Multi-objective optimization method for relieving single-particle soft error of large-scale integrated circuit

Abstract

The invention discloses a multi-objective optimization method for relieving a single-particle soft error of a large-scale integrated circuit, and the method achieves the soft error prediction of the large-scale integrated circuit based on a graph neural network algorithm, and obtains a circuit soft error rate and a failure type. Based on a soft error prediction result and a genetic algorithm, a multi-objective optimized Pareto optimal solution set under various reinforcement measures (such as TMR and reinforcement FF unit replacement) is realized, so that the soft error of a large-scale integrated circuit is relieved, and the overhead of extra area and power consumption caused by reinforcement design is balanced. The optimization strategy of the large-scale integrated circuit is formulated through the combination of the two algorithms, and the comprehensive performance of the large-scale integrated circuit for resisting the single-particle soft error, the area and the power consumption is improved through the optimal strategy of the combination of multiple reinforcing measures on the premise of ensuring that the circuit avoids serious harm caused by the single-particle effect. The method can be used for providing guidance for radiation hardening design and optimization of a large-scale integrated circuit.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of reliability evaluation and optimization of radiation effects of integrated circuits for spaceflight, and particularly relates to a multi-objective optimization method for relieving single event soft errors of large-scale integrated circuits. BACKGROUND

[0002] In the space application of integrated circuits, soft errors caused by single event effects are one of the main factors leading to abnormal functions or failures of systems, and therefore, design reinforcement is needed to improve the reliability of integrated circuits in space applications. For example, triple modular redundancy (TMR), cell library with reinforcement structure, and coding reinforcement. These reinforcement designs can greatly ensure that the spaceflight integrated circuits are not disturbed by radiation effects in the space environment, but they inevitably bring higher area and power consumption overheads. Moreover, as the scale of integrated circuits gradually increases, it becomes very difficult to achieve evaluation and balance of various indicators relying only on manpower and experience.

[0003] The anti-radiation performance evaluation of spaceflight integrated circuits is an important basis for guiding reinforcement optimization design. At present, in view of the problems of complex soft error characteristics and large simulation resource overhead of large-scale integrated circuits (integrated circuits with more than one million gates), the soft error rate prediction method based on machine learning has been developed at home and abroad, but the prediction range is mostly limited to the analysis and prediction of the overall soft error rate of the chip, and the positioning and danger level division of sensitive nodes are ignored, which cannot guide the reinforcement optimization design. In addition, in terms of reinforcement methods, different reinforcement methods have different area and power consumption overheads. In most cases, the reinforcement method with high performance has more area and overhead. A single reinforcement scheme may make the design meet a certain indicator as much as possible, but it cannot achieve comprehensive improvement of multiple targets.

[0004] Based on the anti-radiation performance of integrated circuits, there is a certain restrictive relationship between the area and power consumption design targets. How to efficiently realize the performance evaluation and comprehensive optimization design of large-scale integrated circuits is one of the key problems to be solved in the field of spaceflight integrated circuit design and optimization. SUMMARY

[0005] The technical problem solved by the present application is to overcome the shortcomings of the prior art and provide a multi-objective optimization method for relieving single event soft errors of large-scale integrated circuits, which solves the problem that the anti-single event soft error performance, area and power consumption three indicators are difficult to balance when the large-scale integrated circuits for spaceflight are designed and optimized, and realizes comprehensive performance improvement of the circuit under the premise of ensuring the anti-radiation performance of the spaceflight integrated circuit.

[0006] The technical solution of the present application is a multi-objective optimization method for relieving single event soft errors of large-scale integrated circuits, which comprises the following steps:

[0007] S1, represent the target large-scale integrated circuit gate-level netlist Netlist as an undirected graph G=(V, E), wherein the cells in the netlist Netlist are converted into nodes V, and the line nets are converted into edges E;

[0008] S2, define the characteristics of the nodes V by using cell characteristics and structure characteristics, and aggregate the characteristics of each node by using a pool aggregator, so as to convert the undirected graph G=(V, E) into a GraphSAGE graph;

[0009] S3, take the GraphSAGE graph generated in step S2 as input, and respectively use the trained neural network model to predict the soft error sensitive node list of the target circuit netlist, the overall soft error rate of the chip, the area, and the power consumption; the soft error sensitive node list includes dangerous sensitive points and non-dangerous sensitive points;

[0010] S4, TMR reinforcement is performed on the dangerous sensitive points in the soft error sensitive node list of the target circuit netlist, and a gate-level netlist Netlist_H after reinforcement is generated, and then the overall soft error rate of the chip, the area, and the power consumption are predicted again;

[0011] S5, according to the sensitive node list generated in step S3 and the Netlist_H generated in step S4, the non-dangerous sensitive points in the sensitive node list are replaced by reinforcement cells selected from a reinforcement cell library, and the NSGA II fast non-dominated genetic multi-objective optimization method is used for iterative calculation to obtain a pareto optimal solution set, and a reinforcement cell replacement strategy is generated according to the optimal solution set;

[0012] S6, the reinforcement cell replacement strategy determined in step S5 is used to replace the cells in the Netlist_H generated in step S4, and the soft error, the area, and the power consumption of the netlist after reinforcement are obtained.

[0013] Preferably, the specific steps of step S5 are as follows:

[0014] S5.1, extract the non-dangerous sensitive points in the sensitive node list;

[0015] S5.2, construct a reinforcement cell library, the reinforcement cell library at least includes three reinforcement measures: TMR, FFH, and FF0, TMR represents triple modular redundancy, FFH is a reinforcement cell, and FF0 represents that the cell is not reinforced; each reinforcement measure is represented by a reinforcement measure vector, and the reinforcement measure vector has a size of 1xM;

[0016] S5.3, construct a reinforcement coding matrix used to represent the reinforcement strategy, the reinforcement coding matrix has N rows and M columns, and N rows identify the number of non-dangerous sensitive points.

[0017] S5.4、with the reinforcement coding matrix X as the decision variable, with the soft error, area, and power consumption as the optimization target, defining the target weight, constructing the total target function, taking the number of "not reinforcing FFO for the unit" less than <the number to be replaced as the constraint condition, applying the NSGA II fast non-dominated genetic algorithm multi-objective optimization method, performing a limited number of iteration calculations, and obtaining a pareto optimal solution set;

[0018] S5.5、According to the pareto optimal solution set obtained in step S5.4, generate a reinforcement unit replacement strategy, replace the units in the Netlist_H generated in step S4, and obtain the soft error, area, and power consumption of the reinforced netlist.

[0019] Preferably, the target function is:

[0020] minf M (X)=af SER (X)+bf A (X)+cf P (X)

[0021] Wherein, f SER (X), f A (X), f P (X) are the soft error rate, area and power consumption optimized based on the reinforcement coding matrix X, a, b, c are the weights of the soft error rate, area and power consumption, and f M (X) is the total target function.

[0022] Preferably, the NSGA II fast non-dominated genetic algorithm combined with the GNN neural network prediction model multi-objective optimization method is as follows:

[0023] S5.1, extract non-dangerous sensitive points in the sensitive list;

[0024] S5.2, randomly generate a reinforcement coding matrix X, and evaluate the optimized soft error rate f SER (X), area f A (X), and power consumption f P (X);

[0025] S5.3, the randomly generated reinforcement coding matrix is a particle, and the initial population P0 is initialized, and the population size is K;

[0026] S5.4, apply the NSGA II fast non-dominated genetic algorithm to the particles in the initial population P0 for fast non-dominated sorting;

[0027] S5.5, select, cross, and mutate the population after non-dominated sorting to generate the next generation population Q0, and the population size is K.

[0028] S5.6, adopt elite preservation strategy, the parent P0 is fused with the offspring population Q0 and generates R0, at this time, the population number is 2K, the individual in population R0 is rapidly non-dominated sorted;

[0029] S5.7, the crowdedness of the sorted population is calculated, the first K population is taken to generate new parent population P1, P1 is used as the initial population, and steps S5.5-S5.9 are repeatedly executed until the maximum iteration number is met, and new parent population P1 is the pareto optimal solution.

[0030] Preferably, the sorting principle is: for any optimization strategy X i , X j , if the following conditions are met:

[0031]

[0032] X i dominates X j , denoted as Where X i is the dominated solution, and X j is the non-dominated solution.

[0033] Preferably, the number of particles in the step S5.3 is greater than or equal to 10000.

[0034] Preferably, the iteration number is greater than or equal to 100.

[0035] The trained neural network model is a GNN neural network model.

[0036] Another technical solution provided by the application is to provide a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the steps of the method of the first technical solution.

[0037] The fourth technical solution provided by the application is a terminal device, which includes a memory and a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to realize the steps of the method of the first technical solution.

[0038] Compared with the prior art, the application has the following advantages:

[0039] (1), the application realizes efficient prediction of three targets of large-scale integrated circuit single particle soft error, area and power consumption based on GNN, and combined with NSGA II algorithm, fast solution of multiple targets can be realized.

[0040] (2), the application realizes the prediction of the sensitive unit danger level based on GNN prediction, and carries out the "two-step" reinforcement design strategy according to the danger level, so that the comprehensive performance of the circuit can be improved under the premise of ensuring the radiation resistance of the integrated circuit;

[0041] (3), the application converts the reinforcement measures into a matrix representation through matrix encoding, supports multi-objective optimization solution of multiple reinforcement measures, and can be expanded according to actual design needs, and when applied to NSGA II algorithm fusion, matrix operation can be applied to realize fast evaluation of the objective function;

[0042] (4), the application applies the NSGA II fast non-dominated genetic algorithm to reserve the population diversity by using the elite reservation strategy, obtains the Pareto front solution under the three targets of the target circuit single event soft error rate, area and power consumption, and the reinforcement strategy obtained through optimal solution mapping can effectively guide the comprehensive optimization design of large-scale integrated circuits. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The embodiment of the application is based on the multi-objective optimization algorithm flowchart of GNN and NSGA II.

[0044] Fig. 2(a) is a GraphSAGE node feature aggregation and embedding of the embodiment of the application;

[0045] Fig. 2(b) is a neural network prediction model of the embodiment of the application;

[0046] Figure 3 The multi-objective optimization algorithm flowchart based on NSGA II. DETAILED DESCRIPTION

[0047] The application will be described in detail below in combination with specific embodiments.

[0048] The application provides a multi-objective optimization method for relieving single event soft error of large-scale integrated circuits, which comprises the following steps:

[0049] S1, the target large-scale integrated circuit gate netlist Netlist is represented as an undirected graph G=(V, E), wherein the unit in the netlist Netlist is converted into a node V, and the line net is converted into an edge E;

[0050] The structure and characteristics of the target large-scale integrated circuit netlist are extracted by using VPI, the unit name, the input and output connection names connected to the unit are extracted, the unit name is converted into a node, and the connection relationship between the units is converted into an edge connected to the node, so as to realize the conversion of the node and the edge;

[0051] S2, the undirected graph G=(V, E) generated according to step S1 is converted into a GraphSAGE graph by defining the node V features with structure features and unit features and applying a pool aggregator to realize node feature aggregation.

[0052] The structure features and unit features include the following aspects:

[0053] 1) The structure features mainly refer to the connection relationship between units based on the circuit netlist extraction, mainly including hierarchical structure, upstream and downstream, etc.

[0054] 2) The unit features mainly refer to the intrinsic properties of units extracted based on the unit library, mainly including unit type, port name, soft error rate, area, power consumption, etc.

[0055] 3) GraphSAGE uses node features to learn an embedding function that can be extended to invisible nodes. Based on the obtained target node v feature, the node information of the neighbor nodes of the target node v is obtained by sampling, and then the new representation of the target node v is obtained by aggregating the representations of the neighbor nodes.

[0056] S3, the GraphSAGE graph generated in step S2 is taken as input, and trained neural network models are used to realize the prediction of the soft error sensitive node list of the target circuit netlist, the overall soft error rate of the chip, the area, and the power consumption, and the features include the following aspects:

[0057] The neural network model has four, respectively used to predict the soft error sensitive node list, the chip soft error rate, the area, and the power consumption;

[0058] The neural network of the soft error sensitive node list distinguishes between failure types, including two types of dangerous nodes and non-dangerous nodes. Through dynamic fault injection simulation, based on the failure types caused by some units after fault injection simulation, a training sample set is constructed, and the failure types caused by the remaining units are predicted through the neural network model, and finally the soft error sensitive node list of the whole chip is obtained. The soft error sensitive node list includes dangerous sensitive points and non-dangerous sensitive points;

[0059] For M dangerous sensitive point lists FFS, represented as {FFS1, FFS2,..., FFS M}。

[0060] For N non-dangerous sensitive point lists FFR, represented as {FFR1, FFR2,..., FFR N}。

[0061] Dangerous sensitive points are defined as units that cause failure types defined as dangerous;

[0062] The neural network model for predicting the overall single-event soft error rate of the chip comprises a failure masking rate and an overall soft error rate of the chip. The failure masking rate is obtained based on failure injection simulation. The failure masking rate of the remaining units is predicted based on the probability of failure caused by the part of the units after the failure injection simulation, and a neural network pre-model is applied to finally obtain the failure masking rate of all the units. The overall soft error rate SER of the chip is realized based on the failure masking rate C0 The SER can be expressed as follows:

[0063] SER C0 =∑FF i _SER×DR i

[0064] wherein, FF i _SER is the intrinsic soft error rate of the i-th unit, DR i is the masking rate corresponding to the i-th unit, and i = 1 ~ M + N.

[0065] The neural network model for predicting the chip area is trained based on the unit area and the layout file, and the area A C0 of the chip after layout and routing is predicted.

[0066] The neural network model for predicting the chip power consumption is trained based on the unit power consumption and the synthesis report file, and the chip power consumption P C0 is predicted.

[0067] S4, according to the soft error sensitive node list of the target circuit netlist generated in step S3, the dangerous sensitive points are TMR reinforced, and a reinforced circuit netlist Netlist_H is generated. Then, the three target values of the chip soft error rate, area and power consumption are re-predicted, which specifically includes the following aspects:

[0068] According to the soft error sensitive node list of the target circuit netlist generated in step S3, the units {FFS1, FFS2,..., FFS M} whose failure types are defined as dangerous are extracted for TMR reinforcement, wherein the failure type is defined by the designer or user according to the application scenario; the failure type can be functional interruption and the like.

[0069] The neural network model is applied to the netlist Netlist_H after TMR reinforcement to re-predict the chip soft error rate SER CH , area A CH and power consumption P CH ;

[0070] For the nodes replaced by TMR, it can be determined that they will not cause chip failure.

[0071] S5, generating the sensitive node list according to step S3 and the Netlist_H generated in step S4, taking the overall chip soft error, area and power consumption as the optimization target, replacing the non-dangerous sensitive points in the sensitive node list with the selected reinforcement unit in the reinforcement unit library for reinforcement optimization, iteratively calculating using the NSGA II fast non-dominated genetic multi-objective optimization method to obtain a pareto optimal solution set, and generating a reinforcement unit replacement strategy according to the optimal solution set;

[0072] This step includes the following aspects:

[0073] 1) According to the sensitive list FFR generated in step S3, extract the non-dangerous sensitive points in the list as the basis for subsequent reinforcement replacement;

[0074] 2) According to the Netlist_H generated in step S4, take this netlist as the object of multi-objective optimization, and solve the optimization strategy of the three targets of SER, area and power consumption;

[0075] 3) Replace the design optimization with the reinforcement library unit, build a reinforcement unit library, and the reinforcement unit library includes at least three reinforcement measures: TMR, FFH and FF0, TMR represents triple modular redundancy, FFH is a reinforcement unit, and FF0 represents no reinforcement for the unit; wherein FFH can be a specific set of S-class reinforcement units, or a comprehensive performance representation of reinforcement units; each reinforcement measure is represented by a 1xM reinforcement measure vector;

[0076] 4) If FFH is a specific set of S-class reinforcement units, the reinforcement strategy can be further represented as {TMR, FFH1, FFH2,..., FFS, FF0}.

[0077] The method of iteratively calculating using the NSGA II fast non-dominated genetic multi-objective optimization method is as follows:

[0078] A reinforcement coding matrix is constructed to represent the reinforcement strategy, and the reinforcement coding matrix is N rows by M columns, N rows identify N as the number of non-dangerous sensitive points, and M columns are the size of the reinforcement measure vector;

[0079] The reinforcement coding matrix can be customized according to the reinforcement strategy. If the reinforcement strategy is {TMR, FFH, FF0}, TMR is [1 0 0], FFH is [0 1 0], FF0 is [0 0 1], and the non-dangerous unit to be replaced is N unit set {FF1, FF2,..., FF N}, the corresponding reinforcement strategy FF HS can be represented as an N x 3 matrix. For example:

[0080]

[0081] If the reinforcement strategy is FFH containing S type reinforcement units, the non-dangerous units to be replaced are N unit set {FF1, FF2,..., FF N}, the corresponding reinforcement strategy FF HS Can be represented as an N x (S + 2) matrix. For example:

[0082]

[0083] The reinforcement encoding matrix X is taken as the decision variable, the soft error, area and power consumption are taken as the optimization target, the target weight is defined, the total target function is constructed, the number of units not reinforced FF0 is less than the number of units to be replaced as a constraint condition, the NSGA II fast non-dominated genetic algorithm multi-objective optimization method is applied, the iteration calculation is carried out for a limited number of times, and the pareto optimal solution set is obtained;

[0084] The soft error rate, area and power consumption three target functions are f SER (FF HS ), f A (FF HS ), f P (FF HS ), and the total target function f M (X) is represented as:

[0085] The target function is:

[0086] minf M (X)=af SER (X)+bf A (X)+cf P (X)

[0087] Wherein, f SER (X), f A (X), f P (X) are the soft error rate, area and power consumption based on the reinforcement encoding matrix X optimization, a, b, c are the weight of soft error rate, area and power consumption, f M (X) is the total target function. The weight of three target functions can be defined.

[0088] The variable X is the reinforcement encoding matrix;

[0089] The constraint condition is that the number of rows of the matrix X needs to be equal to the number of non-dangerous units to be replaced, and FF0 < N in the strategy adopted;

[0090] The multi-objective optimization method combining the NSGA II fast non-dominated genetic algorithm with the GNN neural network prediction model is as follows:

[0091] S5.1, extract non-dangerous sensitive points in the sensitive list;

[0092] S5.2, randomly generate reinforcement encoding matrix X, and evaluate the optimized soft error rate f SER (X), area f A (X), power consumption f P (X);

[0093] The optimized soft error rate f SER (X), area f A (X), power consumption f P (X) Target evaluation algorithm is described as follows:

[0094] a. For N non-dangerous sensitive point units {FF1, FF2, …, FF N}, the corresponding masking rate matrix DR I of each unit, based on the reinforcement encoding matrix X optimized soft error rate f SER (X) is expressed as:

[0095] f SER (X) = [SER TMR SER FFH SER FF0 ]*X T *DR I

[0096] Where SER TMR represents the TMR structure soft error rate, SER FFH represents the reinforcement unit soft error rate, SER FF0 represents the unit soft error rate without reinforcement measures.

[0097] b. For N non-dangerous sensitive point units {FF1, FF2, …, FF N}, apply a 1XN matrix A with all elements being 1 to the optimization strategy FF HS Each column is summed, and the area f A (X) based on the reinforcement encoding matrix X optimized is expressed as:

[0098] f A (X) = A*X*[A TMR A FFH A FF0 ] T

[0099] Where A is a 1XN matrix with all elements being 1, A TMR represents the area of TMR structure, A FFH represents the area of FFH unit, A FF0 represents the area of the unit without reinforcement measures, that is, the area of the original unit

[0100] c. For N non-dangerous sensitive point units {FF1, FF2,..., FF N}, apply a 1xN matrix A with all elements being 1 to the optimization strategy FF HS Sum each column, and the optimized power consumption based on the hardened encoding matrix X is represented as:

[0101] f P (X)=A*X*[P TMR P FFH P FF0 ] T

[0102] Wherein, P TMR represents the power consumption of the TMR structure, P FFH represents the power consumption of the hardened unit FFH, and P FF0 represents the power consumption of the unit without adopting the hardened measure, i.e. the power consumption of the original unit.

[0103] S5.3, randomly generate a hardened encoding matrix X j , j = 1 ~ 10000 are particles, and the initial population P0 has a population size of K;

[0104] S5.4, apply the NSGA II fast non-dominated genetic algorithm to perform fast non-dominated sorting on the particles in the initial population P0;

[0105] The application of the NSGA II fast non-dominated genetic algorithm includes fast non-dominated sorting. For any hardened encoding matrix X m , the hardened encoding matrix X n satisfies the following conditions:

[0106]

[0107] The hardened encoding matrix X m is said to dominate the hardened encoding matrix X n , denoted as Wherein, the hardened encoding matrix X m is a dominated solution, and the hardened encoding matrix X n is a non-dominated solution.

[0108] S5.5, apply selection, crossover, mutation and other operations to the sorted population, wherein the crossover operation applies the simulated binary crossover (SBX), and the mutation operation applies the polynomial mutation method (Ploynomial utation), to generate the next generation population Q0, which has a population size of K;

[0109] S5.6, adopt elite reservation strategy, the parent P0 is fused with the offspring population Q0 and generates R0, at this time the population number is 2K, the individual in population R0 is rapidly non-dominated sorting.

[0110] S5.7, the population after sorting is calculated for crowdedness, the first K population before sorting is taken to generate new parent population P1, P1 is used as the population in initialization, the steps S5.4-S5.7 are repeatedly executed, and the iteration operations such as crossover and mutation are continued, until the maximum iteration number is satisfied, and new parent population P1 is the pareto optimal solution.

[0111] The pareto optimal solution and optimal solution set: for matrix X, there is no other Z * So that Z * Dominates Z, Z is called the pareto optimal solution, and the set of all optimal solutions is the pareto optimal solution set.

[0112]

[0113] Wherein, Ω represents the set of all possible solutions, Z represents any one solution, and F(Z) is a multi-objective function.

[0114] S6, the reinforcement unit replacement strategy determined in the step S5 is used to replace the Netlist_H generated in the step S4, and the soft error, area and power consumption of the netlist after reinforcement are obtained.

[0115] According to the optimal solution FF HS And the corresponding relationship of the replacement unit, the corresponding reinforcement measures are implemented on Netlist_H.

[0116] The application also provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by a processor to realize the steps of the method. Figure 1 The steps of the method.

[0117] The application provides a terminal device, which comprises a memory and a processor and a computer program stored in the memory and executable on the processor, and the processor realizes the steps of the method when executing the computer program. Figure 1 The steps of the method.

[0118] In summary, the present application relates to single particle effect soft error simulation analysis and prediction, anti-radiation reinforcement design and performance optimization, in particular, for the design requirements of large-scale integrated circuit anti-radiation reinforcement design and comprehensive performance optimization, based on NSGA II and GNN neural network, a multi-objective optimization algorithm for relieving single particle soft error of large-scale integrated circuit is proposed, the solution of Pareto optimal solution is used to guide the circuit design optimization, and the comprehensive performance of the integrated circuit for spaceflight in radiation resistance, area and power consumption is improved.

[0119] The contents not described in detail in the specification of the present application are the known technology of those skilled in the art.

Claims

1. A multi-objective optimization method for mitigating single-event soft errors in large-scale integrated circuits, characterized in that... Includes the following steps: S1. Represent the target large-scale integrated circuit gate-level netlist as an undirected graph G = (V, E), where the cells in the netlist are converted into nodes V and the wire nets are converted into edges E; S2. Define the features of node V using unit features and structural features, and apply a pool aggregator to aggregate the features of each node, converting the undirected graph G = (V, E) into a GraphSAGE graph. S3. Using the GraphSAGE graph generated in step S2 as input, the trained neural network model is used to predict the soft error sensitive node list of the target circuit netlist, the overall soft error rate of the chip, the area, and the power consumption; the soft error sensitive node list includes dangerous sensitive points and non-dangerous sensitive points. S4. Perform TMR hardening on the dangerous and sensitive points in the soft error sensitive node list of the target circuit netlist, and generate the gate-level netlist Netlist_H of the hardened circuit netlist. After that, re-predict the overall soft error rate, area and power consumption of the chip. S5. Based on the sensitive node list generated in step S3 and the Netlist_H generated in step S4, with the overall chip soft error, area and power consumption as optimization objectives, for non-dangerous sensitive points in the sensitive node list, hardening optimization is performed by selecting hardening units from the hardening unit library for replacement. The NSGA II fast non-dominated genetic multi-objective optimization method is used for iterative calculation to obtain the Pareto optimal solution set, and a hardening unit replacement strategy is generated based on the optimal solution set. S6. The hardened unit replacement strategy determined in step S5 is used to harden and replace the Netlist_H generated in step S4, and the soft errors, area and power consumption of the hardened netlist are obtained.

2. The multi-objective optimization algorithm for mitigating single-event soft errors in large-scale integrated circuits according to claim 1, characterized in that... The specific steps of step S5 are as follows: S5.1 Extract non-dangerous sensitive points from the list of sensitive nodes; S5.2 Construct a reinforcement unit library, which includes at least three reinforcement measures: TMR, FFH and FF0. TMR represents triple modular redundancy, FFH is a reinforcement unit, and FF0 means that the unit is not reinforced. Each reinforcement measure is represented by a reinforcement measure vector, and the size of the reinforcement measure vector is 1×M. S5.3 Construct a reinforcement coding matrix to represent the reinforcement strategy. The reinforcement coding matrix has N rows and M columns, and N rows indicate the number of non-dangerous sensitive points. S5.

4. Using the hardening coding matrix X as the decision variable, soft error, area, and power consumption as optimization objectives, define the objective weights and construct the overall objective function. With the constraint that the number of "not hardening the unit FF0" is less than the number of units that need to be replaced, apply the NSGA II fast non-dominated genetic algorithm multi-objective optimization method to perform a limited number of iterations to obtain the Pareto optimal solution set. S5.5 Based on the Pareto optimal solution set obtained in step S5.4, generate a hardening cell replacement strategy, perform cell hardening replacement on the Netlist_H generated in step S4, and obtain the soft errors, area, and power consumption of the hardened netlist.

3. The multi-objective optimization algorithm for mitigating single-event soft errors in large-scale integrated circuits according to claim 1, characterized in that, The objective function is: minf M (X)=af SER (X)+bf A (X)+cf P (X) Among them, f SER (X), f A (X), f P (X) represents the soft error rate, area, and power consumption optimized based on the hardened coding matrix X, respectively, where a, b, and c are the weights of the soft error rate, area, and power consumption, respectively, and f M (X) is the overall objective function.

4. The multi-objective optimization algorithm for mitigating single-event soft errors in large-scale integrated circuits according to claim 1, characterized in that, The multi-objective optimization method combining the NSGA II fast non-dominated genetic algorithm with the GNN neural network prediction model is as follows: S5.1 Extract non-dangerous sensitive points from the sensitive list; S5.2 Randomly generate the hardened coding matrix X and evaluate the optimized soft error rate f. SER (X), area f A (X), power consumption f P (X); S5.

3. Use the randomly generated reinforced coding matrix as particles, initialize the population P0, and set the population size to K; S5.4 Apply the NSGA II fast non-dominated genetic algorithm to perform fast non-dominated sorting of particles in the initial population P0; S5.5 Perform selection, crossover, and mutation operations on the non-dominated sorted population to generate the next generation population Q0 with a population size of K; S5.

6. Using an elite retention strategy, the parent population P0 and the offspring population Q0 are merged to generate R0. At this time, the population size is 2K. The individuals in the population R0 are then sorted quickly using a non-dominated method. S5.7 Calculate the crowding degree of the sorted population, take the K populations before sorting to generate a new parent population P1, use P1 as the initial population, repeat steps S5.5 to S5.9 until the maximum number of iterations is met, and make the new parent population P1 the Pareto optimal solution.

5. The multi-objective optimization algorithm for mitigating single-event soft errors in large-scale integrated circuits according to claim 1, characterized in that... The sorting principle is as follows: for any hardened coding matrix X m Strengthening the coding matrix X n If the following conditions are met: The hardened coding matrix X is called m Dominant reinforcement coding matrix X n , denoted as X m >X n Among them, the hardening coding matrix X m To determine the dominant solution, the encoding matrix X is reinforced. n It is a non-dominant solution.

6. The multi-objective optimization algorithm for mitigating single-event soft errors in large-scale integrated circuits according to claim 1, characterized in that, In step S5.3, the number of particles is greater than or equal to 10,000.

7. The multi-objective optimization algorithm for mitigating single-event soft errors in large-scale integrated circuits according to claim 1, characterized in that, The number of iterations is greater than or equal to 100.

8. The multi-objective optimization algorithm for mitigating single-event soft errors in large-scale integrated circuits according to claim 1, characterized in that, The trained neural network model is a GNN neural network model.

9. A computer-readable storage medium storing a computer program, characterized in that... When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

10. A terminal device, comprising a memory and a processor, and a computer program stored in the memory that can run on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 8.

Images