Integrated circuit fault testability prediction method based on graph neural network

Through the integrated circuit fault measurability prediction method based on graph neural network, the CNF fault data set is encoded into an undirected graph using the NeuroSAT model, which solves the problems of low computing efficiency and waste of resources in the traditional method, and achieves efficient fault measurability Sex prediction and test vector generation reduce testing costs.

CN118981992APending Publication Date: 2024-11-19NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411118022.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

The traditional fault measurability analysis method has low computational efficiency and huge resource consumption in integrated circuit testing, so it cannot effectively predict unpredictable faults, resulting in wasting computing resources and time on unpredictable faults.

Method used

Using the integrated circuit fault measurability prediction method based on graph neural network, the CNF fault data set is encoded into an undirected graph by constructing the NeuroSAT graph neural network model, and the fault measurability prediction is performed using multi-layer perceptron and message delivery mechanisms. Decide whether to perform automatic test vector generation or Boolean satisfactory analysis.

Benefits of technology

Effectively capture the complex relationship between circuit topology and fault characteristics, avoid invalid expenses on unmeasured faults, improve test efficiency and resource utilization, and reduce test costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981992A_ABST
    Figure CN118981992A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of integrated circuit testability design, and discloses an integrated circuit fault testability prediction method based on a graph neural network, which comprises the following steps: reading a network table file, and converting a circuit fault into a CNF formula; constructing a special CNF fault data set, and dividing the special CNF fault data set into a training set, a verification set and a test set according to a preset proportion; a NeuroSAT graph neural network model is constructed and trained; carrying out fault enumeration on the new circuit netlist file and carrying out testability prediction by using the trained model; and determining whether to perform automatic test vector generation or Boolean satisfiability analysis based on the prediction result. According to the invention, untestable faults in the circuit can be identified in advance, useless test vector generation on the untestable faults is avoided, and the test cost is reduced. Experimental analysis proves that the method disclosed by the invention is greatly reduced in solving time and backtracking rate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of integrated circuit testability design, and in particular relates to an integrated circuit fault testability prediction method based on a graph neural network. Background Art

[0002] ATPG (Automatic Test Pattern Generation) is the process of automatically generating test vectors used in integrated circuit testing. An important step is to determine whether a specific fault is testable, that is, whether there is a test vector that can detect the fault. However, with the continuous increase in the scale and complexity of integrated circuits, traditional fault testability analysis methods face challenges such as low computational efficiency and huge resource consumption.

[0003] In recent years, the application of Boolean Satisfiability Problem (SAT) solving technology in the field of ATPG has made significant progress. By converting the circuit fault detection problem into a Boolean Satisfiability Problem, the SAT solving technology uses an efficient CNF (Conjunctive Normal Form) solver to obtain the correct variable input, thereby obtaining the correct test vector. However, for untestable faults, the existing technology cannot determine the testability of the fault before ATPG, resulting in a large waste of computing time and resources. Therefore, how to accurately and efficiently predict the testability of the fault, generate test vectors in a targeted manner, avoid wasting computing resources and time on untestable faults, and reduce the test cost while ensuring accuracy is an urgent problem to be solved. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides an integrated circuit fault testability prediction method based on graph neural network, which can solve the problem of resource waste on untestable faults in the traditional automatic test vector generation process.

[0005] The present invention provides a method for predicting the fault testability of integrated circuits based on a graph neural network, comprising the following steps:

[0006] Step 1: read the netlist file, create a CNF fault data set, and randomly divide the CNF fault data set into a training set, a validation set, and a test set according to a preset ratio;

[0007] Step 2: construct a NeuroSAT graph neural network model, input the training set and the validation set into the NeuroSAT network model for training and validation, obtain a training weight file, and adjust the parameters based on the loss function to optimize the NeuroSAT network model;

[0008] Step 3: Read in the new netlist file, enumerate the faults and use the optimized NeuroSAT network model to infer and obtain the fault testability;

[0009] Step 4: According to the fault testability, decide whether to perform automatic test vector generation or Boolean satisfiability analysis, and further optimize the fault testability prediction model according to the result.

[0010] Further, in step 1, when making the CNF fault data set, each logic gate in the circuit is converted into an equivalent CNF clause, and a sensitization constraint clause and a fault activation clause are added to each gate in the fault fan-out cone;

[0011] The CNF fault data set is divided into a training set, a validation set and a test set.

[0012] Furthermore, the NeuroSAT graph neural network model includes an input processing layer, two initialization vector layers, two multi-layer perceptrons, two normalization layers and an output layer;

[0013] The input processing layer encodes the CNF formula into an undirected graph including literal nodes and clause nodes;

[0014] The initialization vector layer is used to initialize the embedding vectors of all text nodes and clause nodes in the undirected graph;

[0015] The multilayer perceptron is used to process the messages of the initialized text nodes and perform T rounds of iterations;

[0016] The normalization layer receives messages from adjacent text nodes after T rounds of iterations and updates the embedding;

[0017] The output layer calculates the average of all updated embedding word votes as the final prediction output.

[0018] Furthermore, the initialization vector layer uses a predefined vector L init The initialization text node expression is as follows, where L (t) represents the text embedding matrix at the tth iteration, 2n represents the number of words, d represents the embedding dimension,

[0019] Initialize the predefined vector C used by the vector layer init The initialization clause node expression is as follows, where C (t) represents the clause embedding matrix at the tth iteration, m represents the number of clauses, d represents the embedding dimension,

[0020]

[0021] The update clause node embedding expression is as follows, where C_u represents the network of the normalized layer update clause embedding, C_h (t) : The hidden state of the network, M T represents the transpose of the adjacency matrix, L_msg represents the multi-layer perceptron layer processing the transmission of text messages;

[0022] The expression for updating the text node embedding is as follows, where the L_u normalization layer represents the network for updating the text embedding, and L_h (t) represents the hidden state of the network, Flip represents the function of exchanging complementary words, M represents the adjacency matrix, and C_msg represents the transmission of clause messages processed by the multi-layer perceptron layer;

[0023]

[0024] The loss function used in the output layer is the sigmoid cross entropy loss function, which is expressed as follows, where L represents the loss function, i.e., cross entropy, y represents the true label, and σ(x) represents the sigmoid activation output by the model:

[0025] L=-[yln(σ(x))+(1-y)ln(1-σ(x))].

[0026] Furthermore, step 3 is specifically as follows: when using the optimized NeuroSAT network model for reasoning, each fault point in the new netlist file is converted into a CNF formula and encoded into an undirected graph containing text nodes and clause nodes; the encoded undirected graph is input into the optimized NeuroSAT network model to obtain the corresponding fault testability prediction value.

[0027] Furthermore, step 4 is specifically as follows:

[0028] For the fault points determined to be testable faults, automatic test vector generation is performed;

[0029] For faults that are determined to be untestable, a Boolean satisfiability analysis is performed and the test vector generation process is skipped.

[0030] The beneficial effects of the present invention are as follows: the present invention introduces a dedicated CNF fault data set, and by encoding the circuit fault into a CNF formula, the data set retains the logical characteristics of the fault, so that the model can learn a deeper fault mode; the present invention introduces an improved NeuroSAT network model, and on the basis of the original NeuroSAT graph neural network model, an initialization vector layer, a multi-layer perceptron layer and a normalization layer are set, and the improved NeuroSAT network encodes the SAT problem into an undirected graph containing text nodes and clause nodes. This representation method retains the structural information of the problem and helps the network learn a deeper logical relationship; the message passing mechanism of the graph neural network is adopted, so that the network can reason on the graph structure, which is very suitable for processing combinatorial problems such as SAT; using multiple rounds of message passing iterations, the network can simulate the search process. The present invention uses the graph neural network to optimize the digital circuit back-end test of SAT ATPG, which can effectively capture the complex relationship between circuit topology and fault characteristics, effectively avoid the ineffective expenditure on untestable faults, and solve the efficiency and resource consumption problems faced by traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the integrated circuit fault testability prediction method based on graph neural network according to the present invention;

[0032] Figure 2 A schematic diagram for generating a CNF fault data set in the present invention;

[0033] Figure 3 A schematic diagram of a CNF file of a circuit fault of the present invention;

[0034] Figure 4 A schematic diagram of graph structure message transmission of an integrated circuit fault testability prediction method based on a graph neural network according to an embodiment of the present invention;

[0035] Figure 5 A diagram showing the statistical data of the results of Boolean satisfiability solution (SAT);

[0036] Figure 6 Schematic diagram of statistical data of the learning process of the method described in the present invention. DETAILED DESCRIPTION

[0037] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0038] Example 1

[0039] like Figure 1 As shown, the method for predicting the fault testability of integrated circuits based on graph neural network described in the present invention comprises the following steps:

[0040] Step 1, read the netlist file, create a CNF fault data set, and randomly divide the CNF fault data set into a training set, a validation set, and a test set according to the preset ratio of 80%, 10%, and 10%;

[0041] When making the CNF fault data set, each logic gate in the circuit is converted into an equivalent CNF clause, and a sensitization constraint clause and a fault activation clause are added to each gate in the fault fan-out cone;

[0042] Step 2: construct a NeuroSAT graph neural network model, input the training set and the validation set into the NeuroSAT network model for training and validation, obtain a training weight file, and adjust the parameters based on the loss function to optimize the NeuroSAT network model;

[0043] Step 3: Read in the new netlist file, enumerate the faults and use the optimized NeuroSAT network model to infer and obtain the fault testability;

[0044] Step 4: According to the fault testability, decide whether to perform automatic test vector generation or Boolean satisfiability analysis, and further optimize the fault testability prediction model according to the result.

[0045] Wherein, in step 2, the NeuroSAT graph neural network model includes an input processing layer, two initialization vector layers L_init and C_init, two multi-layer perceptrons L_msg and C_msg, two normalization layers L_u and C_u, and an output layer; wherein L_init and C_init are respectively the embedding vectors of the initialized text nodes and clause nodes; L_msg and C_msg respectively represent the processing of messages transmitted from text nodes to clause nodes and the processing of messages transmitted from clause nodes to text nodes; L_u and C_u represent the updating of embedding vectors of text nodes and clause nodes;

[0046] The input processing layer encodes the SAT problem into an undirected graph including text nodes and clause nodes;

[0047] The initialization vector layer is used to initialize the embedding vectors of all text nodes and clause nodes in the undirected graph;

[0048] The multilayer perceptron is used to process the messages of the initialized text nodes and perform T rounds of iterations;

[0049] The normalization layer receives messages from adjacent text nodes after T rounds of iterations and updates the embedding;

[0050] The output layer calculates the average of all updated embedding word votes as the final prediction output.

[0051] By adjusting the embedding dimension and the number of information transmissions, the cross entropy loss function is calculated to evaluate and optimize the model performance, and the model parameters are updated through the back propagation algorithm; through this process, the present invention can systematically obtain the training weight file, and optimize the embedding dimension and the number of iterations, thereby improving the performance of the model in the fault testability prediction task.

[0052] Initialize the predefined vector L used by the vector layer init The initialization text node expression is as follows, where L (t) represents the text embedding matrix at the tth iteration, 2n represents the number of words, d represents the embedding dimension,

[0053]

[0054] Initialize the predefined vector C used by the vector layer init The initialization clause node expression is as follows, where C (t) represents the clause embedding matrix at the tth iteration, m represents the number of clauses, d represents the embedding dimension,

[0055]

[0056] The update clause node embedding expression is as follows, where C_u represents the network of the normalized layer update clause embedding, C_h (t) : The hidden state of the network, M T represents the transpose of the adjacency matrix, L_msg represents the multi-layer perceptron layer processing the transmission of text messages;

[0057] The expression for updating the text node embedding is as follows, where the L_u normalization layer represents the network for updating the text embedding, and L_h (t) represents the hidden state of the network, Flip represents the function of exchanging complementary words, M represents the adjacency matrix, and C_msg represents the transmission of clause messages processed by the multi-layer perceptron layer;

[0058]

[0059] The loss function used in the output layer is the sigmoid cross entropy loss function, which is expressed as follows, where L represents the loss function, i.e., cross entropy, y represents the true label, and σ(x) represents the sigmoid activation output by the model:

[0060] L=-[yln(σ(x))+(1-y)ln(1-σ(x))].

[0061] Step 3 is specifically as follows: when using the optimized NeuroSAT network model for reasoning, each fault point in the new netlist file is converted into a CNF formula and encoded into an undirected graph containing text nodes and clause nodes; the encoded undirected graph is input into the optimized NeuroSAT network model to obtain the corresponding fault testability prediction value.

[0062] Step 4 is as follows:

[0063] For the fault points determined to be testable faults, automatic test vector generation is performed;

[0064] For faults that are determined to be untestable, a Boolean satisfiability analysis is performed and the test vector generation process is skipped.

[0065] After automatic test vector generation, fault simulation is performed to verify the validity of the generated test vectors; based on the fault simulation and Boolean satisfiability solution analysis results, the quality of the NeuroSAT network model is evaluated, and these results can be selectively used to further optimize the model.

[0066] In order to facilitate the understanding of the present invention, the following four aspects are explained separately as follows:

[0067] (1) Integrated circuit fault testability prediction process based on graph neural network.

[0068] A complete automatic test vector generation process is divided into four stages, namely, reading the netlist, generating a fault list, generating test vectors, and detecting faults; because the traditional automatic test vector generation system cannot predict the fault testability in advance, it leads to the waste of resources on untestable faults. To solve this problem, the present invention proposes a method for predicting the fault testability of integrated circuits based on graph neural networks. The NeuroSAT network model adopted is combined with a dedicated CNF fault data set. By utilizing the advantages of the NeuroSAT model in processing structured data, it can effectively capture the complex relationship between circuit topology and fault characteristics. Compared with the traditional SAT solver, the NeuroSAT graph neural network encodes the SAT problem into an undirected graph containing text nodes and clause nodes. This representation method retains the structural information of the problem, helps to fully understand the circuit logic characteristics, enables the model to learn deeper fault characteristics, and effectively reduces unnecessary test vector generation overhead, thereby optimizing the overall test process and improving test efficiency and quality.

[0069] (2) CNF fault dataset generation process.

[0070] Reference Figure 2The process of generating CNF fault data set is as follows: parse the netlist file, extract the circuit structure information, identify all logic gates, input and output ports, and establish the internal connection relationship of the circuit. Then convert the circuit logic into CNF clauses, and perform CNF conversion on each logic gate: for example, AND gate (A∧B)→C is converted into OR gate: (a∨b)→c is converted to Next, enumerate each possible failure point, consider stuck-at-0 and stuck-at-1 failures, and add failure activation clauses; for stuck-at-0 failure: add Clauses, where f is the fault activation variable and x is the original variable. For stuck-at-1 faults: add (f→x) clauses. Finally, generate fault propagation path constraints, add sensitization constraint clauses for each gate in the fan-out cone and combine the clauses of the CNF formula. Figure 2 As shown in the figure, after the CNF fault injection is completed, it is divided into two steps: one is to convert the CNF formula into an LCG graph structure for graph neural network learning, and the other is to use the solver to solve the CNF to obtain the label corresponding to the data. Then, the CNF fault data set is divided into a training set, a validation set, and a test set at a ratio of 80%, 10%, and 10%, respectively.

[0071] Reference Figure 3 Taking the s38584 netlist in the benchmark test circuit ISCAS-89 as an example, after executing the conversion script, some CNF formulas generated by circuit fault conversion are as follows Figure 3 As shown, the file format is the international standard of DIMACS.

[0072] It should be noted that the advantage of using CNF formulas as graph neural network datasets is that the CNF format, as a standard logical representation method, can describe arbitrarily complex Boolean functions. This allows the dataset to adapt to integrated circuits of various types and sizes, from simple combinational logic to complex sequential circuits, which can be uniformly represented. As the standard input format for most SAT solvers, this compatibility can leverage existing efficient SAT solving techniques to generate accurate fault testability labels. The CNF representation retains the topological structure information of the circuit, and the relationships and constraints between variables are explicitly encoded in the clauses, which provides an ideal input format for models such as graph neural networks that can utilize structural information.

[0073] (3) Graph neural network model training process.

[0074] Reference Figure 4 , which represents the process of NeuroSAT model training SAT solution by message passing on the graph structure. Figure 3As shown, the SAT problem is expressed as {1|2,1|2}, which is equivalent to ∨x2). Where x1, x2, is a literal, c1=(x1∨x2), First, the network model initializes all text nodes with the same vector and all clause nodes with the same vector. Then, the iterative process (t→t+1) updates the clause embedding: for each clause c, messages from adjacent texts are collected and aggregated, and then the clause embedding is updated. The iterative process also updates the text embedding: for each text l, messages from adjacent clauses are collected and aggregated, and then the text embedding is updated. Finally, after T iterations, the model calculates the voting value for each text l and finally outputs the result.

[0075] It should also be noted that the advantage of training the SAT solution mode through message passing is that the original structural information of the problem can be retained, which is particularly suitable for processing problems with relational or structured data. Through iterative message passing, the model can gradually build a global understanding from local information, simulating the process of human solving complex problems. Using the same message passing and update function to process all nodes reduces the number of parameters that need to be learned and improves the efficiency and generalization ability of the model. The graph structure and message passing mechanism strengthen the inductive bias of the model, allowing the model to learn the essential characteristics of the problem rather than overfitting the surface features.

[0076] Furthermore, the message passing process can be visualized and explained, which helps to understand the decision-making process of the model. The model is highly flexible and scalable, and can handle graphs of different sizes and structures. By adjusting the number of iterations, a trade-off can be made between reasoning depth and computational cost. The entire message passing process is differentiable, allowing end-to-end gradient descent learning. Through multiple rounds of message passing, the model can effectively capture long-range dependencies, which is a significant advantage when dealing with complex logical problems. This learning process allows the model to dynamically adapt to the complexity of the input, providing an efficient and flexible framework for dealing with complex structured problems.

[0077] (4) Boolean satisfiability analysis process.

[0078] Reference Figure 5 For the faults that are determined to be untestable, Boolean satisfiability analysis can be performed based on the statistical data of the results output by the solver. Figure 5As shown in the runtime analysis section, the model spends most of its time parsing (99%) and searching (91.25%); in particular, most of the time during the search process (86.52%) is in an unstable state, which indicates that this is a relatively difficult problem and the solver needs to perform a lot of exploration and backtracking. According to the data provided in the statistics section, 73 conflicts occurred during the solution process, 281 decisions were made, 94% of the variables were determined in the preprocessing stage, and 98.63% of the conflicts generated learning clauses. The solver learned a total of 72 clauses, containing 1074 literals, 84.92% of the learned clauses were simplified during the backoff, and the message propagation speed was 1.03 million times per second. It is worth noting that there was 1 restart, with a restart interval of 73 conflicts, and 9 clauses (accounting for 0.1% of all clauses) were deleted during the propagation process because they were redundant. The resource usage section shows that the process time and actual time of the entire solution process are close to 0, and the maximum memory usage is 128.05MB.

[0079] It should also be noted that each line of the solution process represents an important event: including, '*' indicates that an unsatisfiable clause is found; '{' and '}' indicate entering and exiting the decision layer; numbers show the current running statistics, such as the number of conflicts, the number of learned clauses, etc. When closing the proof file, the solver reports that 185 clauses were added to the proof and 34 clauses were deleted. In general, the result statistics output by the solver show the detailed process of the model when dealing with a relatively difficult Boolean satisfiability problem. The solver spent a lot of time searching, experienced many conflicts and restarts, and learned many new clauses.

[0080] Reference Figure 6 In the process of Boolean satisfiability analysis, the quality of NeuroSAT network model can be better evaluated based on the statistical data of model learning process. Figure 6As shown, the second column represents the solution running time, the fourth column represents the current decision level, and the decision level of '9' reflects the depth of the solver in the search tree. A high decision level indicates that the solver is exploring a deep search path. The problem may be complex and require multiple decisions or backtracking to find a solution or conflict. The fifth column represents the number of restarts, and the eighth column represents the number of redundant clauses learned. The ninth column represents the average viscosity, which is an indicator used in the CDCL (Conflict-Driven Clause Learning) SAT solver to evaluate the quality of learned clauses. The average viscosity reflects the overall quality of the learned clauses. A high viscosity of '99%' means that they can only have an impact at higher decision levels. By observing the changes in the decision level and average viscosity, we can understand the efficiency and learning quality of the solver during the search process. Ideally, the decision level should be kept within a reasonable range, while the average viscosity is low, which means that the solver can effectively explore the search space and learn high-quality clauses. The solver can adjust its learning strategy or parameters in turn to achieve the best adaptability of the decision level and average viscosity.

[0081] In order to better verify and illustrate the technical effects used in the method of the present invention, this embodiment selects the traditional Boolean satisfiability automatic test vector generation process and the model trained on the ordinary SR data set to compare and test the method of the present invention to demonstrate the superiority of the graph neural network testability prediction. The test results are compared by means of scientific demonstration to verify the real effect of the method of the present invention.

[0082] The traditional Boolean satisfiability automatic test vector generation process cannot verify the testability of the fault to be tested. In order to verify that the method of the present invention has higher efficiency than the traditional method and the ordinary graph neural network model solver, this embodiment will use the traditional Boolean satisfiability automatic test vector generation, the ordinary graph neural network model solver and the method of the present invention to perform data measurement and comparison on the benchmark circuit test samples respectively.

[0083] Table 1: Benchmark circuit test sample data statistics.

[0084]

[0085]

[0086] Referring to Table 1, compared with the traditional method and the common graph neural network model solving method, the backtracking probability of the S35932 circuit is reduced by up to 17%, and the decision-making level is reduced by 45.2%. The backtracking probability of the S38417 circuit is reduced by up to 10%, and the decision-making level is reduced by 48.4%. Therefore, the fault testability prediction based on the CNF fault data set can effectively reduce the backtracking probability and optimize the decision-making strategy.

[0087] Preferably, compared with the traditional method and the common graph neural network model solution method, the present invention reduces the solution time for the S35932 circuit by up to 16%. The solution time for the S38417 circuit is reduced by up to 12%. Therefore, the fault testability prediction based on the CNF fault data set can effectively reduce the solution time, thereby achieving the purpose of reducing the running time of ATPG and saving test costs.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to be a further limitation of the present invention. All equivalent changes made using the contents of the present specification and drawings are within the protection scope of the present invention.

Claims

1. A method for predicting the fault testability of integrated circuits based on graph neural networks, characterized in that: The following steps are involved: Step 1: read the netlist file, create a CNF fault data set, and randomly divide the CNF fault data set into a training set, a validation set, and a test set according to a preset ratio; Step 2: construct a NeuroSAT graph neural network model, input the training set and the validation set into the NeuroSAT network model for training and validation, obtain a training weight file, and adjust the parameters based on the loss function to optimize the NeuroSAT network model; Step 3: Read in the new netlist file, enumerate the faults and use the optimized NeuroSAT network model to infer and obtain the fault testability; Step 4: According to the fault testability, decide whether to perform automatic test vector generation or Boolean satisfiability analysis, and further optimize the fault testability prediction model according to the result.

2. According to claim 1, a method for predicting integrated circuit fault testability based on graph neural network is characterized in that: In step 1, when making the CNF fault data set, each logic gate in the circuit is converted into an equivalent CNF clause, and a sensitization constraint clause and a fault activation clause are added to each gate in the fault fan-out cone; The CNF fault data set is divided into a training set, a validation set and a test set.

3. According to claim 2, a method for predicting the fault testability of integrated circuits based on graph neural networks is characterized in that: The NeuroSAT graph neural network model includes an input processing layer, two initialization vector layers, two multi-layer perceptrons, two normalization layers and an output layer; The input processing layer encodes the CNF formula into an undirected graph including literal nodes and clause nodes; The initialization vector layer is used to initialize the embedding vectors of all text nodes and clause nodes in the undirected graph; The multilayer perceptron is used to process the messages of the initialized text nodes and perform T rounds of iterations; The normalization layer receives messages from adjacent text nodes after T rounds of iterations and updates the embedding; The output layer calculates the average of all updated embedding word votes as the final prediction output.

4. The method for predicting integrated circuit fault testability based on graph neural network according to claim 3, characterized in that: Initialize the predefined vector L used by the vector layer init The initialization text node expression is as follows, where L (t) represents the text embedding matrix at the tth iteration, 2n represents the number of words, d represents the embedding dimension, Initialize the predefined vector C used by the vector layer init The initialization clause node expression is as follows, where C (t) represents the clause embedding matrix at the tth iteration, m represents the number of clauses, d represents the embedding dimension, The update clause node embedding expression is as follows, where C_u represents the network of the normalized layer update clause embedding, C_h (t) : The hidden state of the network, M T represents the transpose of the adjacency matrix, L_msg represents the multi-layer perceptron layer processing the transmission of text messages; The expression for updating the text node embedding is as follows, where the L_u normalization layer represents the network for updating the text embedding, and L_h (t) represents the hidden state of the network, Flip represents the function of exchanging complementary words, M represents the adjacency matrix, and C_msg represents the transmission of clause messages processed by the multi-layer perceptron layer; The loss function used in the output layer is the sigmoid cross entropy loss function, which is expressed as follows, where L represents the loss function, i.e., cross entropy, y represents the true label, and σ(x) represents the sigmoid activation output by the model: L=-[yln(σ(x)))+(1-y)ln(1-σ(x))].

5. The method for predicting integrated circuit fault testability based on graph neural network according to claim 1, characterized in that: Step 3 is specifically as follows: when using the optimized NeuroSAT network model for reasoning, each fault point in the new netlist file is converted into a CNF formula and encoded into an undirected graph containing text nodes and clause nodes; the encoded undirected graph is input into the optimized NeuroSAT network model to obtain the corresponding fault testability prediction value.

6. The method for predicting integrated circuit fault testability based on graph neural network according to claim 1, characterized in that: Step 4 is as follows: For the fault points determined to be testable faults, automatic test vector generation is performed; For faults that are determined to be untestable, a Boolean satisfiability analysis is performed and the test vector generation process is skipped.

Citation Information

Cited By

  • Automatic test vector generation method and system based on decision tree

    CN119780685A

  • Automatic Test Vector Generation Method and System Based on Decision Tree

    CN119780685B

  • Digital circuit test fan-in decision-making method based on relational graph convolution and reinforcement learning

    CN122613138A