Self-adaptive fidelity model scheduling method and system for circuit parameter optimization

Abstract

The invention discloses an adaptive fidelity model scheduling method and system for circuit parameter optimization. The method comprises the following steps: selecting initial sample points to train a multi-precision agent model cluster; according to the parameter distribution characteristics of the to-be-evaluated individual, determining an adaptive target agent model from the multi-precision agent model cluster to perform performance evaluation on the to-be-evaluated individual; based on the evaluation error of the target agent model and the convergence state of the optimization process, judging whether a trigger threshold value of the high-precision circuit simulator is reached or not; if the triggering threshold value is reached, calling a high-precision circuit simulator to carry out performance verification on the to-be-evaluated individual, and obtaining generated verification data; and if the triggering threshold is not reached, obtaining an evaluation result output by the target agent model. According to the method, calling of a high-precision simulator is reduced from each individual to a key individual, conversion from whole-process high-precision to on-demand high-precision dynamic adaptation is achieved, and the large-scale circuit multi-target parameter optimization requirement under the advanced technology can be met.

Description

Technical Field

[0001] This application relates to the field of integrated circuit technology, specifically to an adaptive fidelity model scheduling method and system for circuit parameter optimization. Background Technology

[0002] In the field of integrated circuit design automation (EDA), circuit parameter optimization is a crucial step in ensuring that chip performance (such as timing, power consumption, and noise) meets standards. As chip manufacturing processes evolve towards advanced technologies, the complexity of circuit structures increases exponentially. The requirements for evaluation accuracy and efficiency in multi-objective parameter optimization (such as simultaneously achieving timing convergence and low power consumption) are becoming increasingly stringent. It is necessary to ensure that the optimization results match the actual circuit performance through accurate evaluation, while also controlling the evaluation time to avoid extending the design cycle. In existing technologies, performance evaluation of circuit parameter optimization requires the continuous use of a high-precision circuit simulator (such as HSPICE) to verify the performance of each individual to be evaluated in the optimization iteration. While this approach can guarantee evaluation accuracy, a single calculation by a high-precision simulator can take several minutes to several hours. Furthermore, multi-objective optimization algorithms (such as genetic algorithms and particle swarm optimization algorithms) require hundreds to thousands of iterations, with each iteration containing dozens to hundreds of individuals to be evaluated. This results in an excessively long overall optimization process, severely reducing design efficiency and making it particularly difficult to adapt to the rapid optimization requirements of large-scale circuits under advanced processes.

[0003] Therefore, how to reduce the overall optimization time and improve design efficiency while ensuring the accuracy of circuit parameter optimization and evaluation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, it is desirable to provide an adaptive fidelity model scheduling method and system for circuit parameter optimization. This method can dynamically schedule a multi-precision proxy model cluster and flexibly trigger high-precision simulation by combining evaluation error and optimization convergence state. While ensuring that the optimization results conform to the actual circuit performance, it can significantly reduce the time consumption of high-precision simulation and achieve a dynamic balance between optimization accuracy and efficiency.

[0005] In a first aspect, embodiments of this application provide an adaptive fidelity model scheduling method for circuit parameter optimization, comprising: An initial set of sample points is selected to train a cluster of multi-precision surrogate models; the cluster of multi-precision surrogate models includes surrogate models with different prediction accuracies and computational costs. Based on the parameter distribution characteristics of the individual to be evaluated, a suitable target agent model is determined from the multi-precision agent model cluster to evaluate the performance of the individual to be evaluated. Based on the evaluation error of the target proxy model and the convergence status of the optimization process, it is determined whether the trigger threshold of the high-precision circuit simulator has been reached. If the trigger threshold is reached, the high-precision circuit simulator is invoked to perform performance verification on the individual to be evaluated and to obtain the generated verification data. If the trigger threshold is not reached, obtain the evaluation result output by the target agent model.

[0006] In one embodiment, after invoking the high-precision circuit simulator to perform performance verification on the individual to be evaluated and obtaining the generated verification data, the method further includes: Inject the verification data into the sample pool; The multi-precision agent model cluster is dynamically updated based on the sample pool.

[0007] In one embodiment, injecting the verification data into the sample pool includes: The verification data is injected into the sample pool according to the error weight allocation strategy.

[0008] In one embodiment, selecting initial sample points to train the multi-precision surrogate model cluster includes: Based on the sensitive region identification results in the circuit parameter space, a stratified sampling strategy is adopted to select initial sample points; the sample density in the sensitive region is higher than that in the non-sensitive region. A multi-precision proxy model cluster is trained based on the initial sample points.

[0009] In one embodiment, the step of determining a suitable target agent model from the multi-precision agent model cluster to perform performance evaluation on the individual to be evaluated based on the parameter distribution characteristics of the individual to be evaluated includes: The Euclidean distance between the individual to be evaluated and the known sample points, as well as their density distribution in the parameter space, are determined as the parameter distribution features; Based on the parameter distribution characteristics, determine whether the individual to be evaluated belongs to a dense sample area; If it belongs to the dense sample region, select a low-precision proxy model from the multi-precision proxy model cluster; If it belongs to a sparse sample region, select a high-precision proxy model from the multi-precision proxy model cluster.

[0010] In one embodiment, determining whether the trigger threshold of the high-precision circuit simulator has been reached based on the evaluation error and convergence state of the optimization process of the target proxy model includes: If the target agent model is a probabilistic model, the prediction variance of the model output is used as the evaluation error; If the target proxy model is a deterministic model, calculate the absolute deviation rate between the evaluation result and the historical high-precision simulation data under the same parameters, and use the absolute deviation rate as the evaluation error.

[0011] In one embodiment, determining whether the trigger threshold of the high-precision circuit simulator has been reached based on the evaluation error and convergence state of the optimization process of the target proxy model includes: Calculate the rate of change of the optimal objective function value in N consecutive iterations, and use the rate of change as the convergence metric of the convergence state.

[0012] In one embodiment, calculating the rate of change of the optimal objective function value over N consecutive iterations includes: Calculate the difference between the optimal objective function value in the i-th iteration and the optimal objective function value in the iN-th iteration; The absolute value of the ratio of the difference to the optimal objective function value of the iNth iteration is taken as the rate of change.

[0013] In one embodiment, the adaptive fidelity model scheduling method for circuit parameter optimization further includes: Based on the evaluation error and the convergence state, the trigger threshold is dynamically adjusted according to the adjustment rules. The adjustment rules include: the larger the evaluation error, the lower the trigger threshold; and the larger the convergence quantification value of the convergence state, the higher the trigger threshold.

[0014] Secondly, embodiments of this application provide an adaptive fidelity model scheduling system for circuit parameter optimization, comprising: A cluster generation module is used to select initial sample points to train a multi-precision surrogate model cluster; the multi-precision surrogate model cluster contains surrogate models with different prediction accuracies and computational costs. The adaptation model evaluation module is used to determine the appropriate target agent model from the multi-precision agent model cluster based on the parameter distribution characteristics of the individual to be evaluated, and to evaluate the performance of the individual to be evaluated. The simulator trigger judgment module is used to determine whether the trigger threshold of the high-precision circuit simulator has been reached based on the evaluation error and the convergence status of the optimization process of the target proxy model; if the trigger threshold is reached, the performance verification module is triggered; if the trigger threshold is not reached, the result acquisition module is triggered. The performance verification module is used to call the high-precision circuit simulator to perform performance verification on the individual to be evaluated and obtain the generated verification data. The result acquisition module is used to acquire the evaluation results output by the target agent model.

[0015] The adaptive fidelity model scheduling method for circuit parameter optimization provided in this application breaks away from the existing evaluation mode that relies solely on high-precision simulators by selecting initial sample points to train a cluster of multi-precision surrogate models with different prediction accuracies and computational costs. Low-precision models can quickly complete a large number of basic evaluations to reduce costs and improve efficiency, while high-precision models can balance accuracy and speed in critical scenarios. Then, the appropriate target surrogate model is determined based on the parameter distribution characteristics of the individuals to be evaluated. Low-precision models are selected for dense sample areas to avoid redundant high-precision simulations, while high-precision models are selected for sparse sample areas to reduce prediction bias, thus achieving differentiated allocation of evaluation resources. The high-precision simulator trigger threshold is determined based on the evaluation error of the target surrogate model and the convergence status of the optimization process. Based on the threshold result, the method selects to call the high-precision simulator or obtain the evaluation result of the surrogate model, realizing flexible switching of evaluation modes. High costs are only consumed when necessary to ensure accuracy, while low-cost and fast evaluation continues in other scenarios, further compressing the optimization time.

[0016] This method reduces the use of high-precision simulators from individual components to key components, enabling a dynamic adaptation from end-to-end high precision to on-demand high precision. This comprehensively solves the problems of low optimization efficiency and difficulty in balancing accuracy and efficiency in existing technologies, and can adapt to the multi-objective parameter optimization needs of large-scale circuits under advanced processes.

[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating an adaptive fidelity model scheduling method for circuit parameter optimization provided in an embodiment of this application is shown. Figure 2 This illustration shows a schematic diagram of an adaptive fidelity model scheduling system for circuit parameter optimization provided in an embodiment of this application. Detailed Implementation

[0019] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present application will now be described in detail with reference to the accompanying drawings and embodiments. Although the embodiments of this application provide method operation instruction steps as shown in the following embodiments or drawings, more or fewer operation instruction steps may be included in the method based on conventional or non-inventive effort. In steps where there is no logically necessary causal relationship, the execution order of these steps is not limited to the execution order provided in the embodiments of this application. In actual processing or when the device executes, the method may be executed sequentially or in parallel according to the method shown in the embodiments or drawings.

[0021] Example 1: This embodiment proposes an adaptive fidelity model scheduling method for circuit parameter optimization. Please refer to [link / reference]. Figure 1 , Figure 1 A flowchart illustrating the adaptive fidelity model scheduling method for circuit parameter optimization provided in this embodiment is shown. Figure 1 As shown, the method mainly includes: Step 101: Select initial sample points to train the multi-precision surrogate model cluster.

[0022] To ensure evaluation accuracy, existing technologies require calling a high-precision simulator for each individual to be evaluated in the optimization iteration. However, a single calculation by a high-precision simulator can take several minutes to several hours. Furthermore, multi-objective optimization algorithms need to complete hundreds to thousands of iterations, with each iteration containing dozens to hundreds of individuals to be evaluated. This results in an extremely long overall optimization time, which cannot meet the rapid design requirements of large-scale circuits under advanced processes.

[0023] To address this, this method selects initial sample points to construct and train an ensemble containing multiple surrogate models, i.e., a multi-precision surrogate model cluster. Each surrogate model within the cluster possesses different prediction accuracies and computational costs. For example, some surrogate models employ simplified machine learning algorithms (such as linear regression and lightweight neural networks), which, while having relatively lower prediction accuracy, offer fast model training and inference speeds and consume fewer computational resources (low computational cost). Other surrogate models employ complex algorithms (such as deep neural networks and Gaussian process regression), which, while offering prediction accuracy approaching that of high-precision circuit simulators (such as HSPICE), require longer model training and inference times and have higher computational resource demands (high computational cost). These surrogate models with varying accuracies and costs collaboratively form a cluster, creating a gradient evaluation toolkit ranging from low-precision to low-cost to high-precision to high-cost.

[0024] It should be noted that this embodiment does not limit the training method of each proxy model in the multi-precision proxy model cluster. You can refer to the introduction of relevant technologies, which will not be repeated here.

[0025] Step 102: Based on the parameter distribution characteristics of the individual to be evaluated, determine the appropriate target agent model from the multi-precision agent model cluster to evaluate the performance of the individual to be evaluated.

[0026] The parameter distribution characteristics of the individual to be evaluated include the positional attributes of the circuit parameter combination to be evaluated in the entire circuit parameter space and the distribution of surrounding samples. The specific parameter types are not limited in this embodiment.

[0027] Based on the above parameter distribution characteristics, specific proxy models that match the evaluation needs of the individuals to be evaluated are selected from the gradient model library within the cluster, ranging from low precision to low cost to high precision to high cost.

[0028] The circuit parameters of the individual to be evaluated are input into the selected target surrogate model. The surrogate model quickly outputs the circuit performance prediction results (such as timing delay, power consumption, etc.) corresponding to the parameter combination based on the parameter-performance mapping relationship learned during the training process, thus replacing the traditional high-precision circuit simulator to complete the performance evaluation.

[0029] This step, based on the core principle of parameter distribution characteristics, links the selection of evaluation tools with the actual attributes of the individuals to be evaluated, breaking the rigid logic of a one-size-fits-all approach. It allows individuals with different characteristics to be evaluated to be matched with more suitable models, enabling the efficiency advantages of low-precision models and the accuracy advantages of high-precision models to be brought into play in a targeted manner, avoiding the waste of cluster resources and avoiding the waste of accuracy or loss of efficiency caused by neglecting individual differences in existing technologies.

[0030] Step 103: Based on the evaluation error of the target surrogate model and the convergence status of the optimization process, determine whether the trigger threshold of the high-precision circuit simulator has been reached. If the trigger threshold has been reached, proceed to step 104; otherwise, proceed to step 105.

[0031] The evaluation error of the target surrogate model refers to the quantified deviation between the performance prediction result of the target surrogate model for the current individual to be evaluated and the actual performance (or historical high-precision simulation data). For example, the prediction variance of a probabilistic model, the absolute deviation rate between the prediction value of a deterministic model and historical simulation data. This error directly reflects the reliability of the surrogate model's prediction results; the larger the error, the lower the credibility of the prediction results. Secondly, the convergence state of the optimization process refers to the quantified degree of the parameters approaching the optimal solution during the circuit parameter optimization iteration process. For example, the rate of change of the optimal objective function value in N consecutive iterations (the smaller the rate of change, the closer the parameters are to the optimal solution, and the higher the degree of convergence), and the concentration of parameter distribution of individuals in the population (the higher the concentration, the higher the degree of convergence). This state reflects the stage characteristics of the optimization process; the higher the degree of convergence, the smaller the space for subsequent parameter adjustment.

[0032] A threshold standard related to evaluation error and convergence state is pre-set. The quantized values ​​of the currently calculated evaluation error and convergence state are substituted into the judgment logic. If the evaluation error exceeds a certain value or the convergence state does not reach a certain level, it is determined that the trigger threshold has been reached and the high-precision circuit simulator needs to be called. Otherwise, it is determined that the threshold has not been reached and there is no need to call it. The evaluation result output by the target proxy model can be directly used as the simulation result output.

[0033] Step 104: Call the high-precision circuit simulator to verify the performance of the individual to be evaluated and obtain the generated verification data.

[0034] Launch professional circuit simulation tools with industrial-grade precision, such as HSPICE and Spectre. Import the individual circuit to be evaluated (i.e., the specific combination of circuit parameters to be evaluated in the optimization iteration, such as transistor aspect ratio, resistor and capacitor values, etc.) into the simulator format and configure a simulation environment consistent with the actual application (such as supply voltage, temperature, and input signals). Solve the circuit physical equations and simulate the real operating characteristics of the devices (such as MOSFET IV curves and the influence of parasitic parameters) through the simulator to simulate the actual operation of the individual circuit to be evaluated. Calculate and output accurate values ​​of core performance indicators such as timing delay, power consumption, and noise margin to verify the reliability of the previous proxy model evaluation results. The system performs several tests, including: determining whether the deviation between the surrogate model's predicted values ​​and the simulation's true values ​​is within an acceptable range; collecting all performance-related data generated by the simulator, including quantified values ​​of core indicators (serving as a true benchmark for correcting the surrogate model and determining the optimization direction), as well as auxiliary data such as voltage waveforms and current change curves at key nodes (used to analyze the causes of performance deviations). This high-precision, high-reliability data is stored in a sample pool according to a preset format, providing true samples for subsequent surrogate model updates and a reliable basis for adjusting parameters in the optimization algorithm. Ultimately, this achieves the core objective of calibrating lightweight evaluation results with real circuit performance data, ensuring the accuracy and reliability of circuit parameter optimization.

[0035] Step 105: Obtain the evaluation results output by the target agent model.

[0036] When the trigger threshold of the high-precision circuit simulator is not reached, the lightweight performance evaluation result of the target proxy model is directly obtained as the current simulation result.

[0037] Based on the above introduction, the adaptive fidelity model scheduling method for circuit parameter optimization provided in this embodiment breaks the existing evaluation mode that relies solely on high-precision simulators by selecting initial sample points to train a cluster of multi-precision proxy models with different prediction accuracies and computational costs. Low-precision models can quickly complete a large number of basic evaluations to reduce costs and improve efficiency, while high-precision models can balance accuracy and speed in key scenarios. Then, the appropriate target proxy model is determined according to the parameter distribution characteristics of the individual to be evaluated, realizing differentiated allocation of evaluation resources. Based on the evaluation error of the target proxy model and the convergence status of the optimization process, the trigger threshold of the high-precision simulator is determined. According to the threshold result, the high-precision simulator is called or the evaluation result of the proxy model is obtained, realizing flexible switching of evaluation mode. High cost is only consumed to ensure accuracy when necessary, while low-cost and fast evaluation continues in other scenarios, further compressing the optimization time. Thus, while ensuring that the optimization results are consistent with the actual circuit performance, the high-precision simulation time is significantly reduced, achieving the goal of dynamic balance between optimization accuracy and efficiency.

[0038] Example 2: To improve the predictive reliability of the multi-precision surrogate model cluster, after calling a high-precision circuit simulator to verify the performance of the individual to be evaluated and obtaining the generated verification data, the following steps can be further performed: Step 106: Inject the validation data into the sample pool.

[0039] The verification data with truth value attributes output by the high-precision simulator, including the parameter combination of the individual to be evaluated and the corresponding accurate performance indicators, such as timing delay and power consumption, are supplemented into the sample pool storing historical sample data according to a preset format, so that the coverage of the sample pool continues to expand as the optimization process progresses, especially filling the sparse areas in the parameter space.

[0040] Step 107: Dynamically update the multi-precision agent model cluster based on the sample pool.

[0041] Using the updated sample pool as the training data basis, the models (low-precision and high-precision proxy models) in the multi-precision proxy model cluster are dynamically updated to avoid the model prediction accuracy from decreasing with optimization iterations due to insufficient coverage of the initial sample pool. By continuously injecting high-precision validation data, the model cluster evolves synchronously with the optimization process, and the prediction reliability of the multi-precision proxy model cluster continues to improve with the optimization process.

[0042] In this case, since the high-precision circuit simulator has small error and high prediction accuracy, in order to update the prediction data generated by the limited-focus high-precision circuit simulator, step 106, which injects the verification data into the sample pool, can specifically be: injecting the verification data into the sample pool according to the error weight allocation strategy.

[0043] When supplementing the sample pool with the validation data (including the parameter combinations and accuracy performance indicators of the individuals to be evaluated) output by the high-precision simulator, not all validation data are given equal importance. Instead, the data weight is determined based on the evaluation error of the target surrogate model corresponding to the individual to be evaluated. If the previous evaluation error of the target surrogate model for that individual is large (e.g., the deviation between the predicted value and the simulation true value exceeds a preset threshold), then the validation data is assigned a higher weight (e.g., the weight coefficient is set to 1.2 to 1.5); if the evaluation error is small, then a lower weight is assigned (e.g., the weight coefficient is set to 0.6 to 0.9). The combination of data and weight is stored in the sample pool in a preset format, so that the contribution of different data in the sample pool to the subsequent model training varies with the size of the error.

[0044] High-weight, high-error data can guide the model to prioritize correcting regions with large prediction biases in the parameter space (such as sparse sample regions), which is more efficient than equal-weighted injection. Low-weight, low-error data, although still participating in training, have a reduced impact on model parameter adjustment, which can reduce the model's overfitting to already fully covered regions and enhance the model's generalization ability. This allows the multi-precision surrogate model cluster to evolve more accurately as the optimization process progresses, continuously reducing the dependence on high-precision simulators and balancing the accuracy and efficiency of circuit parameter optimization.

[0045] Example 3: The above embodiments do not limit the process of training the multi-precision proxy model cluster. To address the problem of weak model training foundation caused by traditional random sampling of initial samples, this embodiment proposes a sample selection method that reduces the total number of initial samples while ensuring model training accuracy. Step 101, selecting initial sample points to train the multi-precision proxy model cluster, can be divided into the following two sub-steps: Step 11: Based on the sensitive region identification results in the circuit parameter space, a stratified sampling strategy is used to select initial sample points.

[0046] First, through circuit simulation or parameter sensitivity analysis (such as Monte Carlo analysis or Sobol exponential analysis), identify sensitive regions (such as regions where small changes in transistor size parameters lead to large fluctuations in timing delay) and less sensitive regions in the circuit parameter space that significantly affect performance indicators (timing, power consumption, etc.). Then, based on the identification results, a stratified sampling strategy is used to select initial sample points, dividing the parameter space into different levels according to sensitive and non-sensitive regions. More dense sample points are arranged in sensitive regions (such as the number of samples per unit space being 2 to 3 times that in non-sensitive regions), while relatively sparse sample points are arranged in non-sensitive regions. This ensures that the initial samples can fully cover the key sensitive regions while avoiding sample redundancy in non-sensitive regions.

[0047] Step 12: Train a multi-precision proxy model cluster based on the initial sample points.

[0048] Using these differentially distributed initial sample points as training data, a cluster of multi-precision surrogate models with different precision and computational costs (such as low-precision linear regression models and high-precision deep neural network models) is constructed and trained, enabling each model to learn the parameter-performance mapping rules of sensitive regions from the training stage.

[0049] Based on the above description, compared with random sampling, the stratified sampling method provided in this embodiment significantly improves the coverage of the initial sample in the sensitive area, enabling the multi-precision surrogate model cluster to accurately capture the performance change patterns of key areas from the training stage, thus reducing the initial prediction error. The combination of high-density samples in sensitive areas and sparse samples in non-sensitive areas reduces the total number of initial samples while ensuring the model training accuracy, thereby reducing the computational cost of data collection and model training. The multi-precision surrogate model cluster trained based on this initial sample has higher reliability in evaluating individuals to be evaluated in sensitive areas during subsequent optimization iterations, which can reduce the additional calls to the high-precision simulator due to insufficient initial training, balancing model training efficiency and subsequent evaluation accuracy, and providing a solid starting point for the entire adaptive scheduling process.

[0050] Example 4: The above embodiments do not limit the specific implementation of determining the suitable target proxy model from the multi-precision proxy model cluster. In order to ensure that the selection logic is deeply compatible with the gradient characteristics of the multi-precision proxy model cluster, this embodiment proposes an adaptation screening rule. Specifically, step 102 determines the suitable target proxy model from the multi-precision proxy model cluster based on the parameter distribution characteristics of the individual to be evaluated, and performs performance evaluation on the individual to be evaluated. This can be performed according to the following sub-steps: Step 21: Determine the Euclidean distance between the individual to be evaluated and the known sample points, as well as their density distribution in the parameter space, as parameter distribution features.

[0051] First, the Euclidean distance between the individual to be evaluated (a certain combination of circuit parameters) and all known sample points in the sample pool (including initial sample points and subsequently injected validation data) is calculated (to quantify the spatial proximity between the individual and historical samples), and the number of known sample points in the unit parameter space around the individual is counted (to determine the density distribution). These two indicators are used together as parameter distribution features to characterize the positional features of the individual in the parameter space.

[0052] Step 22: Determine whether the individual to be evaluated belongs to a dense sample region based on the parameter distribution characteristics; if it belongs to a dense sample region, select a low-precision proxy model from the multi-precision proxy model cluster; if it belongs to a sparse sample region, select a high-precision proxy model from the multi-precision proxy model cluster.

[0053] Based on the parameter distribution characteristics, the region to which an individual belongs is determined. If the Euclidean distance is small and the surrounding sample density is high (i.e., a dense sample region, indicating that historical samples fully represent the performance pattern of this region), then a low-precision surrogate model (such as a linear regression model) is selected from the multi-precision surrogate model cluster for performance evaluation. If the Euclidean distance is large and the surrounding sample density is low (i.e., a sparse sample region, indicating that historical samples do not fully represent the performance pattern of this region), then a high-precision surrogate model (such as a deep neural network model) is selected for evaluation.

[0054] The above-mentioned solution provided in this embodiment accurately defines regional characteristics through the dual indicators of Euclidean distance and density distribution, avoiding the judgment bias of a single indicator. At the same time, selecting a low-precision model in dense areas can reduce the time of a single evaluation by more than 50%, while selecting a high-precision model in sparse areas can reduce the prediction error by 30% to 50%. While ensuring the reliability of the evaluation, it maximizes the compression of computational costs, allowing the efficiency advantage of the low-precision model and the accuracy advantage of the high-precision model to be fully utilized in the appropriate region. This reduces the extra calls to the high-precision simulator due to improper model selection, provides key support for the accuracy-efficiency balance of circuit parameter optimization, and reduces the dependence on computing resources, thus adapting to the actual needs of large-scale circuit optimization.

[0055] Example 5: The above embodiments indicate that, based on the evaluation error of the target proxy model and the convergence state of the optimization process, it is determined whether the trigger threshold of the high-precision circuit simulator has been reached. The specific quantitative calculation method of the evaluation error and the convergence state is not limited. This embodiment proposes a specific calculation method. Other calculation methods can refer to the description in this embodiment, and will not be repeated here.

[0056] One method for calculating the evaluation error is as follows: if the target surrogate model is a probabilistic model, the prediction variance of the model output is used as the evaluation error; if the target surrogate model is a deterministic model, the absolute deviation rate between the evaluation result and the historical high-precision simulation data under the same parameters is calculated, and the absolute deviation rate is used as the evaluation error.

[0057] When determining the trigger threshold for a high-precision simulator, the type of the target surrogate model is first distinguished. If the target surrogate model is a probabilistic model (such as a Gaussian process regression model), this type of model inherently possesses the ability to output prediction uncertainty. Therefore, its output prediction variance (quantifying the uncertainty of the model regarding the current evaluation result; the larger the variance, the lower the prediction reliability) is directly used as the evaluation error. If the target surrogate model is a deterministic model (such as linear regression or ordinary neural network models), this type of model only outputs a single predicted value. The absolute deviation rate between its output evaluation result and existing historical high-precision simulation data with the same parameter combination needs to be calculated, and this deviation rate is used as the evaluation error. Subsequently, the convergence status of the optimization process (such as the rate of change of the optimal value in continuous iterations) is combined to jointly determine whether the current evaluation error reaches the threshold standard for triggering the high-precision simulator.

[0058] In the quantitative calculation scheme for evaluation error provided in this embodiment, the probabilistic model directly reuses its own output prediction variance without additional calculation, which can reduce the computational cost. The deterministic model obtains the deviation rate by comparing with the historical true value, ensuring the objectivity of error measurement.

[0059] One way to calculate the convergence state is as follows: calculate the rate of change of the optimal objective function value in N consecutive iterations, and use the rate of change as the convergence metric of the convergence state.

[0060] When determining the trigger threshold for a high-precision simulator, for the iterative process of circuit parameter optimization (such as iterations based on genetic algorithms or particle swarm optimization), the optimal objective function value output in each of N consecutive iterations (N is a preset positive integer, usually set to 5 to 20 depending on the optimization scale) is selected. The rate of change is calculated based on the output optimal objective function value, and this rate of change is used as the convergence metric for the convergence state. The smaller the rate of change, the smaller the improvement of the optimal performance index in consecutive iterations, and the closer the optimization process is to convergence (the parameters have gradually approached the optimal solution); the larger the rate of change, the more significantly the optimal performance index is still improving, and the optimization process is in the exploratory stage before convergence. Subsequently, the evaluation error of the target surrogate model is combined to jointly determine whether the threshold standard for triggering the high-precision simulator has been reached.

[0061] Calculating the rate of change of the optimal objective function value over N consecutive iterations may specifically include the following sub-steps: Calculate the difference between the optimal objective function value in the i-th iteration and the optimal objective function value in the iN-th iteration; take the absolute value of the ratio of the difference to the optimal objective function value in the iN-th iteration as the rate of change.

[0062] In the iterative process of circuit parameter optimization, when it is necessary to calculate the rate of change of the optimal objective function value for N consecutive iterations, first determine the current iteration number as i (e.g., the 20th iteration), and locate the optimal objective function value output by the i-th iteration (e.g., the optimal timing delay value of 45ps) and the optimal objective function value output by the iN-th iteration (e.g., the 10th iteration when N=10) (e.g., the optimal timing delay value of 55ps). Obtain the difference between the two by subtraction (45ps-55ps=-10ps).

[0063] The difference (-10ps) is then divided by the optimal objective function value (55ps) of the iNth iteration to obtain a ratio (-10ps / 55ps≈-0.182). The absolute value of this ratio (0.182) is then taken as the rate of change of the optimal objective function value over N consecutive iterations (18.2%). The rate of change directly reflects the relative change of the optimal performance index from the iNth iteration to the ith iteration, providing a specific quantitative basis for determining the convergence state.

[0064] Calculating the ratio based on the optimal value of the iNth iteration, rather than the absolute difference, can eliminate the influence of the magnitude difference in the performance indicators of different circuits, making the rate of change comparable among different performance indicators, and making the quantification results more universal.

[0065] Example 6: To avoid the deviation not being corrected in time due to an excessively high threshold in the early stage of optimization, or the simulation resources being wasted due to an excessively low threshold in the later stage of optimization, and to ensure that the trigger threshold is deeply matched with the real-time evaluation requirements and the iteration stage, the following steps can be further performed on the basis of the above embodiments: dynamically adjust the trigger threshold according to the adjustment rules based on the evaluation error and the convergence status.

[0066] Specifically, during the circuit parameter optimization iteration process, instead of using a fixed trigger threshold to determine whether to call the high-precision simulator, the trigger threshold is adjusted in real time according to a preset adjustment rule based on both the evaluation error of the target surrogate model (such as the prediction variance of a probabilistic model and the absolute deviation rate of a deterministic model) and the convergence metric of the optimization process (such as the absolute value of the rate of change of the optimal objective function value over N consecutive iterations). When the evaluation error is larger (indicating lower predictive reliability of the surrogate model and requiring stricter precision control), the trigger threshold is lowered (e.g., from triggering when the error is >15% to triggering when the error is >10%), making it easier for the high-precision simulator to be triggered to correct deviations. When the convergence metric is larger (indicating a smaller rate of change, the optimization process is closer to convergence, and the parameters tend to be stable), the trigger threshold is raised (e.g., from triggering when the error is >10% to triggering when the error is >20%), reducing unnecessary calls to the high-precision simulator to save costs.

[0067] By adjusting the threshold in conjunction with the evaluation error, high-precision simulation can be activated in a timely manner when the prediction reliability is low, keeping the deviation in the optimization process within a smaller range and ensuring optimization accuracy. In addition, by adjusting the threshold in conjunction with the convergence quantification value, the frequency of simulation calls can be reduced in the later stages of optimization (high convergence) (e.g., only triggered when the error is >20%), which reduces the number of high-precision simulator calls, significantly compresses the overall optimization time, and allows high-precision simulation resources to be focused on the truly needed scenarios, adapting to the dynamic needs of optimizing circuit parameters of different scales and processes.

[0068] Example 7: This embodiment provides an adaptive fidelity model scheduling system for circuit parameter optimization, such as... Figure 2 As shown, the system mainly includes: a cluster generation module 201, an adaptation model evaluation module 202, a simulator trigger judgment module 203, a performance verification module 204, and a result acquisition module 205.

[0069] Among them, the cluster generation module 201 is used to select initial sample points to train a multi-precision surrogate model cluster; the multi-precision surrogate model cluster contains surrogate models with different prediction accuracies and computational costs; The adaptation model evaluation module 202 is used to determine the appropriate target agent model from the multi-precision agent model cluster based on the parameter distribution characteristics of the individual to be evaluated, and to evaluate the performance of the individual to be evaluated. The simulator trigger judgment module 203 is used to determine whether the trigger threshold of the high-precision circuit simulator has been reached based on the evaluation error and the convergence status of the optimization process of the target proxy model. If the trigger threshold is reached, the performance verification module is triggered; if the trigger threshold is not reached, the result acquisition module is triggered. The performance verification module 204 is used to call a high-precision circuit simulator to perform performance verification on the individual to be evaluated and to obtain the generated verification data. The result acquisition module 205 is used to acquire the evaluation results output by the target agent model.

[0070] It should be noted that the adaptive fidelity model scheduling system for circuit parameter optimization provided in this embodiment and the adaptive fidelity model scheduling method for circuit parameter optimization provided in the above embodiments can be referred to each other, and the repeated parts will not be repeated in this embodiment.

[0071] The adaptive fidelity model scheduling system for circuit parameter optimization provided in this embodiment features a cluster generation module that constructs a multi-precision proxy model cluster. This breaks away from the limitations of existing technologies that rely solely on high-precision simulators. It provides flexible tools for subsequent evaluation through a gradient model library ranging from low-precision to low-cost to high-precision to high-cost, reducing reliance on high-consumption simulation resources from the outset. The adaptive model evaluation module selects target proxy models based on the parameter distribution characteristics of the individuals to be evaluated, avoiding indiscriminate evaluation of all individuals. The simulator trigger judgment module dynamically determines the trigger threshold based on evaluation error and convergence status, enabling on-demand invocation of high-precision simulators. This avoids the accumulation of deviations due to threshold rigidity in the early stages of optimization and reduces unnecessary simulation consumption in the later stages. The conditional trigger design of the performance verification module and the result acquisition module forms a closed loop of proxy model evaluation-simulation verification / result output. While ensuring the truth value of evaluations in key scenarios, it maximizes the compression of overall optimization time. Furthermore, the clear division of labor and collaborative linkage of each module allows for adaptation to parameter optimization needs of circuits with different processes and scales, effectively balancing the accuracy and efficiency of circuit parameter optimization. This addresses the core pain points of existing technologies, such as long optimization cycles, high resource consumption, and poor adaptability.

[0072] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operational instructions of possible implementations of methods and systems according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two connected blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operational instructions, or using a combination of dedicated hardware and computer instructions.

[0073] The units or modules described in the embodiments of this application can be implemented in software or hardware. The described units or modules can also be located in a processor. The names of these units or modules do not, in certain circumstances, constitute a limitation on the unit or module itself.

[0074] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. An adaptive fidelity model scheduling method for circuit parameter optimization, characterized in that, include: An initial set of sample points is selected to train a cluster of multi-precision surrogate models; the cluster of multi-precision surrogate models includes surrogate models with different prediction accuracies and computational costs. Based on the parameter distribution characteristics of the individual to be evaluated, a suitable target agent model is determined from the multi-precision agent model cluster to evaluate the performance of the individual to be evaluated. Based on the evaluation error of the target proxy model and the convergence status of the optimization process, it is determined whether the trigger threshold of the high-precision circuit simulator has been reached. If the trigger threshold is reached, the high-precision circuit simulator is invoked to perform performance verification on the individual to be evaluated and to obtain the generated verification data. If the trigger threshold is not reached, obtain the evaluation result output by the target agent model.

2. The method as described in claim 1, characterized in that, After invoking the high-precision circuit simulator to perform performance verification on the individual to be evaluated and obtaining the generated verification data, the process further includes: Inject the verification data into the sample pool; The multi-precision agent model cluster is dynamically updated based on the sample pool.

3. The method as described in claim 2, characterized in that, Injecting the verification data into the sample pool includes: The verification data is injected into the sample pool according to the error weight allocation strategy.

4. The method as described in claim 1, characterized in that, The process of selecting initial sample points to train a multi-precision surrogate model cluster includes: Based on the sensitive region identification results in the circuit parameter space, a stratified sampling strategy is adopted to select initial sample points; the sample density in the sensitive region is higher than that in the non-sensitive region. A multi-precision proxy model cluster is trained based on the initial sample points.

5. The method as described in claim 1, characterized in that, The step of determining a suitable target agent model from the multi-precision agent model cluster based on the parameter distribution characteristics of the individual to be evaluated, and then performing performance evaluation on the individual to be evaluated, includes: The Euclidean distance between the individual to be evaluated and the known sample points, as well as their density distribution in the parameter space, are determined as the parameter distribution features; Based on the parameter distribution characteristics, determine whether the individual to be evaluated belongs to a dense sample area; If it belongs to the dense sample region, select a low-precision proxy model from the multi-precision proxy model cluster; If it belongs to a sparse sample region, select a high-precision proxy model from the multi-precision proxy model cluster.

6. The method as described in claim 1, characterized in that, The determination of whether the trigger threshold of the high-precision circuit simulator has been reached, based on the convergence state of the evaluation error and optimization process of the target proxy model, includes: If the target agent model is a probabilistic model, the prediction variance of the model output is used as the evaluation error; If the target proxy model is a deterministic model, calculate the absolute deviation rate between the evaluation result and the historical high-precision simulation data under the same parameters, and use the absolute deviation rate as the evaluation error.

7. The method as described in claim 1, characterized in that, The determination of whether the trigger threshold of the high-precision circuit simulator has been reached, based on the convergence state of the evaluation error and optimization process of the target proxy model, includes: Calculate the rate of change of the optimal objective function value in N consecutive iterations, and use the rate of change as the convergence metric of the convergence state.

8. The method as described in claim 7, characterized in that, The calculation of the rate of change of the optimal objective function value over N consecutive iterations includes: Calculate the difference between the optimal objective function value in the i-th iteration and the optimal objective function value in the iN-th iteration; The absolute value of the ratio of the difference to the optimal objective function value of the iNth iteration is taken as the rate of change.

9. The method according to any one of claims 1 to 8, characterized in that, Also includes: Based on the evaluation error and the convergence state, the trigger threshold is dynamically adjusted according to the adjustment rules. The adjustment rules include: the larger the evaluation error, the lower the trigger threshold; and the larger the convergence quantification value of the convergence state, the higher the trigger threshold.

10. An adaptive fidelity model scheduling system for circuit parameter optimization, characterized in that, include: A cluster generation module is used to select initial sample points to train a multi-precision surrogate model cluster; the multi-precision surrogate model cluster contains surrogate models with different prediction accuracies and computational costs. The adaptation model evaluation module is used to determine the appropriate target agent model from the multi-precision agent model cluster based on the parameter distribution characteristics of the individual to be evaluated, and to evaluate the performance of the individual to be evaluated. The simulator trigger judgment module is used to determine whether the trigger threshold of the high-precision circuit simulator has been reached based on the evaluation error and the convergence status of the optimization process of the target proxy model. If the trigger threshold is reached, the performance verification module is triggered; if the trigger threshold is not reached, the result acquisition module is triggered. The performance verification module is used to call the high-precision circuit simulator to perform performance verification on the individual to be evaluated and obtain the generated verification data. The result acquisition module is used to acquire the evaluation results output by the target agent model.

Images