature prior to avoid unnecessary overhead from complex models.

[0035] In another embodiment, the prior injection module injects network prior scores. This can be transformed into a network ranking factor, enabling higher-risk networks to be routed earlier; or into a routing unit cost modulation term, allowing high-risk networks to better avoid congested areas; or into a search budget modulation term, giving high-risk networks more candidate path attempts or a larger search range. For existing commercial or open-source routers, this invention can be integrated by preprocessing network ranking files, cost configuration files, region penalty maps, search budget configurations, or routing constraint parameters, without requiring modification to the router's core algorithm.

[0036] The key to this invention lies in the complete technical chain formed between single-feature baseline evaluation, board-level evaluation of the learning model, predictability gain calculation, threshold gating, dual prior selection, and unified injection. Any implementation that compares the board-level discrimination gain of the learning model relative to the inexpensive single-feature baseline before generating the routing prior, and determines whether to enable the learning prior or revert to the single-feature prior based on this gain, constitutes an equivalent implementation of the technical concept of this invention.

[0037] like Figure 5 As shown, this embodiment also provides a gain-gated printed circuit board routing prior-based on-demand activation system. The network feature extraction module parses the distribution of networks, pins, bounding boxes, local neighborhoods, obstacles, and existing resources from the PCB design data, and extracts inexpensive geometric features for each network. These inexpensive geometric features may include the number of pins, half-perimeter line length, bounding box area, bounding box aspect ratio, local pin density, local obstacle density, local routing requirement density, and the number of available channels in the network's area. The tag construction module constructs network-level tags based on historical routing results or trial routing results in the current process, such as whether the network is not fully routed, whether it belongs to a high-risk network in a congested area, whether multiple disconnections and reroutings have occurred, or whether it has resulted in routing residuals.

[0038] The single-feature baseline evaluation module calculates the discriminative power of each candidate single feature. Since different features may be positively or negatively correlated with the difficult-to-route label, the system considers both the forward and reverse ranking of single features and selects the better one as the discriminative power of that single feature. This optimal single feature is not a simple statistical comparison, but rather the lowest-cost, explainable prior source corresponding to the PCB routing scenario. For example, pin count, bounding box area, or local pin density can already reflect the difficulty of network routing in certain board-level designs. The learning model evaluation module trains a multi-feature learning model using the same network features and labels, but its evaluation must adopt the same board-level partitioning caliber as the single-feature baseline, such as leaving one board for cross-validation or dividing the training and validation sets by board. This avoids similar networks within the same board appearing simultaneously in training and validation, thereby reducing the risk of evaluation results being amplified by data leakage from the same board and making the gating decision closer to the actual usability of the learned prior on the new PCB.

[0039] The gain calculation and gating module calculates the predictability gain based on the optimal single-feature baseline discriminative power and the discriminative power of the learned model. .when Not less than the threshold When this occurs, it indicates that the learning model has sufficient discriminative improvement compared to relatively inexpensive single features, and the system activates the prior learning generation module; when Less than the threshold When this occurs, it indicates that the benefits of complex models are insufficient compared to single-feature models. The system then activates the single-feature prior generation module and blocks the learning of priors from entering the prior injection module. This gating decision is not a replacement for the router itself, but rather a prior source selection step before the router receives priors. Its output will further change the network ordering, cost modulation, avoidance cues, or search budget used by the router.

[0040] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for enabling prior knowledge of printed circuit board routing on demand based on gain gating, characterized in that, include: Obtain design data and routing results for multiple printed circuit boards; Based on the design data, extract the network-level geometric features of each network, and construct network-level discrimination labels for each network based on the wiring results; Based on the network-level geometric features and the network-level discriminant labels, the discriminant power of at least one single geometric feature baseline and the discriminant power of the multi-feature learning model are evaluated under the board-level evaluation caliber, respectively, and the baseline discriminant power and the learning discriminant power are obtained accordingly. Calculate the predictability gain based on the learned discriminant power and the baseline discriminant power; Based on the comparison result between the predictability gain and the preset threshold, a gating indicator is generated; Based on the gating indication, a learning prior or a single-feature prior is selected as the prior to be injected; wherein, when the predictability gain does not reach the preset threshold, the learning prior is blocked and the system reverts to the optimal single-feature baseline to generate the single-feature prior that is isomorphic to the learning prior; The prior to be injected is injected into the router. The prior to be injected is used to generate network sorting priority, modulate routing cost, generate congestion avoidance cues, or configure search budget to guide the router to perform printed circuit board routing.

2. The method according to claim 1, characterized in that, The acquisition of design data and wiring results for multiple printed circuit boards includes: Obtain the netlist, pin coordinates, layer information, prohibited routing areas, obstacles, and routing rules of the printed circuit board; Obtain the wiring results of the printed circuit board, including historical wiring results, trial wiring output results, and intermediate wiring results in the electronic design automation pipeline.

3. The method according to claim 1, characterized in that, Based on the design data, network-level geometric features of each network are extracted, including: The network, pins, bounding boxes, local neighborhoods, obstacles, and existing resource distribution are analyzed from the design data to extract the inexpensive geometric features of each network and form a network-level feature vector.

4. The method according to claim 1, characterized in that, The corresponding baseline discriminative power and learned discriminative power are obtained, including: At least one candidate single feature is selected from the network-level geometric features. The forward ranking discriminant power and the reverse ranking discriminant power of each candidate single feature are calculated under the board-level evaluation caliber. The better one is taken as the discriminant power of the candidate single feature, and the optimal value among the discriminant powers of each candidate single feature is taken as the baseline discriminant power. The learned discriminative power is obtained by predicting the network-level discriminative label using the multi-feature learning model under the same board-level evaluation caliber.

5. The method according to claim 4, characterized in that, The board-level evaluation criteria include leave one board for cross-validation and dividing the training set and validation set by board, wherein networks within the same board do not appear in the training set and validation set simultaneously.

6. The method according to claim 5, characterized in that, The process of obtaining the learned discriminative power includes: The leave-one-board cross-validation method is adopted, in which one board is used as the validation set and the remaining boards are used as the training set. On each of the validation sets, the multi-feature learning model trained on the training set is used to make predictions, and the discriminative power on that validation set is calculated. The average of the discriminative power on all validation sets is calculated as the learned discriminative power.

7. The method according to claim 1, characterized in that, Based on the gating indication, a learned prior or a single-feature prior is selected as the prior to be injected, including: When the predictability gain reaches the preset threshold, the network risk score is output by the multi-feature learning model as the learning prior. When the predictability gain does not reach the preset threshold, the learning prior is blocked from entering the wirer, and the single feature prior, which is isomorphic to the learning prior, is generated by the optimal single feature baseline and its direction.

8. The method according to claim 1, characterized in that, The prior to be injected into the router is modulated according to the prior to be injected, including: The prior to be injected is converted into network prior scores; Obtain the base cost of the router; Based on the network prior score and the preset prior injection coefficient, a prior modulation term is generated; The sum of the base cost and the prior modulation term is used as the modulated wiring cost.

9. The method according to claim 1, characterized in that, Performing printed circuit board routing based on the modulated routing cost includes: Perform global routing, detailed routing, or sequential routing based on the modulated routing cost; Record the routing results and feed them back to the label construction stage for subsequent board-level evaluation or model updates.

10. A gain-gated printed circuit board routing prior-on-demand activation system, characterized in that, include: The network feature extraction module is used to extract network-level geometric features from printed circuit board design data; The tag construction module is used to construct network-level discrimination tags based on the wiring results. The single-feature baseline evaluation module is used to evaluate the discriminative power of at least one single geometric feature baseline under the plate-level evaluation caliber, and obtain the baseline discriminative power; The learning model evaluation module is used to evaluate the discriminative power of the multi-feature learning model under the same board-level evaluation caliber, and obtain the learning discriminative power. The gain calculation and gating module is used to calculate the predictability gain based on the learned discriminative power and the baseline discriminative power, and generate a gating indication based on the comparison result of the predictability gain and a preset threshold. The learning prior generation module is used to generate a learning prior based on the gating indication when the predictability gain reaches the preset threshold. A single-feature prior generation module is used to block the learning prior and generate a single-feature prior that is identical to the learning prior when the predictability gain does not reach the preset threshold, based on the gating indication. The prior injection module is used to inject the learned prior or the single-feature prior into the router through a unified interface, and modulate the routing cost according to the injected prior. A routing execution module is used to perform printed circuit board routing according to the modulated routing cost.

Images