A method and system for evaluating integrated circuits

Abstract

The present invention relates to a method, system, device and storage medium for evaluating integrated circuits. The method includes: obtaining the circuit topology and process parameters of the integrated circuit to be evaluated; constructing an equivalent circuit model of the integrated circuit based on the circuit topology; generating multiple sets of predicted process parameter fluctuation data according to the current process parameters and production environment factor data; applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters; analyzing the multiple sets of performance parameters and extracting key reliability features; and calculating the reliability score of the integrated circuit based on the extracted key reliability features. By constructing an equivalent circuit model and using a process parameter fluctuation prediction model, the method can accurately evaluate the performance of the integrated circuit under different process conditions, thereby providing a reliability score and providing an important reference for the design and optimization of the integrated circuit.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated circuits, and in particular to a method and system for evaluating integrated circuits. Background Art

[0002] In the prior art, reliability assessment of integrated circuits usually relies on physical testing and historical data analysis methods. Physical testing methods assess the reliability of integrated circuits by actually manufacturing and measuring the performance of integrated circuits under various environmental and operating conditions. Historical data analysis methods use accumulated failure data and experience to predict the reliability of newly designed integrated circuits. These methods play an important role in assessing the reliability of integrated circuits, helping engineers estimate the life and stability of products during the design stage.

[0003] However, the existing technologies mainly have the following problems: physical testing methods are time-consuming and costly, and cannot quickly feedback design problems in the early stages of design; historical data analysis methods rely on a large amount of failure data and experience, and are difficult to adapt to new materials, new processes, and newly designed integrated circuits. These limitations make the existing reliability assessment methods unable to provide sufficient flexibility and accuracy when facing complex and changeable modern integrated circuit designs.

[0004] Therefore, there is an urgent need to develop a new method for evaluating integrated circuits. Summary of the invention

[0005] The present application provides a method and system for evaluating integrated circuits to improve the accuracy of integrated circuit reliability assessment.

[0006] The present application provides a method for evaluating an integrated circuit, comprising:

[0007] Obtaining circuit topology and process parameters of the integrated circuit to be evaluated;

[0008] Based on the circuit topology, construct an equivalent circuit model of the integrated circuit;

[0009] Based on the current process parameters and production environment factor data, a pre-trained process parameter fluctuation prediction model is used to generate multiple sets of predicted process parameter fluctuation data;

[0010] Applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters;

[0011] Analyzing the multiple groups of performance parameters to extract key reliability features;

[0012] Based on the extracted key reliability features, a reliability score of the integrated circuit is calculated.

[0013] Furthermore, the process parameter fluctuation prediction model using pre-training includes a data pre-processing module, a multi-modal deep fusion module and an uncertainty assessment module;

[0014] The data preprocessing module is used to preprocess the current process parameters and production environment factor data to obtain preprocessed process parameter data and production environment factor data, and the preprocessing includes cleaning, normalization and data enhancement;

[0015] The multimodal deep fusion module includes a feature extraction subnetwork, a feature fusion subnetwork and a prediction subnetwork; the feature extraction subnetwork extracts features from the preprocessed process parameter data and production environment factor data through multiple convolutional layers to obtain process parameter features and production environment factor features, and in the process of feature extraction, the effectiveness of feature extraction is improved through the attention mechanism layer; the feature fusion subnetwork uses a multi-head attention mechanism to perform weighted fusion on the process parameter features and production environment factor features to generate fusion features, and in the process of weighted fusion, the integrity and effectiveness of the features are maintained through residual connections; the prediction subnetwork adopts a hybrid architecture of an autoregressive model and a variational autoencoder to perform time series prediction and generation on the fusion features to obtain predicted process parameter fluctuation data;

[0016] The uncertainty assessment module adopts the Bayesian deep learning method to perform multiple sampling on the predicted process parameter fluctuation data to generate predicted confidence intervals, and obtain multiple groups of predicted process parameter fluctuation data.

[0017] Furthermore, the feature extraction subnetwork performs feature extraction through a multi-scale convolutional neural network, which includes convolutional layers with different receptive fields to capture process parameters and production environment factor characteristics at different scales, thereby improving the accuracy and robustness of feature extraction.

[0018] Furthermore, the uncertainty assessment module is specifically used for:

[0019] The process parameter fluctuation data output by the prediction sub-network is modeled in Bayesian fashion, and the prediction results are regarded as parameters of probability distribution;

[0020] By using a Bayesian deep learning method, the probability distribution is randomly sampled multiple times to generate multiple possible process parameter fluctuation samples, and each sampling obtains a set of predicted process parameter fluctuation data;

[0021] Perform statistical analysis on the process parameter fluctuation data obtained from each set of sampling, calculate the mean and variance of each set of data, and form a predicted confidence interval;

[0022] The process parameter fluctuation data and statistical results obtained by multiple samplings are stored as multiple groups of predicted process parameter fluctuation data, and distribution characteristic diagrams of the process parameter fluctuations are generated based on these data;

[0023] Combining historical process data with current forecast data, comparative analysis is performed on each group of predicted process parameter fluctuation data to identify possible abnormal fluctuations and mark them in the final forecast results.

[0024] Furthermore, the analyzing the multiple groups of performance parameters to extract key reliability features includes:

[0025] Each set of performance parameters is represented as a vector ,in , indicating the The simulation results of the group performance parameters;

[0026] According to the following formula 1 and formula 2, the mean vector of the corresponding vector of each group of performance parameters is calculated: and the covariance matrix :

[0027]

[0028]

[0029] in, is the number of groups for circuit simulation:

[0030] According to the following formula 3, the covariance matrix is ​​analyzed using principal component analysis Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector :

[0031]

[0032] From the characteristic value and the corresponding eigenvector Select the front The eigenvectors corresponding to the largest eigenvalues ​​form the characteristic matrix ;

[0033] According to the following formula 4, the initial key feature vector of each group of performance parameters is calculated: :

[0034]

[0035] According to the following formula 5, the initial key feature vector is calculated The corresponding regularized eigenvector ;

[0036]

[0037] According to the following formula 6 and formula 7, the regularized mean vector is calculated and the regularized standard deviation vector ;

[0038]

[0039]

[0040] in, is the number of groups for circuit simulation;

[0041] The regularized mean vector and the regularized standard deviation vector , identified as a key reliability feature.

[0042] Furthermore, the step of calculating the reliability score of the integrated circuit based on the extracted key reliability features includes:

[0043] According to the following formula 8, the reliability score of the integrated circuit is calculated :

[0044]

[0045] in, is the regularized mean vector No. Quantity; is the regularized standard deviation vector No. Quantity; It is Weight factors; is a constant; and is the adjustment parameter; is the regularized mean vector Middle The target mean of the components; is the selected front in principal component analysis The number of key features.

[0046] Furthermore, the method for evaluating an integrated circuit further includes:

[0047] Calculated reliability score With a pre-set reliability threshold Make comparisons;

[0048] if , then it is judged that the integrated circuit is reliable;

[0049] if , then it is judged that the integrated circuit is unreliable.

[0050] The present application provides a device for evaluating an integrated circuit, comprising:

[0051] An acquisition unit, used for acquiring a circuit topology and process parameters of the integrated circuit to be evaluated;

[0052] A construction unit, used to construct an equivalent circuit model of the integrated circuit based on the circuit topology structure;

[0053] A generating unit, for generating multiple sets of predicted process parameter fluctuation data using a pre-trained process parameter fluctuation prediction model based on current process parameters and production environment factor data;

[0054] A simulation unit, used for applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters;

[0055] An extraction unit, used to analyze the multiple groups of performance parameters and extract key reliability features;

[0056] The calculation unit is used to calculate the reliability score of the integrated circuit based on the extracted key reliability features.

[0057] The present application provides a system for evaluating an integrated circuit, comprising:

[0058] A server including a central processing unit, memory and storage device for running the integrated circuit reliability assessment software;

[0059] A data acquisition module, used to obtain the circuit topology structure, process parameters and production environment factor data of the integrated circuit to be evaluated, and transmit the data to the server;

[0060] A remote access module configured to allow a user to remotely access and operate the server through a network interface to input information of the integrated circuit to be evaluated and view evaluation results;

[0061] A data processing module is configured in the server, and the data processing module includes:

[0062] The data preprocessing submodule is used to clean, normalize and enhance the received process parameter and production environment factor data;

[0063] A model building submodule, used to build an equivalent circuit model of the integrated circuit based on the circuit topology;

[0064] The prediction submodule is used to generate multiple sets of predicted process parameter fluctuation data based on the current process parameters and production environment factor data using a pre-trained process parameter fluctuation prediction model;

[0065] A simulation submodule, used for applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters;

[0066] A feature analysis submodule, used to analyze the multiple groups of performance parameters and extract key reliability features;

[0067] A score calculation submodule, used to calculate the reliability score of the integrated circuit based on the extracted key reliability features;

[0068] The result storage and display module is configured in the server and is used to store and display the reliability evaluation results, including:

[0069] The data storage submodule is used to store all input data, process intermediate data, and final evaluation results;

[0070] The result display submodule is used to generate and display visual results including reliability scoring and key feature analysis;

[0071] The alarm and notification module is configured to automatically send alarm notifications to designated users when the reliability score falls below a preset threshold.

[0072] The present application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the content of any one of the above methods is implemented.

[0073] The beneficial effects of this application include: (1) By obtaining the circuit topology and process parameters of the integrated circuit to be evaluated and constructing an equivalent circuit model, this method can accurately reflect the actual working state of the circuit during the simulation process, thereby improving the accuracy of reliability evaluation. (2) Using a pre-trained process parameter fluctuation prediction model, multiple sets of predicted process parameter fluctuation data are generated based on the current process parameters and production environment factor data, which can adapt to different process conditions and improve the flexibility and applicability of the evaluation method. (3) By analyzing multiple sets of performance parameters and extracting key reliability features, this method can comprehensively analyze the performance of integrated circuits under different process conditions, providing a more comprehensive and detailed basis for reliability evaluation. (4) Based on the extracted key reliability features, the reliability score of the integrated circuit is calculated. This method can quickly and effectively evaluate the reliability of the integrated circuit, provide timely feedback for design and optimization, and reduce development cycle and cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1This is a flow chart of a method for evaluating an integrated circuit provided in the first embodiment of the present application.

[0075] Figure 2 It is a schematic diagram of a device for evaluating an integrated circuit provided in the second embodiment of the present application. DETAILED DESCRIPTION

[0076] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0077] The first embodiment of the present application provides a method for evaluating an integrated circuit. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 A method for evaluating an integrated circuit is described in detail in the first embodiment of the present application.

[0078] Step S101: Obtaining the circuit topology and process parameters of the integrated circuit to be evaluated.

[0079] First, the circuit topology of the IC to be evaluated needs to be extracted from the design file. The design file usually includes a netlist of the circuit, which lists in detail the connection relationships of all components in the circuit (such as transistors, resistors, capacitors, etc.). When extracting the netlist, professional electronic design automation (EDA) software, such as Cadence, is required. These software can parse the design file and generate a detailed netlist that describes the circuit topology.

[0080] While obtaining the circuit topology, it is also necessary to obtain the process parameters of the integrated circuit. Process parameters refer to the physical and chemical parameters involved in the manufacturing process of integrated circuits, including but not limited to doping concentration, oxide layer thickness, transistor size, channel length, threshold voltage, etc. These parameters are usually provided by the integrated circuit manufacturer and included in the process design kit (PDK). The parameter files provided in the PDK can be obtained by reading files in a specific format (such as SPICE model files).

[0081] When implementing it, first open the design file and use EDA software to extract the circuit netlist. Then, read the process parameter file in the PDK and extract the key parameter information. In order to ensure the accuracy and completeness of the data, the extracted netlist and process parameters should be checked and verified.

[0082] For example, in Cadence software, you can open the design project and use the "Netlist Extraction" function to extract the circuit netlist. For process parameters, you can load the corresponding PDK file and use the "ProcessParameter Extraction" function to obtain various process parameters.

[0083] After the extraction is completed, the circuit topology and process parameters need to be converted into appropriate formats for use in subsequent steps. For example, the netlist can be converted into a standardized SPICE format file, and the process parameters can be organized into an easily accessible database or configuration file.

[0084] Through the above detailed steps, the circuit topology and process parameters of the integrated circuit to be evaluated can be accurately and completely obtained, laying the foundation for the subsequent equivalent circuit model construction and simulation analysis.

[0085] Step S102: constructing an equivalent circuit model of the integrated circuit based on the circuit topology structure.

[0086] First, after obtaining the circuit topology and process parameters, it is necessary to select appropriate tools and methods to build an equivalent circuit model. An equivalent circuit model is a mathematical model used to simulate the behavior of an actual circuit and can exhibit characteristics similar to those of a real circuit in a simulation environment.

[0087] In order to build an equivalent circuit model, you can use the SPICE (Simulation Program with Integrated Circuit Emphasis) simulation tool, which is a simulation tool widely used in integrated circuit design and verification. First, you need to import the extracted circuit topology (netlist) into the SPICE tool. Usually, the netlist file format includes component types, component values, and connection relationships, such as parameters and node connection information of transistors, resistors, capacitors, etc.

[0088] After importing the netlist, the next step is to define the model parameters of each component. These parameters are mainly derived from the process parameter file, including the threshold voltage, mobility, parasitic capacitance, etc. of the transistor. These parameters can be obtained directly from the process design kit (PDK) and defined in the SPICE model file. By calling these parameter files in SPICE, it can be ensured that the equivalent circuit model accurately reflects the characteristics of the actual manufacturing process.

[0089] After defining the model parameters of the components, you also need to set the simulation control parameters, such as power supply voltage, input signal form and frequency, etc. These parameters are used to simulate the working environment and conditions of the actual circuit in order to obtain authentic and reliable results in the simulation.

[0090] Next, you need to use the SPICE simulation tool to generate a simulation netlist file of the equivalent circuit model. This file includes detailed information about all circuit components and their interconnections, and references the previously defined process parameter model. In the SPICE simulation tool, you can specify the type of simulation to be performed, such as DC analysis, AC analysis, transient analysis, etc., by writing a simulation control file.

[0091] After completing the writing of the simulation control file, run the SPICE simulation tool for preliminary simulation to verify the correctness of the equivalent circuit model. By viewing the simulation output results, such as voltage waveform, current waveform, frequency response, etc., confirm whether the model accurately reflects the actual performance of the circuit. If it is found that the simulation results deviate from expectations, it may be necessary to adjust the component model parameters or correct the circuit topology to ensure the accuracy of the model.

[0092] Finally, through multiple iterations and verifications, an accurate and reliable equivalent circuit model is obtained, which provides a basis for subsequent circuit simulation and performance analysis.

[0093] Step S103: Based on the current process parameters and production environment factor data, a pre-trained process parameter fluctuation prediction model is used to generate multiple sets of predicted process parameter fluctuation data.

[0094] First, it is necessary to collect the current process parameters and production environment factor data. These data can be obtained from the manufacturing process database, real-time monitoring system or production log. Process parameters usually include doping concentration, oxide layer thickness, transistor size, etc., while production environment factor data include temperature, humidity, pressure, etc. These data are the input of the process parameter fluctuation prediction model.

[0095] Next, the collected process parameters and production environment factor data need to be preprocessed to ensure the accuracy and consistency of the data. This step usually involves data cleaning, normalization, and data enhancement. Data cleaning is to remove noise and outliers, normalization is to convert data of different dimensions to the same scale, and data enhancement is to generate more training data to improve the robustness of the model.

[0096] The preprocessed data will be input into a pre-trained process parameter fluctuation prediction model. The model has been trained with a large amount of historical data and can accurately predict parameter fluctuations under different process conditions. The process parameter fluctuation prediction model is usually based on deep learning or machine learning algorithms, such as long short-term memory network (LSTM), convolutional neural network (CNN) or random forest.

[0097] During the model operation, the prediction time range and frequency need to be set. The model will generate prediction results of process parameter fluctuations at multiple time points based on the input data. Each set of prediction results represents the possible changes in process parameters at a specific time point. The model will output multiple sets of such prediction data for subsequent simulation and analysis.

[0098] After generating multiple sets of predicted process parameter fluctuation data, these data need to be verified to ensure their rationality and accuracy. This can be done by comparing with historical data or using some statistical methods to detect anomalies in the predicted data. If there is a large deviation in the predicted results, it may be necessary to adjust the model parameters or retrain the model.

[0099] Finally, the verified sets of predicted process parameter fluctuation data will be used in the subsequent circuit simulation steps.

[0100] Furthermore, the process parameter fluctuation prediction model using pre-training includes a data pre-processing module, a multi-modal deep fusion module and an uncertainty assessment module;

[0101] The data preprocessing module is used to preprocess the current process parameters and production environment factor data to obtain preprocessed process parameter data and production environment factor data, and the preprocessing includes cleaning, normalization and data enhancement;

[0102] The multimodal deep fusion module includes a feature extraction subnetwork, a feature fusion subnetwork and a prediction subnetwork; the feature extraction subnetwork extracts features from the preprocessed process parameter data and production environment factor data through multiple convolutional layers to obtain process parameter features and production environment factor features, and in the process of feature extraction, the effectiveness of feature extraction is improved through the attention mechanism layer; the feature fusion subnetwork uses a multi-head attention mechanism to perform weighted fusion on the process parameter features and production environment factor features to generate fusion features, and in the process of weighted fusion, the integrity and effectiveness of the features are maintained through residual connections; the prediction subnetwork adopts a hybrid architecture of an autoregressive model and a variational autoencoder to perform time series prediction and generation on the fusion features to obtain predicted process parameter fluctuation data;

[0103] The uncertainty assessment module adopts the Bayesian deep learning method to perform multiple sampling on the predicted process parameter fluctuation data to generate predicted confidence intervals, and obtain multiple groups of predicted process parameter fluctuation data.

[0104] First, the data preprocessing module is responsible for preprocessing the current process parameters and production environment factor data. The process includes cleaning, normalization, and data enhancement. The cleaning step is used to remove noise and outliers to ensure the accuracy of the data. For example, erroneous data in the acquisition process can be processed by detecting and removing parameter values ​​that are beyond the normal range. The normalization step converts the data to a uniform scale so that different types of data can be processed consistently in subsequent processing. A common method is to scale the data to between 0 and 1 or standardize it to a mean of 0 and a standard deviation of 1. Data enhancement expands the diversity of the data set and improves the robustness of the model by generating virtual data points. This can be achieved by making small random perturbations to the existing data to generate more diverse data samples.

[0105] After preprocessing, the data will be input into the multimodal deep fusion module. This module includes a feature extraction subnetwork, a feature fusion subnetwork, and a prediction subnetwork. The feature extraction subnetwork extracts features from the preprocessed process parameter data and production environment factor data through multiple convolutional layers. The convolutional layer can capture local patterns and important features in the data. For example, through a series of convolution and pooling operations, low-level features and high-level features of the data are extracted. In the process of feature extraction, the attention mechanism layer is introduced to improve the effectiveness of feature extraction. The attention mechanism can dynamically assign the importance of different features and emphasize the features that contribute most to reliability assessment.

[0106] The extracted features will be passed to the feature fusion sub-network, and the multi-head attention mechanism will be used to perform weighted fusion of process parameter features and production environment factor features. The multi-head attention mechanism captures the complex relationship between different features through multiple attention heads, and performs weighted processing to generate fused features. In the process of weighted fusion, the integrity and effectiveness of the features are maintained through residual connections. Residual connections can prevent information loss during the fusion process and ensure that all important features are retained and effectively utilized.

[0107] The fused features are passed to the prediction subnetwork, which uses a hybrid architecture of an autoregressive model and a variational autoencoder. The autoregressive model is used to model time series data and predict future process parameter fluctuations. The variational autoencoder is used to learn the latent distribution of the data and generate new data samples. By combining the two, the prediction subnetwork can capture both time dependency and data distribution characteristics, thereby generating more accurate process parameter fluctuation data.

[0108] Finally, the uncertainty assessment module uses the Bayesian deep learning method to sample the predicted process parameter fluctuation data multiple times and generate the predicted confidence interval. The Bayesian deep learning method can better handle the uncertainty of data by introducing Bayesian statistics to model the probability distribution of model parameters. In actual operation, by sampling the output of the model multiple times, multiple possible process parameter fluctuation data are generated, and their statistical characteristics are calculated, and finally multiple groups of predicted process parameter fluctuation data are obtained. These data can reflect the possible changes under different process conditions and provide rich information for reliability assessment.

[0109] Furthermore, the feature extraction subnetwork performs feature extraction through a multi-scale convolutional neural network, which includes convolutional layers with different receptive fields to capture process parameters and production environment factor characteristics at different scales, thereby improving the accuracy and robustness of feature extraction.

[0110] In the feature extraction stage, the multi-scale convolutional neural network (MSCNN) is used to process the pre-processed process parameters and production environment factor data. This network structure is uniquely designed and can capture features of different scales through convolutional layers with different receptive fields, improving the accuracy and robustness of feature extraction.

[0111] First, the preprocessed data is input into a multi-scale convolutional neural network. The network contains multiple convolutional layers, each with a different receptive field. The receptive field refers to the size of the area where the convolutional layer perceives the input data. Different receptive fields can capture features of different levels and ranges of the data. For example, a smaller receptive field can capture subtle local features, while a larger receptive field can capture global patterns.

[0112] Specifically, the first layer of a multi-scale convolutional neural network may include multiple convolution kernels with small receptive fields, and the convolution operation can extract subtle features in the data, such as local changes and small fluctuations. The second layer may include convolution kernels with medium receptive fields, which can capture features in a larger range, such as medium-scale patterns and trends. The third layer may include convolution kernels with larger receptive fields, which can identify global features, such as overall trends and global changes.

[0113] After each convolution layer, there is usually a pooling layer for dimensionality reduction and computational reduction. The pooling layer downsamples the convolution output, retains important features and discards redundant information. Common pooling operations include maximum pooling and average pooling. Maximum pooling retains the maximum value in each pooling window, while average pooling calculates the average of all values ​​in the window. Through the processing of the pooling layer, the size of the feature map can be further compressed and the efficiency of feature extraction can be improved.

[0114] In order to enhance the effectiveness of feature extraction, the attention mechanism is also introduced in the multi-scale convolutional neural network. The attention mechanism can dynamically adjust the weights of the convolution kernels so that the network can focus on the most important features for a specific task. In specific implementation, the attention mechanism calculates the attention weight of each convolution kernel and emphasizes the features with higher weights, thereby improving the accuracy and robustness of feature extraction.

[0115] In the last layer of the multi-scale convolutional neural network, the extracted multi-scale features will be summarized to form a comprehensive feature vector. This feature vector contains the multi-scale information extracted by the convolution kernels with different receptive fields, and can fully reflect the various characteristics of process parameters and production environment factors. Through this multi-scale feature extraction method, the various performances of integrated circuits under different process conditions and production environments can be effectively captured, providing rich feature data for subsequent prediction and evaluation.

[0116] Furthermore, the uncertainty assessment module is specifically used for:

[0117] The process parameter fluctuation data output by the prediction sub-network is modeled in Bayesian fashion, and the prediction results are regarded as parameters of probability distribution;

[0118] By using a Bayesian deep learning method, the probability distribution is randomly sampled multiple times to generate multiple possible process parameter fluctuation samples, and each sampling obtains a set of predicted process parameter fluctuation data;

[0119] Perform statistical analysis on the process parameter fluctuation data obtained from each set of sampling, calculate the mean and variance of each set of data, and form a predicted confidence interval;

[0120] The process parameter fluctuation data and statistical results obtained by multiple samplings are stored as multiple groups of predicted process parameter fluctuation data, and distribution characteristic diagrams of the process parameter fluctuations are generated based on these data;

[0121] Combining historical process data with current forecast data, comparative analysis is performed on each group of predicted process parameter fluctuation data to identify possible abnormal fluctuations and mark them in the final forecast results.

[0122] First, the uncertainty assessment module performs Bayesian modeling on the process parameter fluctuation data output by the prediction subnetwork. At this stage, the prediction results are regarded as parameters of probability distribution rather than single definite values. Bayesian modeling calculates the posterior distribution by setting the likelihood function of the prior distribution and the observed data, so that the prediction results can reflect the uncertainty and variability in the data. The key to this step is to select an appropriate prior distribution and update it with the observed data to obtain the posterior distribution.

[0123] Next, the above probability distribution is randomly sampled multiple times through the Bayesian deep learning method. In specific operations, random sampling is performed from the posterior distribution to generate multiple possible process parameter fluctuation samples. Each sampling will generate a set of predicted process parameter fluctuation data, reflecting the possible process parameter changes under different conditions and assumptions. This process can generate a large amount of sample data through repeated sampling, providing a rich data basis for subsequent statistical analysis.

[0124] Then, a detailed statistical analysis is performed on the process parameter fluctuation data obtained from each set of sampling. By calculating the mean and variance of each set of data, the statistical characteristics of each parameter under different conditions can be obtained. These statistical characteristics are used to form a predicted confidence interval, which represents the range and trend of process parameter fluctuations at a certain confidence level. Specifically, the mean represents the expected value of the parameter, while the variance reflects the degree and uncertainty of parameter fluctuations.

[0125] The process parameter fluctuation data and statistical results obtained from multiple samplings are stored as multiple sets of predicted process parameter fluctuation data. These data not only retain the specific results of each sampling, but also include statistical characteristics such as the mean, variance, and confidence interval of each set of data. Based on these data, a distribution characteristic diagram of process parameter fluctuations can be generated. The distribution characteristic diagram intuitively shows the change law and probability distribution of process parameters under different conditions, allowing engineers to better understand and analyze the characteristics of process parameter fluctuations.

[0126] Finally, the historical process data and current forecast data are combined to compare and analyze the process parameter fluctuation data of each group. This step identifies possible abnormal fluctuations by comparing historical data with forecast data. If some forecast data deviates significantly from historical data, it may indicate abnormal conditions or potential problems under these conditions. These abnormal fluctuations are marked in the final forecast results so that engineers can focus on these abnormal conditions during the evaluation process and take corresponding improvement measures.

[0127] The following is the reference implementation code of the process parameter fluctuation prediction model:

[0128] import numpy as np

[0129] import pandas as pd

[0130] from sklearn.preprocessing import StandardScaler

[0131] from sklearn.impute import SimpleImputer

[0132] import tensorflow as tf

[0133] from tensorflow.keras.models import Sequential

[0134] from tensorflow.keras.layers import Conv1D, MaxPooling1D, Flatten,Dense, Input, Attention, Add

[0135] from tensorflow.keras.models import Model

[0136] import tensorflow_probability as tfp

[0137] # Data preprocessing module

[0138] def data_preprocessing(data):

[0139] """

[0140] Preprocess the current process parameters and production environment factor data.

[0141] Including data cleaning, normalization and data enhancement.

[0142] """

[0143] # Data cleaning: removing noise and outliers

[0144] imputer = SimpleImputer(strategy='mean')

[0145] data_cleaned = imputer.fit_transform(data)

[0146] # Data normalization: Convert data to a uniform scale

[0147] scaler = StandardScaler()

[0148] data_normalized = scaler.fit_transform(data_cleaned)

[0149] # Data augmentation: generating more diverse data samples

[0150] noise = np.random.normal(0, 0.01, data_normalized.shape)

[0151] data_augmented = data_normalized + noise

[0152] return data_augmented

[0153] # Multimodal deep fusion module- feature extraction subnetwork

[0154] def build_feature_extraction_network(input_shape):

[0155] """

[0156] Construct a multi-scale convolutional neural network for feature extraction.

[0157] """

[0158] inputs = Input(shape=input_shape)

[0159] x1 = Conv1D(32, kernel_size=3, activation='relu', padding='same')(inputs)

[0160] x1 = MaxPooling1D(pool_size=2)(x1)

[0161] x2 = Conv1D(64, kernel_size=5, activation='relu', padding='same')(inputs)

[0162] x2 = MaxPooling1D(pool_size=2)(x2)

[0163] x3 = Conv1D(128, kernel_size=7, activation='relu', padding='same')(inputs)

[0164] x3 = MaxPooling1D(pool_size=2)(x3)

[0165] x = tf.concat([x1, x2, x3], axis=-1)

[0166] # Attention mechanism layer

[0167] attention = Attention()([x, x])

[0168] x = Add()([x, attention])

[0169] x = Flatten()(x)

[0170] model = Model(inputs, x)

[0171] return model

[0172] # Multimodal deep fusion module- feature fusion subnetwork

[0173] def build_feature_fusion_network(input_shape):

[0174] """

[0175] A multi-head attention mechanism is used to weightedly fuse process parameter features and production environment factor features.

[0176] """

[0177] inputs = Input(shape=input_shape)

[0178] # Multi-head attention mechanism

[0179] attention_output = tf.keras.layers.MultiHeadAttention(num_heads=4, key_dim=input_shape[-1])(inputs, inputs)

[0180] attention_output = Add()([inputs, attention_output])

[0181] # Residual connection

[0182] fusion_output = Flatten()(attention_output)

[0183] model = Model(inputs, fusion_output)

[0184] return model

[0185] # Prediction subnetwork - a hybrid architecture of autoregressive model and variational autoencoder

[0186] def build_prediction_network(input_shape):

[0187] """

[0188] Build a hybrid architecture of autoregressive models and variational autoencoders for time series forecasting.

[0189] """

[0190] # Autoregressive model

[0191] inputs = Input(shape=input_shape)

[0192] x = Conv1D(64, kernel_size=3, activation='relu')(inputs)

[0193] x = Flatten()(x)

[0194] # Variational Autoencoder

[0195] latent_dim = 2

[0196] encoder = Dense(latent_dim, activation='relu')(x)

[0197] z_mean = Dense(latent_dim)(encoder)

[0198] z_log_var = Dense(latent_dim)(encoder)

[0199] # Sampling

[0200] def sampling(args):

[0201] z_mean, z_log_var = args

[0202] epsilon = tf.random.normal(shape=(tf.shape(z_mean)[0],latent_dim))

[0203] return z_mean + tf.exp(0.5 * z_log_var) * epsilon

[0204] z = tf.keras.layers.Lambda(sampling, output_shape=(latent_dim,))([z_mean, z_log_var])

[0205] # Decoder

[0206] decoder = Dense(input_shape[0], activation='relu')(z)

[0207] model = Model(inputs, decoder)

[0208] return model

[0209] # Uncertainty Assessment Module - Bayesian Deep Learning Method

[0210] def bayesian_model_prediction(model, data, num_samples=100):

[0211] """

[0212] Bayesian modeling is performed on the predicted process parameter fluctuation data, and multiple sampling is performed to generate the predicted confidence interval.

[0213] """

[0214] predictions = [model.predict(data) for _ in range(num_samples)]

[0215] predictions = np.array(predictions)

[0216] # Calculate mean and variance

[0217] mean_predictions = np.mean(predictions, axis=0)

[0218] std_predictions = np.std(predictions, axis=0)

[0219] return mean_predictions, std_predictions

[0220] # Example data

[0221] data = np.random.rand(100, 10) # 100 samples, 10 features per sample

[0222] # Data preprocessing

[0223] preprocessed_data = data_preprocessing(data)

[0224] # Feature extraction

[0225] feature_extraction_network = build_feature_extraction_network(preprocessed_data.shape[1:])

[0226] extracted_features = feature_extraction_network.predict(preprocessed_data)

[0227] # Feature Fusion

[0228] feature_fusion_network = build_feature_fusion_network(extracted_features.shape)

[0229] fused_features = feature_fusion_network.predict(extracted_features)

[0230] # predict

[0231] prediction_network = build_prediction_network(fused_features.shape)

[0232] predicted_data = prediction_network.predict(fused_features)

[0233] # Uncertainty Assessment

[0234] mean_predictions, std_predictions = bayesian_model_prediction(prediction_network, fused_features)

[0235] # Output

[0236] print("Predicted mean:", mean_predictions)

[0237] print("Prediction standard deviation:", std_predictions)

[0238] The detailed steps for training the process parameter fluctuation prediction model are provided as follows.

[0239] First, collect the current process parameters and production environment factor data from the manufacturing process database, real-time monitoring system or production log. These data include process parameters such as doping concentration, oxide layer thickness, transistor size, and production environment factors such as temperature, humidity, and pressure. Data collection can be automatically acquired through the API interface, or manually exported and imported into the data processing system. The data collection process needs to ensure the integrity and accuracy of the data to avoid loss and errors.

[0240] Next, data preprocessing is performed. The purpose of the data preprocessing module is to ensure the accuracy and consistency of the data. First, data cleaning is performed to remove noise and outliers. For example, for parameter values ​​that are obviously beyond a reasonable range, the median or mean can be used to replace them. Then data normalization is performed to scale the data to the same scale. A common method is to scale the data to between 0 and 1 or standardize it to a mean of 0 and a standard deviation of 1. Data augmentation is to expand the diversity of the data set by generating virtual data points, which can be achieved by making small random perturbations to the existing data.

[0241] After data preprocessing is completed, a multimodal deep fusion model is constructed and trained. The first is the feature extraction subnetwork, which uses a multi-scale convolutional neural network (MSCNN) for feature extraction. MSCNN includes convolutional layers with different receptive fields to capture process parameters and production environment factor characteristics at different scales. After the model is built, the preprocessed data is input into the model for training.

[0242] After feature extraction is completed, it enters the feature fusion sub-network. The feature fusion sub-network uses a multi-head attention mechanism to perform weighted fusion of process parameter features and production environment factor features. After fusion is completed, the fused features are passed to the prediction sub-network, which adopts a hybrid architecture of autoregressive model and variational autoencoder (VAE).

[0243] After the model training is completed, the model is validated. The model performance is evaluated through methods such as cross-validation to ensure that the model has good generalization ability and prediction accuracy. If the model's prediction results deviate greatly from the actual situation, it may be necessary to adjust the model parameters or retrain the model.

[0244] Finally, the trained model is deployed to the production environment. During the deployment process, the model and related preprocessing steps need to be packaged into a service interface for other systems to call. For example, you can use a framework such as Flask or Django to deploy the model as a RESTful API service.

[0245] Through the above steps, the whole process from data collection, preprocessing, model training to deployment can be completed. In practical applications, the model can automatically predict the fluctuation of process parameters and provide a scientific basis for the reliability assessment of integrated circuits.

[0246] Step S104: applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters.

[0247] First, prepare multiple sets of predicted process parameter fluctuation data generated in step S103, each set of data represents the possible process parameter changes at different time points or under different environmental conditions. These data need to be input into the simulation tool in an appropriate format, such as a standardized SPICE format file.

[0248] Next, these predicted process parameter fluctuation data are applied to the equivalent circuit model constructed in step S102. The equivalent circuit model has been defined and verified in the simulation tool, and includes the topological structure and basic component parameters of the circuit. In the SPICE simulation tool, these predicted process parameters can be applied group by group by modifying the model parameter file. For example, if the predicted data indicates that the threshold voltage of a certain transistor will change, it is necessary to adjust the threshold voltage value of the transistor accordingly in the model parameters.

[0249] Then, configure the simulation environment. The configuration of the simulation environment includes setting the power supply voltage, input signal form and frequency, so that the simulation can run under the predetermined conditions. Make sure that the simulation settings are consistent with the actual working conditions so that the simulation results can truly reflect the performance of the circuit in the actual environment.

[0250] When performing simulation, run the simulation tasks for each set of process parameter fluctuation data in sequence. In the SPICE simulation tool, you can write scripts to automatically perform these tasks. Run a simulation once for each set of data, and the tool will output the corresponding performance parameters. These performance parameters usually include voltage, current, delay, power consumption, etc., which characterize the behavior of the circuit under specific process conditions.

[0251] To ensure the accuracy and reliability of the simulation results, you need to perform a preliminary verification of the results after each simulation. You can check the rationality of the simulation output by comparing it with known standard circuit performance data or using built-in verification tools. If an abnormality is found, you need to recheck the input data and simulation settings to ensure that there are no errors or omissions.

[0252] After the simulation is completed, each set of simulation results is saved and summarized. The performance parameters corresponding to each set of predicted process parameters are stored in a structured database for subsequent analysis. In this way, the simulation results corresponding to multiple process parameter fluctuation data can be systematically sorted and analyzed.

[0253] Step S105: Analyze the multiple groups of performance parameters to extract key reliability features.

[0254] First, the multiple sets of performance parameter data obtained in step S104 are collected and sorted. These performance parameter data usually include multiple indicators such as voltage, current, delay, power consumption, etc., which characterize the performance of the circuit under different process parameter fluctuations. In order to facilitate analysis, these data need to be stored in a structured manner, for example, they can be managed using a spreadsheet or database system.

[0255] Next, a preliminary statistical analysis is performed on each set of performance parameters. Basic statistics such as mean, standard deviation, maximum, minimum, and median are calculated for each set of data. These statistics can provide a preliminary understanding of the data distribution. Through these basic statistical analyses, it is possible to preliminarily identify which performance parameters fluctuate greatly under different process conditions and which are relatively stable.

[0256] Then, these performance parameters are analyzed in depth using data analysis and machine learning techniques. First, correlation analysis is performed to find out the relationship between the various performance parameters. For example, by calculating the Pearson correlation coefficient, it is determined whether there is a significant correlation between voltage and power consumption. Next, dimensionality reduction methods such as principal component analysis (PCA) are used to convert the high-dimensional performance parameter data into a low-dimensional feature space for further processing. PCA uses linear transformation to find the principal components with the largest variance in the data, which can be used as preliminary candidates for key features.

[0257] In order to extract key reliability features, cluster analysis methods can be applied. For example, using the K-means clustering algorithm, the performance parameter data is divided into several groups so that the data within the same group are as similar as possible, while the data between different groups are as different as possible. Through cluster analysis, it is possible to identify which combinations of performance parameters show consistent change trends under different process conditions, thereby extracting the feature combinations that have the greatest impact on reliability evaluation.

[0258] In addition, combining fault data for feature selection is also an effective method. By analyzing historical fault data, we can find out which performance parameters show abnormal fluctuations before a fault occurs, and use these parameters as key features to focus on. For example, we can use machine learning algorithms such as random forests or support vector machines (SVMs) to train a classification model to identify fault precursor features.

[0259] After completing the above analysis, the extracted key reliability features need to be verified. These features are applied again for simulation verification to check whether they can accurately reflect the reliability of the circuit under different process conditions. If it is found that the extracted features do not perform well under certain conditions, it may be necessary to re-select and analyze the features until the most reliable features are found.

[0260] Finally, the extracted key reliability features are sorted and recorded to form a reliability feature database, which is not only used for subsequent reliability score calculations, but also provides a reference for future design optimization and process improvement.

[0261] Furthermore, the analyzing the multiple groups of performance parameters to extract key reliability features includes:

[0262] Each set of performance parameters is represented as a vector ,in , indicating the The simulation results of the group performance parameters;

[0263] According to the following formula 1 and formula 2, the mean vector of the corresponding vector of each group of performance parameters is calculated: and the covariance matrix :

[0264]

[0265]

[0266] in, is the number of groups for circuit simulation:

[0267] According to the following formula 3, the covariance matrix is ​​analyzed using principal component analysis Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector :

[0268]

[0269] From the characteristic value and the corresponding eigenvector Select the front The eigenvectors corresponding to the largest eigenvalues ​​form the characteristic matrix ;

[0270] According to the following formula 4, the initial key feature vector of each group of performance parameters is calculated: :

[0271]

[0272] According to the following formula 5, the initial key feature vector is calculated The corresponding regularized eigenvector ;

[0273]

[0274] According to the following formula 6 and formula 7, the regularized mean vector is calculated and the regularized standard deviation vector ;

[0275]

[0276]

[0277] in, is the number of groups for circuit simulation;

[0278] The regularized mean vector and the regularized standard deviation vector , identified as a key reliability feature.

[0279] First, each set of performance parameters is represented as a vector .vector Contains the The simulation results of the group performance parameters, namely This step systematically processes multiple performance parameter data so that each set of performance parameters can be uniformly represented in vector form, which is convenient for subsequent calculation and analysis.

[0280] Next, use Formula 1 to calculate the mean vector of each group of performance parameter vectors The mean vector is the average value of the performance parameters of all simulation results, which is obtained by summing the performance parameter vectors of all simulation groups and dividing by the number of simulation groups. Specifically, in Formula 1 represents the number of groups for circuit simulation, and Represents the performance parameter vector of each set of simulation results. By calculating the mean vector , the average performance of the performance parameters under all simulation conditions can be obtained.

[0281] Then, the covariance matrix is ​​calculated using Formula 2 . Covariance matrix It reflects the linear relationship and fluctuation degree between each performance parameter. , get the deviation vector, then perform outer product operation on these deviation vectors and accumulate them, and finally divide them by Get the covariance matrix. Specifically, in Formula 2 It is The performance parameter vector of the group simulation results, is the mean vector, and Represents the deviation vector. The calculation of the covariance matrix can help understand the variation of performance parameters under different simulation conditions.

[0282] Then, the covariance matrix is ​​obtained by the formula Perform principal component analysis (PCA) to obtain eigenvalues and the corresponding eigenvector The purpose of formula 3 is to perform eigenvalue decomposition on the covariance matrix and extract the main direction of change in the data. Indicates the data variance in the direction of the corresponding eigenvector, and the eigenvector These represent the main directions of change.

[0283] Next, from the eigenvalue and the corresponding eigenvector Select the front The eigenvectors corresponding to the largest eigenvalues ​​form the characteristic matrix . Feature Matrix Contains the previous The most important direction of change can effectively reduce the dimension and retain the most important information in the data. In this way, the data can be simplified while retaining its main features.

[0284] Next, use Formula 4 to calculate the initial key feature vector for each set of performance parameters: Specifically, by transforming the performance parameter vector Subtract the mean vector , get the centralized performance parameter vector, and then use the feature matrix Perform linear transformation to obtain the initial key feature vector The function of formula 4 is to transform the original high-dimensional data into a low-dimensional space for further analysis.

[0285] Then, use Formula 5 to calculate the initial relevant feature vector The corresponding regularized eigenvector By using the initial key feature vector Adding a term proportional to the square of the deviation of the performance parameter vector to can enhance the robustness of the feature vector. is a regularization coefficient used to control the influence of the regularization term, and Represents the squared deviation of the performance parameter vector.

[0286] Next, use Equation 6 to calculate the regularized mean vector By normalizing all feature vectors Sum and divide by the number of simulation groups , we can get the regularized mean vector The function of Formula 6 is to obtain the average value of the regularized feature vectors of all simulation results, reflecting the overall feature performance.

[0287] Finally, the regularized standard deviation vector is calculated using Formula 7 By calculating each regularized eigenvector With the mean vector The square of the difference between , averaged and squared, can be used to obtain the regularized standard deviation vector The function of Formula 7 is to reflect the fluctuation degree of the regularized eigenvector of all simulation results.

[0288] Through the above steps, key reliability features can be extracted, which will be used to evaluate the reliability of integrated circuits and provide a basis for further analysis and optimization. The parameters and vectors in each formula are calculated using simulation data and statistical methods to ensure the accuracy and reliability of the results.

[0289] Step S106: Calculate the reliability score of the integrated circuit based on the extracted key reliability features.

[0290] First, the key reliability feature data extracted in step S105 needs to be sorted and standardized. These key reliability features may include indicators such as voltage stability, power consumption change, and signal delay. Before performing the scoring calculation, it is necessary to ensure that all feature data are on the same scale or dimension to avoid distortion of the calculation results due to differences in data ranges. Standardization can be performed by subtracting the mean and dividing by the standard deviation to convert each feature value into a standard normal distribution.

[0291] Next, select a suitable scoring model to perform a weighted combination of key reliability features to calculate the overall reliability score of the integrated circuit. The choice of scoring model can be determined based on the specific application scenario and reliability requirements. Common scoring models include linear weighted models, logistic regression models, and machine learning-based models. For linear weighted models, each key reliability feature is assigned a weight that reflects the degree of influence of each feature on the overall reliability. The weights can be determined through expert experience or data-driven methods, such as using regression analysis to determine the weight coefficients.

[0292] Substitute the standardized key reliability feature data into the scoring model, calculate the weighted score of each feature, and add up these scores to get the final reliability score. Assuming a linear weighted model is used, the formula can be expressed as:

[0293]

[0294] in, is the final reliability score, It is The weights of the key reliability features, It is The standardized key reliability characteristic value, is the total number of key reliability characteristics. This formula can be implemented in a programming environment to ensure the automation and accuracy of the calculation.

[0295] After the calculation is completed, the reliability score needs to be checked and verified. By comparing the calculated score with the known reliability test data, check whether the score result is reasonable and accurate. If it is found that there is a large deviation between the score result and the actual situation, it may be necessary to readjust the score model or weight coefficient until the score result can accurately reflect the actual reliability of the integrated circuit.

[0296] Finally, the reliability scoring results are collated and a reliability assessment report is generated. The report should include the specific numerical values ​​of the scores, a detailed analysis of each key reliability feature, and possible improvement suggestions. This information will provide an important reference for design and production, helping engineers optimize design and process parameters to improve the overall reliability of integrated circuits.

[0297] Furthermore, the step of calculating the reliability score of the integrated circuit based on the extracted key reliability features includes:

[0298] According to the following formula 8, the reliability score of the integrated circuit is calculated :

[0299]

[0300] in, is the regularized mean vector No. Quantity; is the regularized standard deviation vector No. Quantity; It is Weight factors; is a constant; and is the adjustment parameter; is the regularized mean vector Middle The target mean of the components; is the selected front in principal component analysis The number of key features.

[0301] In formula 8, we first need to understand the regularized mean vector No. Quantity This component represents the mean of the regularized eigenvectors of all simulation results, which is obtained by summing all the regularized eigenvectors and dividing by the number of simulation groups. It reflects the average performance of each key feature.

[0302] Next is the regularized standard deviation vector No. Quantity This component represents the degree of fluctuation of the regularized feature vector under different simulation conditions. It is obtained by calculating the square of the difference between each regularized feature vector and the mean vector, averaging and taking the square root. It reflects the volatility and reliability of the feature.

[0303] In formula 8, It is The weight factor of each feature. This factor is used to measure the importance of each feature in the overall reliability score and can be determined based on empirical or data-driven methods. For example, some features may have a greater impact on reliability and therefore be given a higher weight.

[0304] constant Is a small positive number used to prevent division by zero errors. It ensures that the denominator in the formula does not become zero during calculations, thus ensuring the stability and reliability of the calculations.

[0305] parameter and are adjustment parameters used to adjust the influence of the second part of the formula. These parameters can be determined by experiments or optimization methods to ensure the accuracy and effectiveness of the reliability score.

[0306] is the regularized mean vector Middle The target mean for each component. This value is usually determined based on historical data or design specifications and represents the expected value of each characteristic under ideal conditions.

[0307] is the selected front in principal component analysis This number determines the number of features involved in the reliability score calculation. The most important features ensure that the score is concentrated on the most representative features. Contribution to the overall score.

[0308] Item 2 A Sigmoid function is used to smoothly reflect the difference between the feature value and the target mean, through the parameter and Adjust its impact.

[0309] By combining these components, Equation 8 can provide a comprehensive reliability score , reflecting the overall reliability performance of the integrated circuit under various simulation conditions.

[0310] Furthermore, the method for evaluating an integrated circuit further includes:

[0311] Calculated reliability score With a pre-set reliability threshold Make comparisons;

[0312] if , then it is judged that the integrated circuit is reliable;

[0313] if , then it is judged that the integrated circuit is unreliable.

[0314] In order to further judge the reliability of the integrated circuit, it is also necessary to calculate the reliability score With a pre-set reliability threshold Make a comparison.

[0315] Specifically, we first need to calculate the reliability score of the integrated circuit This score is calculated based on the previous formula 8 by analyzing and processing multiple groups of performance parameters, extracting key reliability characteristics, and then combining these characteristics. Represents the overall reliability performance of the integrated circuit under various simulation conditions.

[0316] Next, this reliability score With a pre-set reliability threshold This threshold becomes an important criterion for circuit reliability. The setting of the threshold needs to take into account a variety of factors to ensure that it can effectively distinguish reliable and unreliable integrated circuits.

[0317] During the comparison process, if the calculated reliability score Greater than or equal to the preset value , it can be judged that the integrated circuit is reliable. This means that the performance of the integrated circuit under various simulation conditions has reached the expected standard and can meet the reliability requirements of design and application.

[0318] On the contrary, if the calculated reliability score Less than the preset value , it can be judged that the integrated circuit is unreliable. This indicates that the performance of the integrated circuit under certain simulation conditions does not meet the expected standards, there may be potential reliability problems, and further optimization and improvement are needed.

[0319] Through this comparison method, the reliability of the integrated circuit can be quickly and accurately evaluated to determine whether it can meet the design and application requirements. In specific implementation, first calculate the reliability score according to the previous step. , and then with the set threshold Compare and judge the reliability of the integrated circuit according to the comparison results. This method is fast and effective, and can provide an important basis for the design and optimization of integrated circuits, ensuring their reliability and stability in practical applications.

[0320] In the above-mentioned embodiment, a method for evaluating an integrated circuit is provided. Correspondingly, the present application also provides an apparatus for evaluating an integrated circuit. Figure 2 , which is a schematic diagram of an embodiment of an apparatus for evaluating integrated circuits of the present application. Since this embodiment, namely the second embodiment, is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The apparatus embodiment described below is merely illustrative.

[0321] A second embodiment of the present application provides a device for evaluating an integrated circuit, comprising:

[0322] An acquisition unit 201 is used to acquire a circuit topology and process parameters of an integrated circuit to be evaluated;

[0323] A construction unit 202, configured to construct an equivalent circuit model of the integrated circuit based on the circuit topology structure;

[0324] A generating unit 203 is used to generate multiple sets of predicted process parameter fluctuation data based on current process parameters and production environment factor data using a pre-trained process parameter fluctuation prediction model;

[0325] A simulation unit 204, configured to apply each set of predicted process parameter fluctuation data to the equivalent circuit model, perform circuit simulation, and obtain multiple sets of performance parameters;

[0326] An extraction unit 205, configured to analyze the plurality of groups of performance parameters and extract key reliability features;

[0327] The calculation unit 206 is used to calculate the reliability score of the integrated circuit based on the extracted key reliability features.

[0328] The third embodiment of the present application provides a system embodiment for evaluating integrated circuits. Since this embodiment, namely the third embodiment, is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The system embodiment described below is only illustrative.

[0329] A third embodiment of the present application provides a system for evaluating an integrated circuit, including:

[0330] A server including a central processing unit, memory and storage device for running the integrated circuit reliability assessment software;

[0331] A data acquisition module, used to obtain the circuit topology structure, process parameters and production environment factor data of the integrated circuit to be evaluated, and transmit the data to the server;

[0332] A remote access module configured to allow a user to remotely access and operate the server through a network interface to input information of the integrated circuit to be evaluated and view evaluation results;

[0333] A data processing module is configured in the server, and the data processing module includes:

[0334] The data preprocessing submodule is used to clean, normalize and enhance the received process parameter and production environment factor data;

[0335] A model building submodule, used to build an equivalent circuit model of the integrated circuit based on the circuit topology;

[0336] The prediction submodule is used to generate multiple sets of predicted process parameter fluctuation data based on the current process parameters and production environment factor data using a pre-trained process parameter fluctuation prediction model;

[0337] A simulation submodule, used for applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters;

[0338] A feature analysis submodule, used to analyze the multiple groups of performance parameters and extract key reliability features;

[0339] A score calculation submodule, used to calculate the reliability score of the integrated circuit based on the extracted key reliability features;

[0340] The result storage and display module is configured in the server and is used to store and display the reliability evaluation results, including:

[0341] The data storage submodule is used to store all input data, process intermediate data, and final evaluation results;

[0342] The result display submodule is used to generate and display visual results including reliability scoring and key feature analysis;

[0343] The alarm and notification module is configured to automatically send alarm notifications to designated users when the reliability score falls below a preset threshold.

[0344] The core of the system is a server, which consists of a central processor, memory and storage devices, running integrated circuit reliability assessment software. The server is responsible for processing the data transmitted from each module and executing complex assessment algorithms.

[0345] First, the data acquisition module is used to collect detailed information about the IC to be evaluated, including circuit topology, process parameters, and production environment factor data. Circuit topology information is usually extracted from design files, process parameters include physical properties such as doping concentration and oxide thickness, and production environment factor data covers external conditions such as temperature and humidity. The module acquires data through sensors or interface devices and transmits it to the server for further processing.

[0346] The remote access module on the server allows users to remotely access and operate the server through the network interface. Users can input the information of the integrated circuit to be evaluated through the remote access module and view the evaluation results at any time. This function greatly enhances the flexibility and user experience of the system, allowing users to operate without having to be on site.

[0347] The data processing module is configured in the server and is responsible for data preprocessing, model building, prediction, simulation, feature analysis and score calculation. The data processing module includes multiple submodules, each with its own responsibilities. The data preprocessing submodule first cleans, normalizes and enhances the received process parameters and production environment factor data. The cleaning step is used to remove noise and outliers in the data, the normalization step converts the data to a uniform scale, and the data enhancement enriches the data set by generating virtual data points to improve the robustness of the model.

[0348] Next, the model building submodule builds an equivalent circuit model of the integrated circuit based on the circuit topology. The equivalent circuit model is a simplified mathematical model used to simulate the behavior of the actual circuit, which is convenient for subsequent simulation and analysis. The prediction submodule uses the pre-trained process parameter fluctuation prediction model to generate multiple sets of predicted process parameter fluctuation data based on the current process parameters and production environment factor data. These predicted data reflect the possible changes in process parameters under different conditions.

[0349] The simulation submodule applies each set of predicted process parameter fluctuation data to the equivalent circuit model, performs circuit simulation, and obtains multiple sets of performance parameters. The simulation process simulates the performance of the circuit under various process conditions and outputs performance indicators such as voltage, current, and power consumption. The feature analysis submodule conducts in-depth analysis of multiple sets of performance parameters and extracts key reliability features. The extracted features include circuit performance stability, failure rate, etc. These features are important bases for evaluating circuit reliability.

[0350] The scoring calculation submodule calculates the reliability score of the integrated circuit based on the extracted key reliability features. The scoring algorithm combines various key features and outputs a reliability score to quantitatively evaluate the reliability level of the circuit.

[0351] The result storage and display module is configured in the server to store and display the evaluation results. The data storage submodule saves all input data, processed intermediate data, and the final evaluation results. The result display submodule generates visualization results including reliability scoring and key feature analysis, which is convenient for users to intuitively understand the evaluation situation.

[0352] Finally, the alarm and notification module automatically sends an alarm notification to the designated user when the reliability score is below the preset threshold. This function ensures that users can detect potential problems in a timely manner and take corresponding measures to further improve the reliability and safety of integrated circuits. The entire system achieves a comprehensive, accurate and efficient assessment of the reliability of integrated circuits through the collaborative work of various modules.

[0353] The fourth embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps provided in the first embodiment of the present application are implemented.

[0354] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for evaluating an integrated circuit, characterized in that: include: Obtaining circuit topology and process parameters of the integrated circuit to be evaluated; Based on the circuit topology, construct an equivalent circuit model of the integrated circuit; Based on the current process parameters and production environment factor data, a pre-trained process parameter fluctuation prediction model is used to generate multiple sets of predicted process parameter fluctuation data; Applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters; Analyzing the multiple groups of performance parameters to extract key reliability features; Calculating a reliability score of the integrated circuit based on the extracted key reliability features; Wherein, the process parameter fluctuation prediction model using pre-training includes a data preprocessing module, a multimodal deep fusion module and an uncertainty assessment module; The data preprocessing module is used to preprocess the current process parameters and production environment factor data to obtain preprocessed process parameter data and production environment factor data, and the preprocessing includes cleaning, normalization and data enhancement; The multimodal deep fusion module includes a feature extraction subnetwork, a feature fusion subnetwork and a prediction subnetwork; the feature extraction subnetwork extracts features from the preprocessed process parameter data and production environment factor data through multiple convolutional layers to obtain process parameter features and production environment factor features, and in the process of feature extraction, the effectiveness of feature extraction is improved through the attention mechanism layer; the feature fusion subnetwork uses a multi-head attention mechanism to perform weighted fusion on the process parameter features and production environment factor features to generate fusion features, and in the process of weighted fusion, the integrity and effectiveness of the features are maintained through residual connections; the prediction subnetwork adopts a hybrid architecture of an autoregressive model and a variational autoencoder to perform time series prediction and generation on the fusion features to obtain predicted process parameter fluctuation data; The uncertainty assessment module adopts the Bayesian deep learning method to perform multiple sampling on the predicted process parameter fluctuation data to generate the predicted confidence interval, and obtain multiple groups of predicted process parameter fluctuation data; The step of analyzing the multiple groups of performance parameters to extract key reliability features includes: Each set of performance parameters is represented as a vector ,in , indicating the The simulation results of the group performance parameters; According to the following formula 1 and formula 2, the mean vector of the corresponding vector of each group of performance parameters is calculated: and the covariance matrix : in, is the number of groups for circuit simulation: According to the following formula 3, the covariance matrix is ​​analyzed using principal component analysis Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector : From the characteristic value and the corresponding eigenvector Select the front The eigenvectors corresponding to the largest eigenvalues ​​form the characteristic matrix ; According to the following formula 4, the initial key feature vector of each group of performance parameters is calculated: : According to the following formula 5, the initial key feature vector is calculated The corresponding regularized eigenvector ; in, is a regularization coefficient; According to the following formula 6 and formula 7, the regularized mean vector is calculated and the regularized standard deviation vector ; in, is the number of groups for circuit simulation; The regularized mean vector and the regularized standard deviation vector , identified as a key reliability feature.

2. The method for evaluating an integrated circuit according to claim 1, characterized in that: The feature extraction subnetwork performs feature extraction through a multi-scale convolutional neural network, which includes convolutional layers with different receptive fields to capture process parameters and production environment factor characteristics at different scales, thereby improving the accuracy and robustness of feature extraction.

3. The method for evaluating an integrated circuit according to claim 1, characterized in that: The uncertainty assessment module is specifically used for: The process parameter fluctuation data output by the prediction sub-network is modeled in Bayesian fashion, and the prediction results are regarded as parameters of probability distribution; By using a Bayesian deep learning method, the probability distribution is randomly sampled multiple times to generate multiple possible process parameter fluctuation samples, and each sampling obtains a set of predicted process parameter fluctuation data; Perform statistical analysis on the process parameter fluctuation data obtained from each set of sampling, calculate the mean and variance of each set of data, and form a predicted confidence interval; The process parameter fluctuation data and statistical results obtained by multiple samplings are stored as multiple groups of predicted process parameter fluctuation data, and distribution characteristic diagrams of the process parameter fluctuations are generated based on these data; Combining historical process data with current forecast data, comparative analysis is performed on each group of predicted process parameter fluctuation data to identify possible abnormal fluctuations and mark them in the final forecast results.

4. The method for evaluating an integrated circuit according to claim 3, characterized in that: The step of calculating the reliability score of the integrated circuit based on the extracted key reliability features comprises: According to the following formula 8, the reliability score of the integrated circuit is calculated : in, is the regularized mean vector No. Quantity; is the regularized standard deviation vector No. Quantity; It is Weighting factors; is a constant; and is the adjustment parameter; is the regularized mean vector Middle The target mean of the components; is the selected predecessor in principal component analysis The number of key features.

5. The method for evaluating an integrated circuit according to claim 4, characterized in that: Also includes: Calculated reliability score With a pre-set reliability threshold Make comparisons; if , then it is judged that the integrated circuit is reliable; if , then it is judged that the integrated circuit is unreliable.

6. A device for evaluating an integrated circuit, characterized in that: include: An acquisition unit, used for acquiring a circuit topology and process parameters of the integrated circuit to be evaluated; A construction unit, used to construct an equivalent circuit model of the integrated circuit based on the circuit topology structure; A generating unit, for generating multiple sets of predicted process parameter fluctuation data using a pre-trained process parameter fluctuation prediction model based on current process parameters and production environment factor data; A simulation unit, used for applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters; An extraction unit, used to analyze the multiple groups of performance parameters and extract key reliability features; A calculation unit, configured to calculate a reliability score of the integrated circuit based on the extracted key reliability features; Wherein, the process parameter fluctuation prediction model using pre-training includes a data preprocessing module, a multimodal deep fusion module and an uncertainty assessment module; The data preprocessing module is used to preprocess the current process parameters and production environment factor data to obtain preprocessed process parameter data and production environment factor data, and the preprocessing includes cleaning, normalization and data enhancement; The multimodal deep fusion module includes a feature extraction subnetwork, a feature fusion subnetwork and a prediction subnetwork; the feature extraction subnetwork extracts features from the preprocessed process parameter data and production environment factor data through multiple convolutional layers to obtain process parameter features and production environment factor features, and in the process of feature extraction, the effectiveness of feature extraction is improved through the attention mechanism layer; the feature fusion subnetwork uses a multi-head attention mechanism to perform weighted fusion on the process parameter features and production environment factor features to generate fusion features, and in the process of weighted fusion, the integrity and effectiveness of the features are maintained through residual connections; the prediction subnetwork adopts a hybrid architecture of an autoregressive model and a variational autoencoder to perform time series prediction and generation on the fusion features to obtain predicted process parameter fluctuation data; The uncertainty assessment module adopts the Bayesian deep learning method to perform multiple sampling on the predicted process parameter fluctuation data to generate the predicted confidence interval, and obtain multiple groups of predicted process parameter fluctuation data; The step of analyzing the multiple groups of performance parameters to extract key reliability features includes: Each set of performance parameters is represented as a vector ,in , indicating the The simulation results of the group performance parameters; According to the following formula 1 and formula 2, the mean vector of the corresponding vector of each group of performance parameters is calculated: and the covariance matrix : in, is the number of groups for circuit simulation: According to the following formula 3, the covariance matrix is ​​analyzed using principal component analysis Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector : From the characteristic value and the corresponding eigenvector Select the front The eigenvector corresponding to the largest eigenvalue forms the characteristic matrix ; According to the following formula 4, the initial key feature vector of each group of performance parameters is calculated: : According to the following formula 5, the initial key feature vector is calculated The corresponding regularized eigenvector ; in, is a regularization coefficient; According to the following formula 6 and formula 7, the regularized mean vector is calculated and the regularized standard deviation vector ; in, is the number of groups for circuit simulation; The regularized mean vector and the regularized standard deviation vector , identified as a key reliability feature.

7. A system for evaluating integrated circuits, characterized in that: include: A server including a central processing unit, memory and storage device for running the integrated circuit reliability assessment software; A data acquisition module, used to obtain the circuit topology structure, process parameters and production environment factor data of the integrated circuit to be evaluated, and transmit the data to the server; A remote access module configured to allow a user to remotely access and operate the server through a network interface to input information of the integrated circuit to be evaluated and view evaluation results; A data processing module is configured in the server, and the data processing module includes: The data preprocessing submodule is used to clean, normalize and enhance the received process parameter and production environment factor data; A model building submodule, used to build an equivalent circuit model of the integrated circuit based on the circuit topology; The prediction submodule is used to generate multiple sets of predicted process parameter fluctuation data based on the current process parameters and production environment factor data using a pre-trained process parameter fluctuation prediction model; A simulation submodule, used for applying each set of predicted process parameter fluctuation data to the equivalent circuit model, performing circuit simulation, and obtaining multiple sets of performance parameters; A feature analysis submodule, used to analyze the multiple groups of performance parameters and extract key reliability features; A score calculation submodule, used to calculate the reliability score of the integrated circuit based on the extracted key reliability features; The result storage and display module is configured in the server and is used to store and display the reliability evaluation results, including: The data storage submodule is used to store all input data, process intermediate data, and final evaluation results; The result display submodule is used to generate and display visual results including reliability scoring and key feature analysis; an alarm and notification module configured to automatically send an alarm notification to a designated user when the reliability score falls below a preset threshold; Wherein, the process parameter fluctuation prediction model using pre-training includes a data preprocessing module, a multimodal deep fusion module and an uncertainty assessment module; The data preprocessing module is used to preprocess the current process parameters and production environment factor data to obtain preprocessed process parameter data and production environment factor data, and the preprocessing includes cleaning, normalization and data enhancement; The multimodal deep fusion module includes a feature extraction subnetwork, a feature fusion subnetwork and a prediction subnetwork; the feature extraction subnetwork extracts features from the preprocessed process parameter data and production environment factor data through multiple convolutional layers to obtain process parameter features and production environment factor features, and in the process of feature extraction, the effectiveness of feature extraction is improved through the attention mechanism layer; the feature fusion subnetwork uses a multi-head attention mechanism to perform weighted fusion on the process parameter features and production environment factor features to generate fusion features, and in the process of weighted fusion, the integrity and effectiveness of the features are maintained through residual connections; the prediction subnetwork adopts a hybrid architecture of an autoregressive model and a variational autoencoder to perform time series prediction and generation on the fusion features to obtain predicted process parameter fluctuation data; The uncertainty assessment module adopts the Bayesian deep learning method to perform multiple sampling on the predicted process parameter fluctuation data to generate the predicted confidence interval, and obtain multiple groups of predicted process parameter fluctuation data; The step of analyzing the multiple groups of performance parameters to extract key reliability features includes: Each set of performance parameters is represented as a vector ,in , indicating the The simulation results of the group performance parameters; According to the following formula 1 and formula 2, the mean vector of the corresponding vector of each group of performance parameters is calculated: and the covariance matrix : in, is the number of groups for circuit simulation: According to the following formula 3, the covariance matrix is ​​analyzed using principal component analysis Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector : From the characteristic value and the corresponding eigenvector Select the front The eigenvector corresponding to the largest eigenvalue forms the characteristic matrix ; According to the following formula 4, the initial key feature vector of each group of performance parameters is calculated: : According to the following formula 5, the initial key feature vector is calculated The corresponding regularized eigenvector ; in, is a regularization coefficient; According to the following formula 6 and formula 7, the regularized mean vector is calculated and the regularized standard deviation vector ; in, is the number of groups for circuit simulation; The regularized mean vector and the regularized standard deviation vector , identified as a key reliability feature.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method described in any one of claims 1 to 5 is implemented.

Images