Circuit PVT performance optimization method and system based on KAN-CNN integration

Abstract

The invention discloses a circuit PVT performance optimization method and system based on KAN-CNN integration. The method comprises the steps that a parameter sample combination data set of a to-be-optimized circuit is constructed, and PVT vector codes are acquired; the parameter sample combination data set of the to-be-optimized circuit is weighted through an adaptive sampling strategy, and a circuit parameter sample combination data set with a weight coefficient is obtained; constructing a circuit PVT performance prediction model by combining a self-attention module, a KAN network structure and a CNN network structure; predicting the circuit parameter sample combination data set with the weight coefficient and the PVT vector code based on the circuit PVT performance prediction model to obtain a circuit PVT performance prediction data table; and performing circuit PVT performance optimization according to the circuit PVT performance prediction data table to obtain an optimized circuit. According to the embodiment of the invention, the prediction efficiency and prediction precision of the model can be improved, and synchronous modeling and optimization of various PVT performance indexes of the circuit are realized. The method can be widely applied to the technical field of analog circuit optimization.

Description

Technical Field

[0001] The present application relates to the technical field of analog circuit optimization, and in particular to a circuit PVT performance optimization method and system based on KAN-CNN integration. Background Art

[0002] With the rapid development of artificial intelligence and electronic design automation technology, size optimization in integrated circuit design has gradually shifted to methods based on agent models. However, existing methods still face challenges in dealing with the nonlinear characteristics of analog circuits and process, voltage, and temperature (PVT) changes. The Gaussian process model is too computationally expensive when dealing with large-scale high-dimensional data, while graph neural networks (GNNs) perform well in capturing local features but have difficulty in fully reflecting the global behavior of complex circuits. In addition, current research is still not perfect in modeling circuit performance across multiple PVT corners, and there are problems such as high model complexity and large training sample requirements.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to propose a circuit PVT performance optimization method and system based on KAN-CNN integration, which can improve the prediction efficiency and prediction accuracy of the model and realize the simultaneous modeling and optimization of multiple PVT performance indicators of the circuit.

[0005] To achieve the above objectives, an embodiment of the present application proposes a circuit PVT performance optimization method based on KAN-CNN integration, the method comprising:

[0006] Construct a parameter sample combination data set of the circuit to be optimized and obtain the PVT vector encoding;

[0007] Assigning weights to the parameter sample combination data set of the circuit to be optimized through an adaptive sampling strategy to obtain a circuit parameter sample combination data set with a weight coefficient;

[0008] Combining the self-attention module, KAN network structure and CNN network structure, a circuit PVT performance prediction model is constructed;

[0009] Based on the circuit PVT performance prediction model, the circuit parameter sample combined data set with weight coefficients and the PVT vector code are predicted to obtain a circuit PVT performance prediction data table;

[0010] The circuit PVT performance is optimized according to the circuit PVT performance prediction data table to obtain an optimized circuit.

[0011] In some embodiments, constructing a parameter sample combination data set of the circuit to be optimized and obtaining a PVT vector code includes:

[0012] Acquiring several adjustable parameters of the circuit to be optimized, and defining value ranges of several of the adjustable parameters;

[0013] Dividing the value ranges of the plurality of adjustable parameters to obtain equally divided value ranges of the adjustable parameters;

[0014] Based on the adjustable parameter value range after the equal division, random sampling is performed, and the sampling results are processed by permutation and combination to obtain a parameter sample combination data set of the circuit to be optimized;

[0015] Several PVT conditions are preset and represented by one-hot encoding to obtain the PVT vector encoding.

[0016] In some embodiments, the step of assigning weights to the parameter sample combination data set of the circuit to be optimized by using an adaptive sampling strategy to obtain a circuit parameter sample combination data set with a weight coefficient includes:

[0017] Performing adaptive weight calculation on the parameter sample combination data set of the circuit to be optimized to obtain a preliminary circuit parameter sample combination data set with weight coefficients;

[0018] Performing weighted calculation on extreme parameters in the parameter sample combination data set of the circuit to be optimized to obtain a weighted calculation result of the extreme parameters;

[0019] The preliminary circuit parameter sample combination data set with weight coefficients is combined with the extreme parameter weighted calculation result to perform a final sampling weight calculation to obtain the circuit parameter sample combination data set with weight coefficients.

[0020] In some embodiments, the expression for performing adaptive weight calculation on the parameter sample combination data set of the circuit to be optimized is specifically as follows:

[0021]

[0022] In the above formula, represents the calculated adaptive weight of target i, L i (θ (t) ) represents the loss function value of target i, L j (θ (t) ) represents the loss function value of target j, θ (t) represents the model parameters at the tth iteration, t represents the number of parameter iterations, m represents the total number of targets, i represents the index of the target, and j represents the index of the target.

[0023] In some embodiments, the expression for weighted calculation of the extreme parameters in the parameter sample combination data set of the circuit to be optimized is specifically as follows:

[0024]

[0025] In the above formula, w l (l k，j ) represents the extreme parameter weight value of length, w w (w k，j ) represents the extreme parameter weight value of width, w(z k ) represents the total weight of extreme parameters, l k，j Indicates the length value of the k-th sample on the j-th feature, l min，j Indicates the minimum length value of the j-th feature, l max，j Indicates the maximum length value of the j-th feature, w k，j Indicates the width value of the kth sample on the jth feature, w min，j Indicates the minimum width value of the j-th feature, w max，j Indicates the maximum width value of the jth feature, z k represents the kth sample, n represents the total number of features, k represents the index of the sample, and j represents the index of the feature.

[0026] In some embodiments, the expression for performing final sampling weight calculation by combining the preliminary circuit parameter sample combination data set with weight coefficients with the extreme parameter weighted calculation result is specifically as follows:

[0027]

[0028] In the above formula, Represents the final sampling weight value, represents the adaptive weight, L i (θ (t) ;z k ) represents the value of the loss function of target i, L i (θ (t) ;z j ) represents the value of the loss function of target j, w(z k ) represents the extreme parameter weight value of sample k, w(z j ) represents the extreme parameter weight value of sample j, z k represents the kth sample, z j represents the jth sample, θ (t) represents the model parameters of the tth iteration, N represents the total number of samples, m represents the total number of features, i represents the index of the feature, j represents the index of the sample, and t represents the parameters of the iteration.

[0029] In some embodiments, the circuit PVT performance prediction model includes a self-attention module, a KAN network structure and a CNN network structure, the output end of the self-attention module is connected to the input end of the KAN network structure, and the output end of the KAN network structure is connected to the input end of the CNN network structure.

[0030] In some embodiments, the circuit parameter sample combined data set with weight coefficients and the PVT vector code are predicted based on the circuit PVT performance prediction model to obtain a circuit PVT performance prediction data table, including:

[0031] Inputting the circuit parameter sample combined data set with weight coefficients and the PVT vector code into the circuit PVT performance prediction model;

[0032] Based on the self-attention module of the circuit PVT performance prediction model, self-attention feature extraction is performed on the circuit parameter sample combination data set with weight coefficients to obtain circuit parameter sample combination feature data;

[0033] Based on the KAN network structure of the circuit PVT performance prediction model, nonlinear transformation and addition calculation are performed on the circuit parameter sample combination characteristic data to obtain the circuit parameter standard performance prediction result;

[0034] Based on the CNN network structure of the circuit PVT performance prediction model, the circuit parameter standard performance prediction results and the PVT vector encoding are sequentially concatenated and convolved to obtain the circuit PVT performance prediction data table.

[0035] In some embodiments, the self-attention module based on the circuit PVT performance prediction model performs self-attention feature extraction on the circuit parameter sample combination data set with weight coefficients to obtain circuit parameter sample combination feature data, including:

[0036] Inputting the circuit parameter sample combined data set with weight coefficients into the self-attention module of the circuit PVT performance prediction model;

[0037] Performing sensitivity analysis on the circuit parameter sample combination data set with weight coefficients to obtain loss gradient weight values ​​of the circuit parameters;

[0038] Normalizing the loss gradient weight value of the circuit parameter to obtain a normalized loss gradient weight value of the circuit parameter;

[0039] A multi-head self-attention mechanism is introduced to perform self-attention calculation on the loss gradient weight value of the normalized circuit parameters to obtain the circuit parameter sample combination feature data.

[0040] To achieve the above purpose, another aspect of the embodiment of the present application proposes a circuit PVT performance optimization system based on KAN-CNN integration, the system comprising:

[0041] The first module is used to construct a parameter sample combination data set of the circuit to be optimized and obtain the PVT vector encoding;

[0042] The second module is used to assign weights to the parameter sample combination data set of the circuit to be optimized through an adaptive sampling strategy to obtain a circuit parameter sample combination data set with a weight coefficient;

[0043] The third module is used to combine the self-attention module, KAN network structure and CNN network structure to build a circuit PVT performance prediction model;

[0044] A fourth module is used to predict the circuit parameter sample combination data set with weight coefficients and the PVT vector code based on the circuit PVT performance prediction model to obtain a circuit PVT performance prediction data table;

[0045] The fifth module is used to optimize the circuit PVT performance according to the circuit PVT performance prediction data table to obtain an optimized circuit.

[0046] The embodiments of the present application include at least the following beneficial effects: The present application provides a circuit PVT performance optimization method and system based on KAN-CNN integration. The scheme constructs a parameter sample combination data set of the circuit to be optimized and obtains the PVT vector code, and further assigns weights to the parameter sample combination data set of the circuit to be optimized through an adaptive sampling strategy, so as to give priority to processing samples that have a greater impact on circuit performance, greatly reduce the number of samples and time required for training, and accelerate the convergence speed of the model. Then, the self-attention module, KAN network structure and CNN network structure are combined to construct a circuit PVT performance prediction model, and the circuit parameter sample combination data set with weight coefficients and PVT vector code are predicted. The multi-head self-attention mechanism is used to effectively capture the complex nonlinear relationship between the key parameters of the circuit and the performance indicators, thereby improving the prediction accuracy of the model. The hierarchical multi-task learning framework is adopted to share the common features extracted by the convolution layer under different PVT angles, so as to realize the simultaneous modeling and optimization of multiple performance indicators. Finally, the circuit PVT performance is optimized according to the circuit PVT performance prediction data table to obtain the optimized circuit. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of a circuit PVT performance optimization method based on KAN-CNN integration provided in an embodiment of the present application;

[0048] Figure 2It is a structural schematic diagram of a circuit PVT performance optimization system based on KAN-CNN integration provided in an embodiment of the present application;

[0049] Figure 3 It is a schematic diagram of the flow chart of circuit PVT performance prediction provided by the embodiment of the present application;

[0050] Figure 4 It is a data processing schematic diagram of the circuit PVT performance prediction model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present application. They are only examples of systems and methods consistent with some aspects of the embodiments of the present application as detailed in the attached claims.

[0052] It is understood that the terms "first", "second", etc. used in this application can be used to describe various concepts in this article, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another concept. For example, without departing from the scope of the embodiment of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein can be interpreted as "at the time of" or "when" or "in response to determination".

[0053] The terms "at least one", "multiple", "each", "any", etc. used in this application, at least one includes one, two or more, multiple includes two or more, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0055] Reference Figure 1 , Figure 1 A flowchart of a circuit PVT performance optimization method based on KAN-CNN integration provided by an embodiment of the present invention, referring to Figure 1 , the method comprises the following steps:

[0056] S100, constructing a parameter sample combination data set of the circuit to be optimized and obtaining a PVT vector code;

[0057] It should be noted that, in some embodiments, step S100 may include: S110, obtaining several adjustable parameters of the circuit to be optimized, and defining several value ranges of the adjustable parameters; S120, dividing the value ranges of the several adjustable parameters to obtain the adjustable parameter value ranges after equal division; S130, performing random sampling based on the adjustable parameter value ranges after equal division, and performing permutation and combination processing on the sampling results to obtain a parameter sample combination data set of the circuit to be optimized; S140, presetting several PVT conditions and performing one-hot encoding representation to obtain the PVT vector encoding.

[0058] In some specific embodiments, the parameter space of the circuit includes multiple adjustable parameters, such as capacitance value, resistance value, transistor size, etc. The value range of each parameter is divided into N equal segments (where N is the number of samples), and a value is randomly selected from each segment. All parameter combinations are arranged to ensure that the values ​​of all parameters are selected at least once during the sampling process, thereby obtaining a parameter sample combination data set of the circuit to be optimized.

[0059] Furthermore, the PVT conditions are vectorized and encoded. Assume that there are three combinations of PVT:

[0060] SF (Slow-Fast), 1.8V, 27℃;

[0061] SS(Slow-Slow), 1.6V, -40℃;

[0062] FF (Fast-Fast), 1.8V, 125℃.

[0063] These conditions can be expressed in the form of One-Hot encoding, for example: 0,1,0,0,0, 0,0,1,0,0, and finally the PVT vector encoding is obtained.

[0064] S200, assigning weights to the parameter sample combination data set of the circuit to be optimized through an adaptive sampling strategy to obtain a circuit parameter sample combination data set with a weight coefficient;

[0065] It should be noted that, in some embodiments, step S200 may include:

[0066] S210, performing adaptive weight calculation on the parameter sample combination data set of the circuit to be optimized to obtain a preliminary circuit parameter sample combination data set with weight coefficients;

[0067] In this embodiment, in order to better utilize data during the training process, the embodiment of the present invention introduces an adaptive sampling strategy, combining multi-objective weights with extreme parameter weighting. It should be noted that adaptive sampling dynamically adjusts the sampling point strategy and selects subsequent sample positions based on the results of previous sampling data, making the optimization process more efficient. Such a strategy can increase the sampling frequency in areas with large performance changes or more critical areas, and reduce the sampling frequency in areas with small changes or unimportant areas.

[0068] First, adaptive weight calculation is performed. Multi-objective optimization usually involves multiple performance indicators (such as power consumption, gain, frequency band, linearity, etc.), and the relative importance of each objective needs to be determined. By assigning weights to each objective, the degree of attention paid by the optimization algorithm to different objectives can be controlled. The expression is:

[0069]

[0070] In the above formula, represents the calculated adaptive weight of target i, L i (θ (t) ) represents the loss function value of target i, L j (θ (t) ) represents the loss function value of target j, θ (t) represents the model parameters at the tth iteration, t represents the number of parameter iterations, m represents the total number of targets, i represents the index of the target, and j represents the index of the target.

[0071] S220, performing weighted calculation on the extreme parameters in the parameter sample combination data set of the circuit to be optimized to obtain a weighted calculation result of the extreme parameters;

[0072] In this embodiment, extreme parameter weighting is further performed. In circuit design, extreme values ​​of certain parameters may have a great impact on performance. The extreme parameter weighting strategy emphasizes paying special attention to parameters that perform abnormally during the simulation process (such as extremely high / low gain, delay, power consumption, etc.). By increasing the weights of these parameters, it is possible to ensure that the circuit performance is kept in a good state under extreme conditions. The expression is:

[0073]

[0074] In the above formula, w l (l k，j ) represents the extreme parameter weight value of length, w w (w k，j ) represents the extreme parameter weight value of width, w(z k ) represents the total weight of extreme parameters, l k，j Indicates the length value of the k-th sample on the j-th feature, lmin,j Indicates the minimum length value of the j-th feature, l max,j Indicates the maximum length value of the j-th feature, w k，j Indicates the width value of the kth sample on the jth feature, w min,j Indicates the minimum width value of the j-th feature, w max，j Indicates the maximum width value of the jth feature, z k represents the kth sample, n represents the total number of features, k represents the index of the sample, and j represents the index of the feature.

[0075] S230 , performing final sampling weight calculation on the preliminary circuit parameter sample combination data set with weight coefficients and the extreme parameter weighted calculation result to obtain the circuit parameter sample combination data set with weight coefficients.

[0076] In this embodiment, the final sampling weight calculation is finally performed, and its expression is:

[0077]

[0078] In the above formula, Represents the final sampling weight value, represents the adaptive weight, L i (θ (t) ;z k ) represents the value of the loss function of target i, L i (θ (t) ;z j ) represents the value of the loss function of target j, w(z k ) represents the extreme parameter weight value of sample k, w(z j ) represents the extreme parameter weight value of sample j, z k represents the kth sample, z j represents the jth sample, θ (t) represents the model parameters of the tth iteration, N represents the total number of samples, m represents the total number of features, i represents the index of the feature, j represents the index of the sample, and t represents the parameters of the iteration.

[0079] S300, combining the self-attention module, KAN network structure and CNN network structure to build a circuit PVT performance prediction model;

[0080] It should be noted that, in some embodiments, the circuit PVT performance prediction model includes a self-attention module, a KAN network structure and a CNN network structure, the output end of the self-attention module is connected to the input end of the KAN network structure, and the output end of the KAN network structure is connected to the input end of the CNN network structure.

[0081] Among them, the self-attention module combines parameter sensitivity weighting and adopts a multi-head attention mechanism. The dimension of each attention head is 15, a total of 3 heads are used, and residual connections are used.

[0082] The KAN network consists of learnable B-spline functions with 40 neurons in each layer and a grid size of 200.

[0083] The CNN network contains several convolutional layers, batch normalization layers and SiLU activation functions; the convolutional kernel size of the convolutional layer is 3 and the stride is 1.

[0084] S400, predicting the circuit parameter sample combination data set with weight coefficients and the PVT vector code based on the circuit PVT performance prediction model to obtain a circuit PVT performance prediction data table;

[0085] It should be noted that in some embodiments, Figure 4 As shown, step S400 may include:

[0086] S410, inputting the circuit parameter sample combined data set with weight coefficients and the PVT vector code into the circuit PVT performance prediction model;

[0087] S420, based on the self-attention module of the circuit PVT performance prediction model, performing self-attention feature extraction on the circuit parameter sample combination data set with weight coefficients to obtain circuit parameter sample combination feature data;

[0088] Specifically, the circuit parameter sample combination data set with weight coefficients is input into the self-attention module of the circuit PVT performance prediction model; a sensitivity analysis is performed on the circuit parameter sample combination data set with weight coefficients to obtain the loss gradient weight values ​​of the circuit parameters; the loss gradient weight values ​​of the circuit parameters are normalized to obtain the normalized loss gradient weight values ​​of the circuit parameters; a multi-head self-attention mechanism is introduced to perform self-attention calculation on the normalized loss gradient weight values ​​of the circuit parameters to obtain the circuit parameter sample combination feature data.

[0089] It should be noted that in the self-attention module, the network processes the input parameters through a multi-head self-attention mechanism. Each input sample is processed in parallel by three attention heads, and the dimension of each head is 15. After the parameters are weighted by the attention mechanism, the residual connection is used to add the input and output.

[0090] In some embodiments, a parameter sensitivity analysis is first performed to calculate the gradient of the loss function for each input parameter, and the expression is:

[0091]

[0092] In the above formula, S i represents the parameter sensitivity weight of the i-th feature, Represents the value of the loss function, x i Represents the parameter value of the i-th feature.

[0093] The sensitivity weight is further normalized, and its expression is:

[0094]

[0095] In the above formula, represents the normalized parameter sensitivity weight, S i represents the parameter sensitivity weight of the i-th feature, n represents the total number of features, and i represents the index of the feature.

[0096] Furthermore, the self-attention calculation of the sensitivity weight is introduced, and its expression is:

[0097]

[0098] In the above formula, Attention(·) represents the calculation of the self-attention mechanism, Q represents the query matrix, K represents the key matrix, V represents the value matrix, S represents the sensitivity weight matrix, softmax(·) represents the normalization calculation, ⊙ represents the element-by-element multiplication, and d k Indicates the dimension of the key.

[0099] Finally, a multi-head self-attention mechanism is introduced. Multiple parallel self-attention heads are used to capture different subspaces of input features. Its expression is:

[0100] MultiHead(Q,K,V,S)=Concat(head 1 , head 2 , …, head h )W 0

[0101] In the above formula, MultiHead(·) represents the calculation of the multi-head self-attention mechanism, Concat(·) represents concatenating the outputs of multiple attention heads, and head h represents the output of the h-th attention head, W 0 The parameter matrix representing the final linear transformation used to integrate the outputs of all attention heads and map them to the target space.

[0102] S430, based on the KAN network structure of the circuit PVT performance prediction model, performing nonlinear transformation and addition calculation on the circuit parameter sample combination characteristic data to obtain a circuit parameter standard performance prediction result;

[0103] In this embodiment, in the standard performance prediction, the KAN model is used to model the typical performance of the circuit and generate basic performance prediction data. The weight of the training data is adjusted through an adaptive sampling strategy to improve the accuracy of key performance. After the model is trained, a standard typical performance prediction data table is output to provide basic data for subsequent PVT modeling, that is, in the KAN network, the circuit parameter sample combination feature data obtained in the previous step is nonlinearly transformed and added through the learned B-spline function, and the standard performance prediction result is output.

[0104] S440. Based on the CNN network structure of the circuit PVT performance prediction model, the circuit parameter standard performance prediction result and the PVT vector code are sequentially concatenated and convolved to obtain the circuit PVT performance prediction data table.

[0105] In some specific embodiments, the typical performance data obtained above, i.e., the circuit parameter standard performance prediction results, are combined with the PVT conditional coding and input into the CNN model. Through the multi-task learning framework, the CNN model trains the performance data under various PVT combinations to generate performance prediction results including those under PVT perturbations.

[0106] That is, the comprehensive input vector obtained by concatenating the standard performance prediction results with the different PVT condition codes is input into CNN, the standard performance and PVT condition codes are processed, and the performance prediction results under different PVT conditions are generated.

[0107] S500, optimizing the circuit PVT performance according to the circuit PVT performance prediction data table to obtain an optimized circuit;

[0108] It should be noted that, in some embodiments, the SPICE simulation tool is used to simulate the analog circuit to obtain the performance data of the circuit under different PVT (process, voltage, temperature) conditions. Set the sampling parameter combination and PVT conditions in the SPICE tool. Simulate each combination separately and extract performance indicators, such as: Gain: the circuit's ability to amplify the signal. Bandwidth: the frequency range supported by the circuit at a specified gain. Power consumption: the total power consumption of the circuit in the working state. Obtain a data table, where each row corresponds to a parameter combination and a performance indicator under PVT conditions.

[0109] In summary, if Figure 3As shown, by introducing an improved adaptive sampling strategy and an extreme parameter weighting mechanism, the embodiment of the present invention enables the model to prioritize samples that have a greater impact on circuit performance, greatly reducing the number of samples and time required for training, and accelerating the convergence speed of the model. Secondly, the use of a sensitivity-guided multi-head self-attention mechanism can effectively capture the complex nonlinear relationship between key circuit parameters and performance indicators, thereby improving the prediction accuracy of the model. In addition, the present invention adopts a hierarchical multi-task learning framework to share the common features extracted by the convolutional layer under different PVT angles, thereby realizing the simultaneous modeling and optimization of multiple performance indicators. This architecture enables the model to exhibit stable prediction capabilities under multiple PVT angles and significantly enhances the generalization ability of the model.

[0110] See also Figure 2 The embodiment of the present application further provides a circuit PVT performance optimization system based on KAN-CNN integration, which can implement the above-mentioned circuit PVT performance optimization method based on KAN-CNN integration, and the system includes:

[0111] The first module 201 is used to construct a parameter sample combination data set of the circuit to be optimized and obtain the PVT vector code;

[0112] The second module 202 is used to assign weights to the parameter sample combination data set of the circuit to be optimized through an adaptive sampling strategy to obtain a circuit parameter sample combination data set with a weight coefficient;

[0113] The third module 203 is used to combine the self-attention module, the KAN network structure and the CNN network structure to build a circuit PVT performance prediction model;

[0114] The fourth module 204 is used to predict the circuit parameter sample combination data set with weight coefficients and the PVT vector code based on the circuit PVT performance prediction model to obtain a circuit PVT performance prediction data table;

[0115] The fifth module 205 is used to optimize the circuit PVT performance according to the circuit PVT performance prediction data table to obtain an optimized circuit.

[0116] It can be understood that the contents of the above method embodiments are all applicable to the present system embodiments, the functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0117] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A circuit PVT performance optimization method based on KAN-CNN integration, characterized in that: The method comprises the following steps: Construct a parameter sample combination data set of the circuit to be optimized and obtain the PVT vector encoding; Assigning weights to the parameter sample combination data set of the circuit to be optimized through an adaptive sampling strategy to obtain a circuit parameter sample combination data set with a weight coefficient; Combining the self-attention module, KAN network structure and CNN network structure, a circuit PVT performance prediction model is constructed; Based on the circuit PVT performance prediction model, the circuit parameter sample combined data set with weight coefficients and the PVT vector code are predicted to obtain a circuit PVT performance prediction data table; The circuit PVT performance is optimized according to the circuit PVT performance prediction data table to obtain an optimized circuit.

2. The method according to claim 1, characterized in that The step of constructing a parameter sample combination data set of the circuit to be optimized and obtaining a PVT vector code includes: Acquiring several adjustable parameters of the circuit to be optimized, and defining value ranges of several of the adjustable parameters; Dividing the value ranges of the plurality of adjustable parameters to obtain equally divided value ranges of the adjustable parameters; Based on the adjustable parameter value range after the equal division, random sampling is performed, and the sampling results are processed by permutation and combination to obtain a parameter sample combination data set of the circuit to be optimized; Several PVT conditions are preset and represented by one-hot encoding to obtain the PVT vector encoding.

3. The method according to claim 1, characterized in that The step of assigning weights to the parameter sample combination data set of the circuit to be optimized by using an adaptive sampling strategy to obtain a circuit parameter sample combination data set with a weight coefficient includes: Performing adaptive weight calculation on the parameter sample combination data set of the circuit to be optimized to obtain a preliminary circuit parameter sample combination data set with weight coefficients; Performing weighted calculation on extreme parameters in the parameter sample combination data set of the circuit to be optimized to obtain a weighted calculation result of the extreme parameters; The preliminary circuit parameter sample combination data set with weight coefficients is combined with the extreme parameter weighted calculation result to perform a final sampling weight calculation to obtain the circuit parameter sample combination data set with weight coefficients.

4. The method according to claim 3, characterized in that The expression for performing adaptive weight calculation on the parameter sample combination data set of the circuit to be optimized is specifically as follows: In the above formula, represents the calculated adaptive weight of target i, L i (θ (t) ) represents the loss function value of target i, L j (θ (t) ) represents the loss function value of target j, θ (t) represents the model parameters at the tth iteration, t represents the number of parameter iterations, m represents the total number of targets, i represents the index of the target, and h represents the index of the target.

5. The method according to claim 3, characterized in that: The expression for weighted calculation of the extreme parameters in the parameter sample combination data set of the circuit to be optimized is specifically as follows: In the above formula, w l (l k,j ) represents the extreme parameter weight value of length, w w (w k,j ) represents the extreme parameter weight value of width, w(z k ) represents the total weight of extreme parameters, l k,j Indicates the length value of the k-th sample on the j-th feature, l min,j Indicates the minimum length value of the j-th feature, l max,j Indicates the maximum length value of the j-th feature, w k,j represents the width value of the kth sample on the hth feature, w min,j Indicates the minimum width value of the j-th feature, w max,j Indicates the maximum width value of the jth feature, z k represents the kth sample, n represents the total number of features, k represents the index of the sample, and j represents the index of the feature.

6. The method according to claim 3, characterized in that: The expression for performing final sampling weight calculation by combining the preliminary circuit parameter sample combination data set with weight coefficients with the extreme parameter weighted calculation result is specifically as follows: In the above formula, Represents the final sampling weight value, represents the adaptive weight, L i (θ (t) ;z k ) represents the value of the loss function of target i, L i (θ (t) ;z j ) represents the value of the loss function of target j, w(z k ) represents the extreme parameter weight value of sample k, w(z j ) represents the extreme parameter weight value of sample j, z k represents the kth sample, z j represents the jth sample, θ (t) represents the model parameters of the tth iteration, N represents the total number of samples, m represents the total number of features, i represents the index of the feature, j represents the index of the sample, and t represents the parameters of the iteration.

7. The method according to claim 1, characterized in that The circuit PVT performance prediction model includes a self-attention module, a KAN network structure and a CNN network structure. The output end of the self-attention module is connected to the input end of the KAN network structure, and the output end of the KAN network structure is connected to the input end of the CNN network structure.

8. The method according to claim 1, characterized in that The circuit parameter sample combined data set with weight coefficients and the PVT vector code are predicted based on the circuit PVT performance prediction model to obtain a circuit PVT performance prediction data table, including: Inputting the circuit parameter sample combined data set with weight coefficients and the PVT vector code into the circuit PVT performance prediction model; Based on the self-attention module of the circuit PVT performance prediction model, self-attention feature extraction is performed on the circuit parameter sample combination data set with weight coefficients to obtain circuit parameter sample combination feature data; Based on the KAN network structure of the circuit PVT performance prediction model, nonlinear transformation and addition calculation are performed on the circuit parameter sample combination characteristic data to obtain the circuit parameter standard performance prediction result; Based on the CNN network structure of the circuit PVT performance prediction model, the circuit parameter standard performance prediction results and the PVT vector encoding are sequentially concatenated and convolved to obtain the circuit PVT performance prediction data table.

9. The method according to claim 8, characterized in that The self-attention module based on the circuit PVT performance prediction model performs self-attention feature extraction on the circuit parameter sample combination data set with weight coefficients to obtain circuit parameter sample combination feature data, including: Inputting the circuit parameter sample combined data set with weight coefficients into the self-attention module of the circuit PVT performance prediction model; Performing sensitivity analysis on the circuit parameter sample combination data set with weight coefficients to obtain loss gradient weight values ​​of the circuit parameters; Normalizing the loss gradient weight value of the circuit parameter to obtain a normalized loss gradient weight value of the circuit parameter; A multi-head self-attention mechanism is introduced to perform self-attention calculation on the loss gradient weight value of the normalized circuit parameters to obtain the circuit parameter sample combination feature data.

10. A circuit PVT performance optimization system based on KAN-CNN integration, characterized in that: The system comprises: The first module is used to construct a parameter sample combination data set of the circuit to be optimized and obtain the PVT vector encoding; The second module is used to assign weights to the parameter sample combination data set of the circuit to be optimized through an adaptive sampling strategy to obtain a circuit parameter sample combination data set with a weight coefficient; The third module is used to combine the self-attention module, KAN network structure and CNN network structure to build a circuit PVT performance prediction model; A fourth module is used to predict the circuit parameter sample combination data set with weight coefficients and the PVT vector code based on the circuit PVT performance prediction model to obtain a circuit PVT performance prediction data table; The fifth module is used to optimize the circuit PVT performance according to the circuit PVT performance prediction data table to obtain an optimized circuit.

Images