A Chip Radiation Hardening Optimization Matching System Based on Multimodal Data

Abstract

A chip anti-radiation hardening technology optimization and matching system based on multi-modal data, comprising: a chip anti-radiation hardening technology parameter acquisition module, which acquires and stores different types of chip anti-radiation hardening technologies and their corresponding data to construct a multi-modal database; a chip anti-radiation multi-modal hardening technology correlation calculation module, which calculates the correlation of the chip anti-radiation hardening technology optimization scheme to determine whether the scheme is effective; a chip anti-radiation hardening technology optimization and matching module, which automatically matches the eligible hardening technologies in the multi-modal database according to the selected chip model and the provided radiation dose and working environment, forms multiple hardening technology schemes and conducts combined optimization and evaluation on them; and a visualization display module, which displays the chip anti-radiation hardening technology optimization and matching results. The present invention realizes the automatic matching and combined optimization of various chip anti-radiation hardening technologies, provides chip anti-radiation hardening scheme optimization options for R & D personnel, and improves the work efficiency of R & D personnel.

Description

Technical Field

[0001] The present invention belongs to the technical field of chip design, and relates to the selection of chip radiation hardening technologies, in particular to an optimization matching system for chip radiation hardening technologies based on multi-modal data. Background Art

[0002] With the continuous development of electronic technology, the integration of electronic devices is getting higher and higher. Especially chip technology provides support for the functions of various electronic devices. However, during long-term use, chips are prone to radiation damage, and in severe cases, even chip failures may occur. Therefore, chip radiation hardening technology is particularly important.

[0003] At present, many chip radiation hardening technologies have been studied, such as compensation circuits, fault injection, fault masking, etc. However, due to the performance differences of different chips and the differences in radiation environments, the same hardening technology has different effects on different chips. The traditional method for matching chip radiation hardening technologies usually uses prior experience to match hardening technologies. This method lacks data support and is difficult to adapt to the selection of hardening technologies under different chips and different radiation environments. Therefore, the effect is not satisfactory.

[0004] At the same time, due to the large number of hardening technologies, R & D personnel often need to spend a lot of time querying and screening technologies, and often cannot effectively match and optimize the selection of different types of hardening technologies, resulting in low work efficiency and poor hardening effects. Summary of the Invention

[0005] In order to overcome the above-mentioned disadvantages of the prior art, the purpose of the present invention is to provide an optimization matching system for chip radiation hardening technologies based on multi-modal data, which realizes automatic matching and combined optimization among multiple hardening technologies, provides options for chip radiation hardening solutions for R & D personnel, and improves the work efficiency of R & D personnel.

[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0007] An optimization matching system for chip radiation hardening technologies based on multi-modal data, comprising:

[0008] A chip radiation hardening technology parameter acquisition module, which acquires and stores different types of chip radiation hardening technologies and their corresponding hardening material data, hardening process data, and circuit structure diagram data, and constructs a multi-modal database;

[0009] A chip radiation multi-modal hardening technology correlation calculation module, which calculates the technical correlation of the chip radiation hardening optimization solution obtained by integrating and correlating the acquired data to determine whether the solution is effective;

[0010] The chip radiation hardening technology optimization and matching module automatically matches the eligible hardening technologies in the multi-modal database according to the selected chip model, the provided radiation dose, and the working environment, forms multiple hardening technology solutions, and uses the above-mentioned chip radiation multi-modal hardening technology correlation calculation module to evaluate all solutions, and then performs combination optimization and multiple iterations to generate a better chip radiation hardening solution;

[0011] The visualization display module displays the results of the chip radiation hardening technology optimization and matching.

[0012] In one embodiment, in the chip radiation hardening technology parameter acquisition module, the means of data acquisition include:

[0013] R & D personnel experimentally test the relevant parameters and experimental results of the chip radiation hardening and manually upload them;

[0014] Use the embedded crawler tool to regularly crawl data from the specified website;

[0015] Extract data from the manufacturer's data system;

[0016] Standardize and normalize the same-modal data collected by different means, and store them in the multi-modal database according to different data structures.

[0017] In one embodiment, the chip radiation multi-modal hardening technology correlation calculation module sequentially executes:

[0018] (1) For the data corresponding to a certain chip radiation hardening technology, use different deep learning methods to perform feature representation respectively, that is, extract feature vectors, where:

[0019] For the material data X1, extract the feature vector E1 of its text information;

[0020] For the process data X2, extract the feature vector E2_1 of its process text materials and the feature vector E2_2 of its process parameters, and fuse them to obtain the feature vector E2;

[0021] For the circuit structure diagram data X3, extract its feature vector E3;

[0022] (2) Use the linear transformation method to perform dimensionality conversion on E1, E2, and E3, and map them to the same feature vector space. The transformed vectors are expressed as: E'1 = W1 * E1 T , E'2 = W2 * E2 T , E'3 = W3 * E3 T , where W1, W2, and W3 are three linear transformations respectively;

[0023] (3)Fuse the feature information of E'1, E'2, and E'3 in the same feature vector space to obtain the concatenated feature vector E;

[0024] (4)Use 2 Transformer layers to perform multiple non-linear transformations on the feature vector E to extract richer features, and then obtain the feature vector G;

[0025] (5)Use a fully connected layer to map the feature vector G to a scalar value, which is the sum of the correlation degrees of the three items of data where G = {g1, g2, …, g i , …, g n}}, g i represents the i-th feature vector, n is the length of the feature vector sequence, w i represents the weight of the i-th feature vector, and b represents the bias term.

[0026] In one embodiment, for the material data X1, the text is encoded by the Bert encoding module to extract E1, that is, E1 = Bert(text), and the process is as follows:

[0027] First, convert the input text into digital form, that is, use the WordPiece tokenizer to tokenize the input text, and map each word to the corresponding number in the vocabulary of the BERT model;

[0028] Then, according to the position encoding formula: positional_encoding[i, 2*j] = sin(i / 10000^(2*j / hidden_size)) and positional_encoding[i, 2*j + 1] = cos(i / 10000^(2*j / hidden_size)) to perform position encoding on the tokens of the text, where i represents the index of the position encoding, j represents the index of the hidden unit, and hidden_size is the number of hidden units in the BERT model, that is, hidden_size = 768;

[0029] Again, input the obtained encoded sequence into the BERT model for forward propagation, and convert it into a series of vectors through 12 layers of self-attention mechanism, fully connected, and pooling operations, where each vector represents a token in the input text;

[0030] Finally, generate the feature vector representation of the entire input text by taking the average of these vectors, with a size of (1, 768).

[0031] In one embodiment, for the process text data of the process data X2, the text is encoded by the Bert encoding module, that is, E2_1 = Bert(text), with a size of (1, 768);

[0032] For the process parameters of process data X2, E2_2 is extracted through the Linear encoding module. The process is as follows:

[0033] First, normalize the numerical range of the process parameters;

[0034] Then, weight the process parameters:

[0035] Secondly, use the principal component analysis method to transform the weighted process parameters into feature vectors;

[0036] Finally, through the fully connected layer, output a feature vector E2_2 with a fixed dimension, sized (1, 512);

[0037] Fuse the feature information by concatenating E2_1 and E2_2 along the second dimension, i.e., E2 = (E2_1, E2_2), sized (1, 1280).

[0038] In one embodiment, for the circuit structure diagram data X3, use the ResNet18 encoding module to perform feature extraction to obtain E3, i.e., E3 = ResNet18(image). The process is as follows:

[0039] First, the input RGB image of 3 * 224 * 224 is fed into the first convolutional layer, which contains 64 convolutional kernels, each with a size of 3×3, a stride of 2, and uses the ReLU activation function, outputting a feature map sized (batch_size, 64, 112, 112);

[0040] Then, the feature map passes through 4 convolutional blocks, each containing two convolutional layers and an identity mapping. Among them, the output channel number of the first convolutional layer of each convolutional block is 64, and the output channel number of the second convolutional layer is 128, 256, or 512. The convolutional kernel size in the convolutional layer is 3×3, the stride is 1, and the ReLU activation function and batch normalization are used;

[0041] Finally, after being processed by the fourth convolutional block, the feature map is fed into the global average pooling layer, which takes the average of all values on each channel of the feature map and outputs a feature vector E3 with a fixed dimension, sized (1, 1000).

[0042] In one embodiment, the dimensions of the three linear transformations W1, W2, and W3 are (256, 512), (256, 1280), and (256, 1000) respectively, and the sizes of the feature vector E and the feature vector G are both (1, 768).

[0043] In one embodiment, the chip radiation hardening technology optimization and matching module sequentially executes:

[0044] (1) Take the selected chip model, the provided irradiation dose, and the working environment as inputs, match the eligible data from the multi-modal database, and obtain a set of candidate hardening technologies, including hardening materials, hardening processes, and circuit structure diagrams.

[0045] (2) Randomly select 1 piece of data from each of the sets of hardening materials, hardening processes, and circuit structure diagrams to form 1 random initial solution, and evaluate the solution. If the evaluation result of the solution is valid, that is, the relevance of the solution is greater than the given threshold, then put it into the initial population. If the evaluation result of the solution is invalid, then randomly select another initial solution and evaluate it again. Repeat this process until there are k valid initial hardening technology solutions in the initial population.

[0046] (3) Use the genetic algorithm to optimize the initial hardening technology solutions in the initial population, continuously iterate to generate new populations. After Z rounds of genetic evolution operations, obtain the population generated in the last round, and select the top m hardening technology optimization solutions from it as the optimal hardening technology solutions for output, where m < k.

[0047] In one embodiment, the method for evaluating the solution is as follows:

[0048] Obtain the relevance S of the solution through the chip anti-irradiation multi-modal hardening technology relevance calculation module. If the relevance S is greater than or equal to the threshold H, then retain the solution. Otherwise, it means the solution is invalid and is discarded. The system will randomly combine and generate solutions multiple times and evaluate them, and retain the valid solutions together to form a population. This population includes k valid hardening technology solutions, where the threshold H is a preset system parameter.

[0049] In one embodiment, the process of the genetic algorithm is as follows:

[0050] ① Consist of k randomly generated valid initial hardening technology solutions to form an initial population.

[0051] ② Calculate the relevance of each solution in the current population respectively, and then use the roulette wheel strategy to select g solutions, where g < k.

[0052] ③ Pair the g selected solutions in pairs. Then, for each pair of solutions, randomly select a position with probability Pc for single-point crossover operation to generate a new pair of solutions. Calculate the relevance of each newly generated solution. If the relevance of the solution is greater than the given threshold H, then retain it. If the relevance of the solution is less than the given threshold, then discard it.

[0053] ④Perform a random mutation operation on each of the newly generated solutions with probability Pd; after the mutation operation, if a new solution is obtained, calculate the correlation degree of the new solution; if the correlation degree of this solution is greater than the given threshold H, retain it, otherwise discard this solution and perform another mutation operation;

[0054] ⑤In this way, a new generation of population is obtained;

[0055] ⑥Then repeat steps ② to ⑤ above on the new population, and end after repeating Z times.

[0056] The present invention collects data on chip anti-radiation multi-modal hardening technology, uses deep learning technology to discover the correlation relationship between different hardening technologies, and further combines different hardening technologies through methods such as genetic algorithms to form an optimal hardening solution. At the same time, it also provides a visual display, enabling R & D personnel to quickly and intuitively understand the advantages and disadvantages of different hardening solutions, thereby saving query time and improving work efficiency.

[0057] Therefore, compared with the prior art, the advantages of the present invention are as follows:

[0058] 1. Based on the matching and optimal selection of hardening technologies for multi-modal data, the optimal hardening solution can be quickly and accurately found.

[0059] 2. Combining deep learning and genetic algorithm technologies, it realizes a more intelligent matching and selection of chip anti-radiation multi-modal hardening technologies, improves the optimization of chip anti-radiation multi-modal hardening technology matching and selection, reduces the work intensity of chip R & D personnel, and improves work efficiency. Brief Description of the Drawings

[0060] Figure 1 It is a system framework diagram of the present invention.

[0061] Figure 2 It is a structural diagram of the chip anti-radiation multi-modal hardening technology correlation degree calculation module of the system of the present invention.

[0062] Figure 3 It is a Linear encoding module.

[0063] Figure 4 It is a structural diagram of the chip anti-radiation hardening optimization matching module of the system of the present invention.

[0064] Figure 5 It is the display effect of hardening technology 1 in solution 1 with the highest correlation degree in the embodiment of the present invention.

[0065] Figure 6 It is the display effect of hardening technology 2 in solution 1 with the highest correlation degree in the embodiment of the present invention.

[0066] Figure 7This is the display effect of the reinforcement technology 4 in Scheme 1 with the highest correlation degree in the embodiments of the present invention. Detailed implementation manners

[0067] The following will describe in detail the implementation manners of the present invention with reference to the accompanying drawings and embodiments.

[0068] The present invention provides an optimization matching system for chip anti-radiation reinforcement technology based on multi-modal data, which can generate all anti-radiation reinforcement technologies related to the queried chip and automatically realize the combination optimization of multiple reinforcement technologies. As Figure 1 shown, it mainly includes the following parts:

[0069] I. Chip anti-radiation reinforcement technology parameter acquisition module

[0070] This module is responsible for collecting and storing different types of chip anti-radiation reinforcement technologies and related data files, including information such as the corresponding reinforcement material data, reinforcement process data, and circuit structure diagram data, and constructing a multi-modal database.

[0071] II. Chip anti-radiation multi-modal reinforcement technology correlation degree calculation module

[0072] This module integrates and correlates the data obtained from the chip anti-radiation reinforcement technology parameter acquisition module to construct a chip anti-radiation reinforcement optimization scheme. In the present invention, the integrated data can be analyzed for the correlation between data through intelligent technologies such as deep learning.

[0073] III. Chip anti-radiation reinforcement technology optimization matching module

[0074] This module is responsible for automatically matching the eligible reinforcement technologies in the multi-modal database according to the chip model selected by the R & D personnel and the provided conditions such as irradiation dose and working environment, and conducting a combined evaluation of the matched reinforcement technologies, and then providing an optimization scheme. The R & D personnel can select the optimized reinforcement technologies according to the results, such as reinforcement materials and reinforcement processes.

[0075] IV. Visualization display module

[0076] This module is responsible for presenting the results of the chip anti-radiation reinforcement technology optimization matching to the R & D personnel. Specifically, it can be comprehensively presented in an intuitive and easy-to-use manner such as text, list, curve, image, chart, graph, and table, enabling scientific research personnel to quickly view and compare the performance indicators of different schemes and make selections and decisions on the final reinforcement scheme. The R & D personnel can perform operations such as screening and sorting on the results according to their own needs.

[0077] According to the above solution, the present invention performs matching and optimization based on multimodal data, which can improve the matching speed and accuracy of the optimal reinforcement solution. At the same time, combined with deep learning, the matching and selection become more intelligent.

[0078] In some specific embodiments of the present invention, the execution steps of the chip radiation-hardening technology parameter acquisition module can be further described as follows:

[0079] (1) Data acquisition can be carried out by means of manual upload, web crawling, and data extraction to obtain data parameters related to the radiation-hardening technology of different types of chips, including information on chip radiation-hardening materials, chip radiation-hardening circuit structure design, chip radiation-hardening processes, etc. Among them, manual upload is mainly for R & D personnel to experimentally test the relevant parameters and experimental results of chip radiation hardening. Web crawling is mainly to use the built-in crawling tool of the acquisition module to regularly collect data related to chip radiation-hardening technology from designated websites. Data extraction is mainly to extract chip-related data from the data systems of relevant manufacturers or units.

[0080] (2) Since there are differences in the same-modal data obtained through different methods, it is necessary to perform standardization and normalization processing on the data to unify data with different units and different orders of magnitude into a standard numerical form. Clean the null values, duplicate values, outliers, etc. existing in the data to ensure the integrity and accuracy of the data.

[0081] (3) Store in the multimodal database according to different data structures.

[0082] In some specific embodiments of the present invention, the execution steps of the chip radiation-hardening multimodal reinforcement technology correlation calculation module can refer to Figure 2 , and are further described as follows:

[0083] (1) For the data corresponding to a certain chip radiation-hardening technology, different deep learning methods are used to perform feature representation respectively, that is, extract feature vectors. Among them,

[0084] For the material data X1, the text content is encoded by the Bert encoding module to generate the corresponding feature vector, denoted as E1, i.e., E1 = Bert(text). In this process, first, the input text is converted into digital form, that is, the WordPiece tokenizer is used to tokenize the input text, and each word is mapped to the corresponding number in the vocabulary (vocab_size = 21128) of the BERT model. Then, according to the position encoding formula: positional_encoding[i,2*j] = sin(i / 10000^(2*j / hidden_size)) and positional_encoding[i,2*j+1] = cos(i / 10000^(2*j / hidden_size)), the tokens of the text are positionally encoded, where i represents the index of the position encoding, j represents the index of the hidden unit, and hidden_size is the number of hidden units in the BERT model, i.e., hidden_size = 768. Again, the obtained encoded sequence is input into the BERT model for forward propagation, and through 12 layers of self-attention mechanisms, fully connected and pooling operations, it is transformed into a series of vectors, where each vector represents a token in the input text. Finally, the feature vector representation of the entire input text is generated by taking the average of these vectors, with a size of (1, 768).

[0085] For the process data X2, it mainly consists of two parts: the text data of the radiation hardening process flow and the process parameters. Among them, for the process text data, the text data of the radiation hardening process flow is encoded by the Bert encoding module to generate the corresponding feature vector representation E2_1, with a size of (1, 768). The process is the same as that of the hardened material text information E1 = Bert(text). For the process parameters, the feature vectors are extracted through the Linear encoding module. As Figure 3 shown, in this process, first, the numerical range of the process parameters is normalized to facilitate the comparison and weighting of different process parameters. Then, the process parameters are weighted to better reflect the influence of different process parameters on the chip in the subsequent process. Secondly, the principal component analysis (PCA) method is used to transform the weighted process parameters into feature vectors. Finally, through the fully connected layer, a feature vector E2_2 with a fixed dimension is output, with a size of (1, 512). Therefore, the feature vector representation of the hardened process data X2 is the feature information fusion by concatenating the feature vectors E2_1 and E2_2 along the second dimension, i.e., E2 = (E2_1, E2_2), with a size of (1, 1280).

[0086] For the circuit structure diagram data X3, use the ResNet18 encoding module to extract features from it, generating the corresponding feature vector representation, denoted as E3, i.e., E3 = ResNet18(image). In this process, first, the input RGB image (3*224*224) is fed into the first convolutional layer, which contains 64 convolutional kernels, each with a size of 3×3, a stride of 2, and uses the ReLU activation function, outputting a feature map with a size of (batch_size, 64, 112, 112). Then, this feature map passes through 4 convolutional blocks, each containing two convolutional layers and an identity mapping (used to correspond the size of the input feature map and the number of channels of the output feature map). Among them, the output number of channels of the first convolutional layer of each convolutional block is 64, and the output number of channels of the second convolutional layer is 128, 256, or 512, depending on the layer where the convolutional block is located. The size of the convolutional kernel in the convolutional layer is 3×3, the stride is 1, and the ReLU activation function and batch normalization are used. Finally, after being processed by the fourth convolutional block, the feature map is fed into the global average pooling layer, which takes the average of all values on each channel of the feature map and outputs a feature vector with a fixed dimension, with a size of (1, 1000).

[0087] (2) Since the dimensions of the feature vectors E1, E2, and E3 are different, linear transformation methods need to be used separately to perform dimension conversion to map them into the same feature vector space. Let the three linear transformations be W1, W2, and W3, and their dimensions be (256, 512), (256, 1280), and (256, 1000) respectively. Then the transformed vector representations are: E'1 = W1 * E1 T , E'2 = W2 * E2 T , E'3 = W3 * E3 T .

[0088] (3) Perform feature information fusion on E'1, E'2, and E'3 in the same feature vector space by concatenating them in the second dimension, i.e., E = [E'1, E'2, E'3]. The size of the concatenated feature vector E is (1, 768).

[0089] (4) Use 2 Transformer layers to perform multiple non-linear transformations on the feature vector E to extract richer features, and then obtain the feature vector G with a size of (1, 768). In this network structure, the first Transformer layer can integrate three different types of feature vectors together and perform mutual fusion to obtain a better feature representation. The second Transformer layer can further enhance the feature representation ability to improve the accuracy of the correlation degree.

[0090] (5) Use a fully connected layer to map the feature vector G output by the Transformer layer to a scalar value, which is the sum of the correlation degrees of the three files. Among them, the feature vector G output by the Transformer layer = {g1, g2, …, g i , …, g n}, g i represents the i-th feature vector, n is the length of the feature vector sequence, w i represents the weight of the i-th feature vector, and b represents the bias term.

[0091] In some specific embodiments of the present invention, the execution steps of the chip radiation hardening technology optimization and matching module can refer to Figure 4 and are further described as follows:

[0092] (1) Take the chip model selected by the R & D personnel and the provided conditions such as radiation dose and working environment as inputs, match the eligible data from the multi-modal database to obtain a set of candidate hardening technologies, including hardening material text information, hardening processes, and circuit structure diagrams, etc.

[0093] (2) Randomly select 1 piece of data from each of the hardening material text information, hardening processes, and circuit structure diagram sets to form 1 random initial solution, form a random initial solution, and evaluate this solution. If the evaluation result of this solution is effective, that is, the correlation degree of this solution is greater than a given threshold, then put it into the initial population. If the evaluation result of this solution is ineffective, then randomly select another initial solution and evaluate it again. Repeat this process until there are k effective initial hardening technology solutions in the initial population.

[0094] Specifically: Obtain the correlation degree S of this solution through the chip radiation hardening multi-modal technology correlation degree calculation module. If the correlation degree S is greater than or equal to the threshold H, then retain this solution; otherwise, it means this solution is ineffective and is discarded. The system will randomly combine and generate solutions multiple times and evaluate them, and retain the effective solutions together to form a population. This population includes k effective hardening technology solutions, where the threshold H is a preset system parameter.

[0095] (3) Use the genetic algorithm to optimize the initial hardening technology solutions in the initial population, continuously iterate to generate a new population. After Z rounds of genetic evolution operations, obtain the population generated in the last round, and select the top m (m < k) optimized hardening technology solutions from it as the optimal hardening technology solutions for output.

[0096] The specific steps are as follows:

[0097] ① A set of k randomly generated effective initial hardening technology solutions constitutes an initial population.

[0098] ② Calculate the correlation degrees of each solution in the current population respectively, and then select g (g < k) solutions using the roulette wheel strategy.

[0099] ③ Pair up the g solutions selected above two by two. Then, for each pair of solutions, randomly select a position with probability Pc to perform a single-point crossover operation to generate a new pair of solutions. For example: Solution 2 (Material B + Process B + Structure A) and Solution 3 (Material A + Process C + Structure D) are crossed to become Solution 2 (Material B + Process C + Structure D) and Solution 3 (Material A + Process B + Structure A). Calculate the correlation degree of each newly generated solution. If the correlation degree of this solution is greater than the given threshold H, it is retained; if the correlation degree of this solution is less than the given threshold, it is discarded.

[0100] ④ Perform a random mutation operation on each of the newly generated solutions above with probability Pd. After the mutation operation, if a new solution is obtained, calculate the correlation degree of the new solution. For example: Mutate "Structure A" in Solution 3 to "Structure C", and finally form a new generation of solution population. If the correlation degree of this solution is greater than the given threshold H, it is retained; otherwise, this solution is discarded and the mutation operation is performed again.

[0101] ⑤ Thus, a new generation of population is obtained.

[0102] ⑥ Then repeat the above steps ② to ⑤ on the new population. After repeating Z times, end the process, and select the top m as the optimal reinforcement technology solutions for output.

[0103] The visualization display module is responsible for presenting the optimized and matched reinforcement solutions to the scientific research personnel, so that the scientific research personnel can more intuitively understand the performance and reliability of the optimized reinforcement solutions. By displaying the information and data of the reinforcement solutions, the scientific research personnel can select the optimal reinforcement solution, so as to achieve the purpose of improving the radiation resistance ability of the chip. This module will present each optimized and matched reinforcement solution in the form of text and lists. For the specific technologies involved in a certain reinforcement solution, the scientific research personnel can choose to view the details. In addition, the scientific research personnel can directly select a certain reinforcement solution as the final choice, or further adjust a certain solution according to their own needs.

[0104] In a specific embodiment of the present invention, the input chip parameters and radiation environment parameters are as follows: "Aerospace radiation-resistant Buck-type DC-DC, requiring input voltage: 5.5V - 17V, output adjustable: 1.2V - 12V, output current: 5A, maximum conversion efficiency greater than 90%. Radiation index requirements: total dose resistance greater than 100krad(Si), single-event LET threshold greater than 75MeV.cm2 / mg".

[0105] According to the above requirements, entity matching in the multi-modal database obtains the following reinforcement technology data:

[0106] (1) Reinforcement technology documentation: "Design and Implementation of Low-Voltage High-Current DC-DC Converters with Anti-Addition Characteristics.pdf", "Research on Key Technologies of Aerospace-Grade BUCK-Type DC-DC Switching Power Supplies.pdf", "Design of a Highly Reliable Radiation-Resistant DC / DC Converter.pdf", "Research and Application of a Single-Channel Key Analog Unit for Aerospace Radiation Resistance.pdf", "5nm Process Technology in DC / DC Conversion Chips.pdf".

[0107] (2) Reinforcement technology process flows: "A Process Flow for Adding a P+ Protection Ring at the Emitter Junction of NPN Bipolar Transistors.txt", "A Process Flow for Radiation Resistance Reinforcement by Adding a Protection Ring on MOSFETs.txt", "A Process Flow for Radiation Resistance Reinforcement by Changing P-Well Process Parameters on Silicon DMOS.txt", "A Process Flow for Radiation Resistance Reinforcement by Changing the Connection Method of the N-Well Bulk Electrode on Silicon CMOS.txt".

[0108] (3) Reinforcement process parameters: "Partial Design Parameters of DC / DC Converters.csv", "Test Results of Electrical Parameters of DC-DC Sample Circuits.csv", "Spraying Process Parameters.csv", "Table of Radiation Resistance Reinforcement Process Parameters for Aerospace Electronic Components.csv".

[0109] (4) Reinforcement circuit design diagrams: "Single-Event Latchup Reinforcement Method.jpg", "Schematic Diagram of Ramp Compensation Circuit.jpg", "Schematic Diagram of Error Amplifier Design.jpg", "Buck-Type DC-DC Topology.jpg".

[0110] Combine the above reinforcement technology data, and calculate the correlation degree through the chip radiation resistance multi-modal reinforcement technology correlation degree calculation module (as Figure 2 shown), so as to screen out the reinforcement technology combinations that meet the threshold requirements (assuming K = 3), and form 10 initial scheme combinations, as shown in Table 1.

[0111] Table 1 Initial Scheme Set

[0112]

[0113]

[0114] In Table 1, each column represents 1 initial scheme combination. Specifically, the 4 files in the parentheses of each column represent 4 different reinforcement technology data, forming 1 initial scheme combination. And the "correlation degree" in each column is obtained by calculating the correlation degree of this combination through the "chip radiation resistance multi-modal reinforcement technology correlation degree calculation module", and the subsequent value is the calculation result.

[0115] Take the initial solution combination as the initial population, and through the genetic optimization method, update and retain the optimal reinforcement technology solution combination. Set that after 10 iterations, the optimization of the genetic algorithm ends, and save the final reinforcement technology solution combination.

[0116] Select the first 5 reinforcement technology solution combinations as the optimal reinforcement technology solutions and output them in the visualization display module. This output is multi-modal data, consisting of Table 2, Figure 5 , Figure 6 , Table 3 and Figure 7 . It should be noted that this multi-modal data is presented completely in the visualization display module of the present invention. Considering the expression method, the present invention shows the tables and pictures separately here, and shows the pdf file and text file in the form of pictures.

[0117] Table 2 Optimal solution combinations finally output

[0118]

[0119] Table 3 Reinforcement technology 3 in Solution 1

[0120]

[0121]

[0122] In Table 2, the 5 optimal solution combinations finally output are shown in descending order. Each solution includes "Reinforcement technology 1, Reinforcement technology 2, Reinforcement technology 3, Reinforcement technology 4".

[0123] The 4 reinforcement technologies in Solution 1 with the highest correlation are shown in the ways of PDF, text, table and picture respectively in practical applications.

[0124] In Figure 5 , the reinforcement technology 1 in Solution 1 with the highest correlation is shown, namely "Research and Application of a Single Channel of the Key Anti-Radiation Simulation Unit for Aerospace.pdf". The PDF is an embedded corresponding browsing plug-in. Since the content is too much to be fully displayed, the plug-in scroll bar is used to assist browsing.

[0125] In Figure 6 , the reinforcement technology 2 in Solution 1 with the highest correlation is shown, namely "Process Flow of Anti-Irradiation Reinforcement by Changing P-Well Process Parameters on Silicon DMOS.txt". The text is an embedded corresponding browsing plug-in. Since the content is too much to be fully displayed, the plug-in scroll bar is used to assist browsing.

[0126] In Table 3, the reinforcement technology 3 in Solution 1 with the highest correlation is shown, namely "Test Results of Electrical Parameters of DC-DC Sample Circuit.csv".

[0127] In Figure 7 it shows the reinforcement technology 4 in the most relevant Solution 1, namely "well contact protection ring structure.jpg", and the meaning of this figure is: in the unit layout, a well contact protection ring is designed, and more contact holes are designed on the protection ring.

[0128] Thus, the present invention realizes the multi-modal presentation of the results.

Claims

1. A chip anti-irradiation hardening technology optimization and matching system based on multi-modal data, characterized in that, Including: A chip radiation-hardening technology parameter acquisition module that collects and stores data on different types of chip radiation-hardening technologies and their corresponding hardening materials, hardening process data, and circuit structure diagram data, and constructs a multi-modal database; A chip radiation-hardening multi-modal technology correlation calculation module that calculates the technical correlation of the chip radiation-hardening optimization scheme obtained by integrating and correlating the collected data to determine whether the scheme is effective; A chip radiation-hardening technology optimization and matching module that automatically matches the eligible hardening technologies in the multi-modal database according to the selected chip model, provided radiation dose, and working environment, forms multiple hardening technology schemes, and uses the chip radiation-hardening multi-modal technology correlation calculation module to evaluate all the schemes, and then performs combined optimization and multiple iterations to generate a better chip radiation-hardening scheme; A visualization display module that displays the results of the chip radiation-hardening technology optimization and matching; Among them, the chip radiation-hardening multi-modal technology correlation calculation module sequentially executes: (1) For the data corresponding to a certain chip radiation-hardening technology, different deep learning methods are used to perform feature representation respectively, that is, feature vectors are extracted, where: For the material data X1, the feature vector E1 of its text information is extracted; For the process data X2, the feature vector E2_1 of its process text materials and the feature vector E2_2 of the process parameters are extracted, and the fused feature vector E2 is obtained; For the circuit structure diagram data X3, its feature vector E3 is extracted; (2) Perform dimensionality conversion on E1, E2, and E3 using a linear transformation method, and map them into the same feature vector space. The transformed vectors are expressed as: E'1 = W1 * E1 T , E'2 = W2 * E2 T , E'3 = W3 * E3 T , where W1, W2, and W3 are three linear transformations respectively; (3) The feature information of E'1, E'2, and E'3 in the same feature vector space is fused to obtain the spliced feature vector E; (4) Two Transformer layers are used to perform multiple non-linear transformations on the feature vector E to extract richer features, and then the feature vector G is obtained; (5) Use a fully connected layer to map the feature vector G to a scalar value, which is the sum of the association degrees of the three items of data where G = {g1, g2, …, g i , …, g n}, g i represents the i-th feature vector, n is the length of the feature vector sequence, w i represents the weight of the i-th feature vector, and b represents the bias term.

2. The chip anti-radiation hardening technology optimization matching system based on multimodal data according to claim 1, characterized in that In the chip radiation-hardening technology parameter acquisition module, the means of data acquisition include: R & D personnel experimentally test the relevant parameters and experimental results of chip radiation hardening and manually upload them; Use the embedded crawling tool to regularly crawl data from the specified website; Extract data from the manufacturer's data system; The same-modal data collected by different means are standardized and normalized, and stored in the multi-modal database according to different data structures.

3. The chip anti-irradiation reinforcement technology optimization and matching system based on multi-modal data according to claim 1, characterized in that, For the material data X1, the text is encoded by the Bert encoding module to extract E1, that is, E1 = Bert(text), and the process is as follows: First, the input text is converted into digital form, that is, the input text is tokenized using the WordPiece tokenizer, and each word is mapped to the corresponding number in the vocabulary of the BERT model; Then, according to the positional encoding formula: positional_encoding[i, 2*j] = sin(i / 10000 ^ (2*j / hidden_size)) and positional_encoding[i, 2*j + 1] = cos(i / 10000 ^ (2*j / hidden_size)), perform positional encoding on the word segments of the text, where i represents the index of the positional encoding, j represents the index of the hidden unit, and hidden_size is the number of hidden units in the BERT model, that is, hidden_size = 768; Next, input the obtained encoded sequence into the BERT model for forward propagation, and convert it into a series of vectors through 12 layers of self-attention mechanism, fully connected layer, and pooling operations, where each vector represents a token in the input text; Finally, generate the feature vector representation of the entire input text by taking the average of these vectors, with a size of (1, 768).

4. The chip anti-radiation hardening technology optimization matching system based on multimodal data according to claim 3, wherein For the process text data of process data X2, encode the text through the Bert encoding module, that is, E2_1 = Bert(text), with a size of (1, 768); For the process parameters of process data X2, extract E2_2 through the Linear encoding module, and the process is as follows: First, normalize the numerical range of the process parameters; Then, weight the process parameters: Secondly, use the principal component analysis method to transform the weighted process parameters into feature vectors; Finally, through the fully connected layer, output a feature vector E2_2 with a fixed dimension, with a size of (1, 512); Concatenate E2_1 and E2_2 along the second dimension for feature information fusion, that is, E2 = (E2_1, E2_2), with a size of (1, 1280).

5. The chip anti-radiation hardening technology optimization matching system based on multi-modal data according to claim 4, characterized in that, For the circuit structure diagram data X3, use the ResNet18 encoding module to extract features to obtain E3, that is, E3 = ResNet18(image), and the process is as follows: First, the input RGB image of 3 * 224 * 224 is fed into the first convolutional layer, which contains 64 convolutional kernels, each with a size of 3×3, a stride of 2, and uses the ReLU activation function, and outputs a feature map with a size of (batch_size, 64, 112, 112); Then, this feature map passes through 4 convolutional blocks, each convolutional block contains two convolutional layers and an identity mapping. Among them, the output channel number of the first convolutional layer of each convolutional block is 64, and the output channel number of the second convolutional layer is 128, 256, or 512. The size of the convolutional kernel in the convolutional layer is 3×3, the stride is 1, and the ReLU activation function and batch normalization are used; Finally, after being processed by the fourth convolutional block, the feature map is fed into the global average pooling layer, which takes the average of all values on each channel of the feature map and outputs a feature vector E3 with a fixed dimension, with a size of (1, 1000).

6. The chip anti-radiation hardening technology optimization matching system based on multimodal data according to claim 5, characterized in that The dimensions of the three linear transformations W1, W2, and W3 are (256, 512), (256, 1280), and (256, 1000) respectively, and the sizes of the eigenvector E and the eigenvector G are both (1, 768).

7. The chip anti-radiation hardening technology optimization matching system based on multimodal data according to claim 1, wherein The chip radiation hardening technology optimization and matching module sequentially executes the following steps: (1) Using the selected chip model, the provided radiation dose, and the working environment as inputs, match the eligible data from the multi-modal database to obtain a set of candidate hardening technologies, including hardening materials, hardening processes, and circuit structure diagrams; (2) Randomly select 1 piece of data from each of the hardening material, hardening process, and circuit structure diagram sets to form 1 random initial plan, and evaluate the plan; if the evaluation result of the plan is effective, that is, the correlation degree of the plan is greater than the given threshold, then put it into the initial population; if the evaluation result of the plan is invalid, then randomly select another initial plan and evaluate it again; Repeat the above steps until there are k effective initial hardening technology plans in the initial population; (3) Use the genetic algorithm to optimize the initial hardening technology plans in the initial population, continuously iterate to generate a new population. After Z rounds of genetic evolution operations, obtain the population generated in the last round, and select the top m hardening technology optimization plans from it as the optimal hardening technology plan for output, where m < k.

8. The chip anti-radiation hardening technology optimization matching system based on multi-modal data according to claim 7, characterized in that, The method for evaluating the plan is as follows: Obtain the correlation degree S of the plan through the chip radiation multi-modal hardening technology correlation degree calculation module. If the correlation degree S is greater than or equal to the threshold H, then retain the plan; otherwise, it means the plan is invalid and is discarded; the system will randomly combine and generate plans multiple times and evaluate them, and retain the effective plans together to form a population. This population includes k effective hardening technology plans, where the threshold H is a preset system parameter.

9. The chip anti-radiation hardening technology optimization matching system based on multimodal data according to claim 7, characterized in that, The process of the genetic algorithm is as follows: ① An initial population is composed of k randomly generated effective initial hardening technology plans; ② Calculate the correlation degree of each plan in the current population respectively, and then select g plans using the roulette wheel strategy, where g < k; ③ Pair the g selected plans in pairs; then each pair of plans randomly selects a position for single-point crossover operation with probability Pc to generate a new pair of plans; calculate the correlation degree of each newly generated plan. If the correlation degree of the plan is greater than the given threshold H, then retain it; if the correlation degree of the plan is less than the given threshold, then discard it; ④ Perform random mutation operations on each of the newly generated plans with probability Pd; after the mutation operation, if a new plan is obtained, calculate the correlation degree of the new plan; if the correlation degree of the plan is greater than the given threshold H, then retain it; otherwise, discard the plan and perform the mutation operation again; ⑤ Thus, a new generation of population is obtained; ⑥ Then repeat the above steps ② to ⑤ on the new population, and end after repeating Z times.

Images