A structural material high-throughput design method based on microstructure image visual analysis

Abstract

The application discloses a structure material high-throughput design method based on microstructure image visual analysis, and comprises the following steps: establishing a data set according to the composition, process, performance and corresponding microstructure image of a material; extracting an organization characteristic parameter by using machine learning on the microstructure image in the data set to obtain a normalized input data set; establishing a machine learning model of material composition / organization characteristic parameter→performance according to the normalized input data set, and training the machine learning model to make the machine learning model reach a preset precision, and completing the machine learning model training; taking the machine learning model as a discrimination model, and establishing an optimization model of performance→composition / organization parameter by using a neural network optimized by a genetic algorithm; inputting a target performance into the optimization model for training to obtain a predicted composition and a predicted organization characteristic parameter, selecting a corresponding process in the data set according to the predicted organization characteristic parameter, and obtaining a final design composition and process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material design, and particularly relates to a structural material high-throughput design method based on microstructure image visual analysis. BACKGROUND

[0002] Materials serve various aspects of national economy, social development, national defense construction and people's life, and are the material guarantee for economic construction, social progress and national security. With the development of science and technology, higher demands are put forward for key metal materials and composite materials widely used in fields such as aerospace, energy equipment, transportation and electronic information, especially for materials in extremely complex service environments, which are required to have more excellent comprehensive performance. Therefore, the material design method for comprehensive performance has attracted widespread attention of researchers.

[0003] At present, material design mainly relies on the traditional experimental trial-and-error method, which is difficult to realize efficient and synchronous optimization of comprehensive performance, and even more difficult to realize precise design of material composition and process, resulting in long research and development cycle, low efficiency and high cost of new materials.

[0004] The proposal of material genetic engineering promotes the development of material big data, and brings material research and development into the fourth paradigm of data-driven. Using machine learning to establish the mapping relationship between material influencing factors (such as composition and process) and target quantities (such as performance, microstructure and phase composition) can realize the prediction of material composition, structure, organization, process and performance and the discovery of new materials. For example, Chinese patent CN 110010210 A provides a multi-component alloy composition design method based on machine learning and oriented to performance requirements, which mines a large amount of existing data about alloy composition and performance, uses machine learning technology to unlock the implicit complex relationship between “composition-performance”, and realizes the purpose of quickly and accurately designing alloy composition according to performance requirements; Chinese patent CN 112582032 A is a method for assisting in designing high-thermal-stability iron-based soft-magnetic amorphous alloy by using an interpretable XGBoost machine learning algorithm, which establishes a prediction criterion with high accuracy, provides a new method for research and development of new iron-based soft-magnetic amorphous alloy, and significantly improves the research and development efficiency and reduces the research and development cost. However, the above-mentioned patents are all suitable for material composition optimization, cannot design the preparation process of the material, ignore the influence of microstructure on material performance, and are difficult to realize the synchronous optimization of multiple performances. SUMMARY

[0005] The present application aims to overcome the deficiencies of the prior art, and provides a structural material high-throughput design method based on microstructure image visual analysis, which comprises the following steps:

[0006] S1, establishing a data set according to the composition, process, performance and corresponding microstructure image of the material;

[0007] S2, extracting the microstructure feature parameters of the microstructure images in the data set by machine learning, establishing an input data set according to the material composition and the microstructure feature parameters, and performing data dimension reduction and normalization on the input data set to obtain a normalized input data set;

[0008] S3, establishing a machine learning model of material composition / microstructure feature parameters→performance according to the normalized input data set, and training the machine learning model to make the machine learning model reach a preset accuracy, thereby completing the training of the machine learning model;

[0009] S4, taking the machine learning model as a discriminant model, establishing an optimization model of performance→composition / microstructure parameters by using a neural network optimized by a genetic algorithm; inputting a target performance into the optimization model for training to obtain predicted composition and predicted microstructure feature parameters, and selecting a process corresponding to the predicted microstructure feature parameters in the data set to obtain a final design composition and process.

[0010] Further, the microstructure images in the data set are extracted by machine learning to obtain microstructure feature parameters, an input data set is established according to the material composition and the microstructure feature parameters, and data dimension reduction and normalization are performed on the input data set to obtain a normalized input data set, including the following processes:

[0011] S21, extracting the microstructure feature parameters of the microstructure images of the same category and the same size by using a machine learning algorithm;

[0012] S22, establishing an input data set according to the material composition and the corresponding microstructure feature parameters, performing dimension reduction on the input data set, and calculating the correlation of each microstructure feature parameter after dimension reduction;

[0013] S23, normalizing the data set after dimension reduction to obtain a normalized input data set.

[0014] Further, the machine learning model of material composition / microstructure feature parameters→performance is established according to the normalized input data set, and the machine learning model is trained to make the machine learning model reach a preset accuracy, thereby completing the training of the machine learning model, including:

[0015] S31, taking the material composition and the corresponding microstructure feature parameter data in the normalized input data set as input and the performance data as output to establish a material composition / microstructure feature parameter→performance model;

[0016] S32, selecting hyperparameters to train the composition / microstructure feature parameter→performance model to make the accuracy of the model reach a preset accuracy.

[0017] Further, the machine learning model is used as a discriminant model, a neural network optimized by a genetic algorithm is used to establish a performance composition / organizational characteristic parameter optimization model; the target performance is input into the optimization model for training to obtain predicted composition and predicted organizational characteristic parameters, and the corresponding process in the data set is selected according to the predicted organizational characteristic parameters to obtain the final design composition and process, including:

[0018] S41, the randomly generated original solution of the set number of weights and threshold values is introduced into an artificial neural network with multi-dimensional target attributes as input and features as output, and the artificial neural network generates pseudo features according to the required attributes;

[0019] S42, the generated pseudo features are used as input, and a discriminator is used to obtain predicted performance;

[0020] S43, the relative error between the target performance and the predicted performance is calculated, and the fitness is obtained according to the reciprocal of the sum of the relative errors:

[0021]

[0022] S44, it is judged whether the fitness is within a preset range, if not, the weights and threshold values corresponding to the n predicted performances with the fitness greater than the set fitness threshold are selected as the female parent, and a new set number of weights and threshold values are generated by using a genetic algorithm through crossover and mutation;

[0023] S45, S41-S44 are repeated until the fitness is within the preset range, the cycle is stopped, and the feature value corresponding to the maximum fitness is selected as the final predicted composition and the final predicted organizational characteristic parameter;

[0024] S46, the process with smaller absolute value of the organizational characteristic parameter in the data set and the final predicted organizational parameter is selected as the final preparation process.

[0025] The beneficial effects of the present application are: the material organization can be quantified, the quantitative relationship between material composition and organization and performance can be established, and new structural materials can be quickly and accurately designed according to given performance requirements, and the material research and development efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The flowchart of the structural material high-throughput design method based on microstructure image visual analysis of the present application;

[0027] Figure 2 The flowchart of the discriminator and the genetic algorithm optimized neural network composition / organizational optimization model described in S4 of the present application;

[0028] Figure 3 The result schematic diagram of the SURF feature point extraction in the present application;

[0029] Figure 4 is a schematic diagram of the correlation results of the component and the tissue characteristic parameter in the step S22 of the present application. DETAILED DESCRIPTION

[0030] The technical solutions of the present application are described in further detail below in combination with the drawings, but the protection scope of the present application is not limited to the following description.

[0031] For the purpose of the present application, the technical solutions and advantages are more clearly and obviously understood, the present application is further described in detail in combination with the drawings and the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application, that is, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0032] Therefore, the detailed description of the embodiments of the present application provided in the drawings below is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. It should be noted that the relationship terms such as "first" and "second" and the like are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations.

[0033] Moreover, the term "comprising", "containing" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the sentence "including a…" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0034] The features and performances of the present application are further described in detail below in combination with the embodiments.

[0035] As shown in the following figure, the structure material high-throughput design method based on microstructure image visual analysis includes the following steps: Figure 1 S1, according to the composition, process, performance and corresponding microstructure image of the material to establish a data set;

[0036]

[0037] ​S2, for the microstructure images in the data set, using machine learning to extract the organization characteristic parameters, establishing the input data set according to the material composition and the organization characteristic parameters, and performing data dimension reduction and normalization on the input data set to obtain the normalized input data set;

[0038] S3, a machine learning model of material composition / organization characteristic parameters→performance is established according to the normalized input data set, and the machine learning model is trained to make the machine learning model reach a preset accuracy, and the machine learning model training is completed;

[0039] S4, using the machine learning model as a discriminant model, a neural network optimized by a genetic algorithm is used to establish an optimization model of performance→composition / organization parameters; the target performance is input into the optimization model for training to obtain predicted composition and predicted organization characteristic parameters, and a process corresponding to the data set is selected according to the predicted organization characteristic parameters to obtain the final design composition and process.

[0040] The microstructure images in the data set are extracted by using machine learning to extract the organization characteristic parameters, the input data set is established according to the material composition and the organization characteristic parameters, and the normalized input data set is obtained by performing data dimension reduction and normalization on the input data set, including the following processes:

[0041] S21, using a machine learning algorithm to extract the organization characteristic parameters of the microstructure images of the same category and the same size;

[0042] S22, establishing an input data set according to the material composition and the corresponding organization characteristic parameters, performing dimension reduction on the input data set, and calculating the correlation of each organization characteristic parameter after dimension reduction;

[0043] S23, normalizing the data set after dimension reduction to obtain the normalized input data set.

[0044] The machine learning model of material composition / organization characteristic parameters→performance is established according to the normalized input data set, and the machine learning model is trained to make the machine learning model reach a preset accuracy, and the machine learning model training is completed, including:

[0045] S31, the material composition and the corresponding organization characteristic parameter data in the normalized input data set are taken as input, and the performance data is taken as output to establish a material composition / organization characteristic parameter→performance model;

[0046] S32, selecting hyperparameters to train the composition / organization characteristic parameter→performance model to make the accuracy of the model reach a preset accuracy.

[0047] The machine learning model is used as a discriminant model, and a neural network optimized by a genetic algorithm is used to establish a performance→component / organizational characteristic parameter optimization model; the target performance is input into the optimization model for training to obtain predicted components and predicted organizational characteristic parameters, and the corresponding process in the data set is selected according to the predicted organizational characteristic parameters to obtain the final designed components and processes, including:

[0048] S41, the randomly generated original solution of the set number of weights and threshold values is introduced into the artificial neural network with multi-dimensional target attributes as input and features as output, and the artificial neural network generates pseudo features according to the required attributes;

[0049] S42, the generated pseudo features are used as input, and the discriminator is used to obtain predicted performance;

[0050] S43, the relative error between the target performance and the predicted performance is calculated, and the fitness is obtained according to the reciprocal of the sum of the relative errors:

[0051]

[0052] S44, it is judged whether the fitness is within the preset range, if not, the n predicted performances corresponding to the weights and threshold values with fitness greater than the set fitness threshold are selected as the female parent, and the genetic algorithm is used to generate new set number of weights and threshold values through crossover and mutation;

[0053] S45, repeating S41-S44 until the fitness is within the preset range, stopping the cycle, and selecting the maximum fitness corresponding feature value as the final predicted component and the final predicted organizational characteristic parameter;

[0054] S46, the process with the organizational characteristic parameter close to the final predicted organizational characteristic in the data set is used as the final preparation process.

[0055] Specifically, a structure material high-throughput design method based on microstructure image visual analysis, which quantifies the microstructure of the structure material, establishes a data set combined with the material component data, uses a machine learning algorithm to establish a quantitative relationship between “component / organization-performance”, and uses a neural network optimized by a genetic algorithm to realize directional design from performance to component / organization, overcoming the complex optimization problem from low dimension to high dimension, realizing fast and accurate design of alloy components and processes according to performance, and accelerating the research and application of new materials. The specific steps include:

[0056] (1) Data collection: through collecting and arranging the data of published literature, the component, organization and performance data of the structure material are obtained, and the data set is established.

[0057] Structural materials include but are not limited to the following: metallic materials (aluminum alloy, copper alloy, magnesium alloy, titanium alloy, high-temperature alloy, high-entropy alloy, etc.); composite materials (metal matrix composite, ceramic matrix composite)

[0058] The performance data of the material include but are not limited to: hardness, fracture toughness, bending strength, yield strength, and a combination of one or more of the above properties.

[0059] The microstructure photos can be light microscope photos, secondary electron photos or backscattered photos under the same magnification;

[0060] (2) Data set establishment: using machine learning algorithms such as SURF or n-PC to extract feature quantities from microstructure photos of the same category and the same size. Integrate material composition and organizational features to establish input data set, use PCA to reduce dimension of the input data set, and calculate the correlation of each feature parameter after dimension reduction, and normalize the data after dimension reduction (0-1).

[0061] (3) Establishing a machine learning model of material composition / organizational parameters→performance: taking the normalized composition and organizational parameters in step (2) as input and performance data as output, selecting a machine learning model and adjusting its hyperparameters, training different performance corresponding models, and establishing a model for predicting performance from composition and organization, denoted as performance prediction model.

[0062] 10-fold cross-validation is used to train the performance prediction model, and the hyperparameters are selected to make the performance prediction model have a prediction accuracy of not less than 0.85 and an error of not more than 10% for each performance.

[0063] (4) Using the performance prediction model established in (3) as the discriminator, a neural network optimized by genetic algorithm is used to establish a performance→composition / organizational parameter optimization model; input the target performance into the optimization model for training to obtain the predicted composition and organizational parameters, and select the corresponding process in the data set according to the organizational parameters to obtain the final design composition and process, as shown in Figure 2 .

[0064] First, randomly generate 100 sets of weight values and threshold values (G) and assign them to an artificial neural network framework containing three hidden layers with 15 neurons in each layer, forming an artificial neural network (ANN) with multi-dimensional target performance as input and composition and organizational features as output. ANN generates a first pseudo-feature (O) according to the required attribute (T), and uses the discriminator to calculate the predicted performance (P) corresponding to the pseudo-feature (O);

[0065] The sum of the relative errors between the target performance (T) and the predicted performance (P) is calculated as the fitness;

[0066] If not, 50 weight values and threshold values (G) with greater performance fitness are selected as parents, and new 100 weight values and threshold values (G') are generated by using genetic algorithm (GA) through crossover and mutation.

[0067] The pseudo-feature and performance prediction process are repeated until the fitness is within the preset range, and the cycle is stopped. The feature value corresponding to the maximum fitness in the final result is selected as the final prediction component and the tissue feature parameter.

[0068] The preparation process of the sample in the data set with tissue characteristics close to the final predicted tissue characteristics is selected as the final preparation process.

[0069] Embodiment:

[0070] The method provided by the application is applied to design the composition and process of powder metallurgy preparation of WC-Co hard alloy, with hardness Hv30 = 1700 kg / mm 2 , fracture toughness K 1C = 11 MPa·m 1 / 2 , and bending strength TRS = 4640 MPa as performance targets. The design method is specifically described as follows:

[0071] (1) Data collection: By collecting and sorting the published compositions of WC-Co-M (M = VC, Cr3C2, TaC, NbC, etc. grain growth inhibitors), raw material states, preparation methods, hardness, fracture toughness, bending strength, and corresponding backscattering tissue photographs, a basic data set is established.

[0072] The backscattering photographs corresponding to the alloy are collected, and four groups of regions with a size of 6um*6um and a resolution of 256pixel*256pixel are selected as initial tissue photographs, as shown in Figure 3 (a).

[0073] (2) Feature extraction is performed on the initial tissue photographs of the WC-Co alloy by using SURF technology. Each feature point extracted by SURF is represented by a set of 64-dimensional vectors, and the feature extraction result is shown in Figure 3 (b). The first 20 feature points are selected, so each group of photographs is used as an initial tissue feature vector with 64*20 = 1280 dimensions. The initial tissue feature vector is processed by using a PCA algorithm for dimension reduction, and the first 20 features after dimension reduction are selected as the final tissue features of the alloy (the tissue features are 21-dimensional in total).

[0074] The composition features (Co content, VC, Cr3C2, TaC, and NbC content, a total of 5 dimensions) and the final tissue features are integrated, and the data sets corresponding to the hardness, fracture toughness, and bending strength are established, respectively.

[0075] The correlation of each feature parameter in different data sets was calculated using Pearson correlation coefficient and maximum mutual information coefficient, respectively, as shown in Figure 4 After confirming that there were no closely related features, each data set was normalized to 0-1.

[0076] (3) Machine learning model of WC-Co alloy composition / structure parameter→hardness (or fracture toughness, bending strength) was established: the normalized composition and structure parameters in step (2) were taken as input, and the hardness (or fracture toughness, bending strength) data were taken as output. Ten different machine learning models were selected, 10-fold cross-validation was adopted to train the performance prediction model, and the hyperparameters were adjusted so that the prediction accuracy of the performance prediction model was not less than 0.85. For hardness, fracture toughness and bending strength, GradientBoostinRegressor was selected as the final algorithm to establish the performance prediction model, and the performance error was not more than 10%.

[0077] (4) The three performance prediction models established in (3) were used as discriminators, and a neural network optimized by genetic algorithm was used to establish a composition and structure feature optimization model; Hv30=1700 kg / mm 2 , fracture toughness K 1C =11 MPa·m 1 / 2 , and bending strength TRS=4640 MPa were taken as target performance and input into the optimization model for training. First, 100 randomly generated original solutions of weight and threshold value (G) were introduced into a structured artificial neural network, which took multi-dimensional target attributes as input and features as output. ANN generated the first pseudo-feature (O) according to the required attribute (T), and used the discriminator to calculate the predicted performance (P) corresponding to the pseudo-feature (O).

[0078] The reciprocal of the sum of the relative errors between the target performance (T) and the predicted performance (P) was taken as the fitness.

[0079] It was judged whether the fitness was within the preset range. If not, the 50 predicted performance pairs corresponding to the weight and threshold value (G) with larger fitness were selected as the parents, and new 100 weight and threshold values (G') were generated by crossover and mutation using genetic algorithm (GA).

[0080] The pseudo-feature and performance prediction process was repeated until the fitness was within the preset range, and the cycle was stopped. The feature value corresponding to the maximum fitness in the final result was selected as the final predicted composition and structure feature parameter.

[0081] The preparation process of the sample with the closest structure feature to the final predicted structure feature in the data set was selected as the final preparation process.

[0082] After returning to normalization after multiple optimization, the results are shown in Table 1. The prediction results are simplified, and finally WC-10Co-0.75Cr3C2 is selected as the final prediction result, the number of feature points extracted by SURF is 300-400, combined with the structure of the data set, the initial particle size of WC is selected as 0.6 μm and 0.8 μm.

[0083] The material is prepared by using low-pressure sintering technology, and the hardness, fracture toughness and bending strength are tested, and the results are shown in Table 2. The error between the performance of the prepared alloy and the target performance is within 10%, which shows that the structure material high-throughput design method based on microstructure image visual analysis provided by the present application has high precision.

[0084] Table 1 Discriminator and genetic algorithm optimized neural network for component and organization feature optimization results

[0085]

[0086] Table 2 Comparison results of WC-Co cemented carbide performance designed according to the present application and target performance

[0087]

[0088] The above only describes the preferred embodiments of the present application, and it should be understood that the present application is not limited to the forms disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concepts described herein by the above-mentioned teaching or related art or knowledge. The modifications and changes made by those skilled in the art without departing from the spirit and scope of the present application shall be within the scope of protection of the appended claims of the present application.

Claims

1. A method for high-throughput design of structural materials based on visual analysis of microstructure images, characterized in that, It comprises the following steps: S1, according to the composition of the material, process, performance and corresponding microstructure image to establish a data set; S2, the microstructure image in the data set, using machine learning to extract the organization characteristic parameter, according to the material composition and organization characteristic parameter to establish input data set, and the input data set is reduced dimension and normalized, obtains the normalized input data set; S3, according to the normalized input data set, the machine learning model of material composition / organization characteristic parameter→performance is established, and the machine learning model is trained, so that the machine learning model reaches the preset accuracy, and the machine learning model training is completed; S4, using the machine learning model as the discriminant model, the performance→component / organization parameter optimization model is established by using the neural network optimized by genetic algorithm; the target performance is input into the optimization model for training to obtain the predicted composition and predicted organization characteristic parameter, and the corresponding process in the data set is selected according to the predicted organization characteristic parameter to obtain the final design composition and process; The step S4 comprises the following steps: S41, the randomly generated original solution of a certain number of weights and thresholds is introduced into the artificial neural network with multidimensional target attributes as input and features as output, and the artificial neural network generates pseudo features according to the required attributes; S42, the generated pseudo features are used as input, and the predicted performance is obtained by using the discriminator; S43, the relative error between the target performance and the predicted performance is calculated, and the fitness is obtained according to the reciprocal of the sum of the relative errors: S44, whether the fitness is in the preset range is judged, if not, the n predicted performances corresponding to the weights and thresholds with the fitness greater than the set fitness threshold are selected as the female parent, and the new weights and thresholds of a certain number are generated by using genetic algorithm through crossover and mutation; S45, S41-S44 are repeated until the fitness is in the preset range, the cycle is stopped, and the feature value corresponding to the maximum fitness is selected as the final predicted composition and the final predicted organization characteristic parameter; S46, the process with the minimum absolute value of the organization characteristic parameter in the data set and the final predicted organization parameter is selected as the final preparation process.

2. The microstructure image-based visual analysis of structural materials high-throughput design method according to claim 1, characterized in that, The microstructure image in the data set is extracted by using machine learning, and the input data set is established according to the material composition and the organization characteristic parameter, and the normalized input data set is obtained by reducing dimension and normalizing the input data set, which comprises the following processes: S21, the same category and size of microstructure image is extracted by using machine learning algorithm; S22, the input data set is established according to the material composition and the corresponding organization characteristic parameter, the input data set is reduced dimension, and the correlation of each organization characteristic parameter after dimension reduction is calculated; S23, the normalized input data set is obtained by normalizing the data set after dimension reduction.

3. The microstructure image-based visual analysis for structural material high-throughput design method according to claim 1, wherein, The machine learning model of material composition / organization characteristic parameter→performance is established according to the normalized input data set, and the machine learning model is trained, so that the machine learning model reaches the preset accuracy, and the machine learning model training is completed, which comprises: S31, taking the material composition in the normalized input data set and the corresponding tissue characteristic parameter data as input, and performance data as output, to establish a composition / tissue characteristic parameter→performance model of the material; S32, selecting hyperparameters to train the composition / tissue characteristic parameter→performance model, so that the accuracy of the model reaches a preset accuracy.

Images