Reverse prediction method for small sample material

Abstract

The invention relates to a reverse prediction method for a small sample material. The method comprises the following steps: acquiring data characteristics of a composite material; preprocessing the data features; building a model group formed by connecting a plurality of multi-input single-output generalized regression neural networks in parallel, and training the model group to characterize an implicit relationship among material components, processes and performances; establishing a reverse prediction model combined by a model group and a genetic algorithm; and comparing and screening out available prediction data and merging the prediction data into a total data set, and by adopting a small sample material reverse prediction method, the experiment trial and error cost can be greatly reduced, and the material design efficiency can be improved.

Description

Technical Field

[0001] The invention belongs to the technical field of material design optimization, and in particular relates to a small sample material reverse prediction method. Background Art

[0002] At present, the design and development of materials still mainly rely on the traditional "trial and error" experimental method to prepare and test samples. This method has a long cycle, low efficiency, high cost and a certain degree of randomness, which is not suitable for the synthesis and development of new materials under the current situation.

[0003] As one of the fastest-growing branches in the field of artificial intelligence, machine learning has received extensive attention from scholars in recent years. Machine learning is applied to the design and development of materials, and experimental data including material components, process parameters, performance characteristics, etc. are trained to establish a correlation model between material components, processes and material properties. Further, the implicit relationship between the parameters is extracted, the influencing factors of the attributes and their changing laws are found, and the prediction model is obtained. In this way, the design optimization of material components and process parameters is completed according to the required performance requirements, greatly improving the efficiency and accuracy of material research and development.

[0004] There are many problems with existing prediction models. For example, CN110010210A uses a double hidden layer BP neural network to achieve the purpose of quickly and accurately designing alloy composition according to performance requirements. However, this method uses a multi-input and multi-output model, which requires a large number of experimental data samples to ensure that the model achieves a high degree of accuracy and has limitations. Another example is CN107577850A, which uses a BP neural network to predict shrinkage defects in TC4 titanium alloy castings, but the cross-validation method is not used to evaluate the network model during the network training process, which increases the randomness of the results. Summary of the invention

[0005] In view of the defects existing in the prior art, the purpose of the present invention is to provide a small sample material reverse prediction method, which can predict the target performance from the given material composition and process parameters, generate a multidimensional initial population including parameters such as composition and process, and use the Euclidean distance between the target performance and the predicted performance as the fitness function to search for the optimal components and processes that meet the target performance and complete the reverse design.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a small sample material reverse prediction method, comprising the steps of: obtaining data features of composite materials; preprocessing the data features; building a model group consisting of multiple multi-input single-output generalized regression neural networks in parallel and training them to characterize the implicit relationship between material composition, process and its performance; establishing a reverse prediction model combining the model group and genetic algorithm; comparing and screening out available prediction data and incorporating them into the total data set.

[0007] Furthermore, the data features include material composition, material process and material properties.

[0008] Furthermore, the material composition is the weight ratio of each type of material required to be added, the material process includes time, temperature, and pressure, and the material properties include elongation at break and hardness.

[0009] Furthermore, the preprocessing of the data features includes data cleaning, missing value processing, feature selection, and feature scaling of the obtained data features.

[0010] Furthermore, the model group consisting of multiple multi-input single-output generalized regression neural networks in parallel includes the steps of: establishing a GRNN network structure; completing the division of the sample set; performing network adaptive training and learning; optimizing the model to improve the performance and generalization ability of the model; and predicting the output performance of the provided material composition and process input parameters.

[0011] Furthermore, when the sample set is divided, the leave-one-out cross-validation method is used to validate the model.

[0012] Furthermore, the establishment of the reverse prediction model combining the model group and the genetic algorithm includes the steps of: initializing the parameters of the genetic algorithm; establishing an initial space; performing a selection operation on the population using the roulette method; setting the crossover probability, randomly selecting two individuals from the population, and then randomly selecting a crossover position to complete a single-point crossover operation; setting the mutation probability, randomly selecting an individual so that it undergoes a single-point mutation under the condition that the mutation probability is met; and continuously optimizing and updating the optimal solution and its fitness value until the maximum number of iterations is met.

[0013] Furthermore, when establishing the reverse prediction model combining the model group and the genetic algorithm, the Euclidean distance between the target performance and the GRNN prediction performance is used as the fitness function.

[0014] Furthermore, the establishing of the initial space specifically includes: using tent chaotic mapping to randomly generate individuals of uniform population size as the initial solution space.

[0015] Furthermore, when the available prediction data are screened out through comparison and incorporated into the total data set, the prediction data that meet the requirements are screened out through comparison and incorporated into the total data set by presetting the acceptable error range of each feature.

[0016] The effect of the present invention is that the generalized neural network GRNN used has the characteristics of strong nonlinear mapping ability and fast learning speed. By adopting the leave-one-out cross-validation method to perform adaptive learning of hyperparameters, the prediction accuracy of the model is improved, so that when the sample data is small, it can also achieve a good prediction effect. The network model can be used to carry out small sample and high-precision material component performance prediction research for the problem of insufficient learning samples caused by long experimental time and difficult data acquisition in the material design process. In addition, according to the independent characteristics of each material performance, a parallel model group is established, and the model group is composed of multiple multi-input single-output GRNN dimension reduction fitting prediction models, which can improve the accuracy of performance prediction; at the same time, a multidimensional initial population including parameters such as composition and process is generated by a genetic algorithm, and the Euclidean distance between the target performance and the GRNN prediction performance is used as the fitness function to search for the optimal components and processes that meet the target performance, and establish a reverse prediction model combined with a neural network model group and a genetic algorithm, which can greatly reduce the cost of experimental trial and error and improve the efficiency of material design. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a principle flow chart of a small sample material reverse prediction method of the present invention;

[0018] Figure 2 It is a schematic diagram of the GRNN network structure;

[0019] Figure 3 It is a schematic diagram of the overall module of the reverse prediction model. DETAILED DESCRIPTION

[0020] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0021] like Figure 1-3 As shown, the present invention proposes a small sample material reverse prediction method, comprising the steps of:

[0022] S1, obtain data characteristics of composite materials;

[0023] Specifically, we sorted out the previous experimental results and data from related databases, extracted data features such as the composition, process parameters and corresponding test performance of composite materials, and completed data preparation.

[0024] In this embodiment, the data features include material composition, material process and material properties. Material composition is the weight percentage of various materials to be added, including but not limited to metal materials, polymer materials, etc.; material process includes but not limited to time, temperature, pressure and other parameters; material properties include but not limited to elongation at break, hardness and other parameters.

[0025] S2, preprocessing of data features;

[0026] Specifically, preprocessing operations such as data cleaning, missing value processing, feature selection, and feature scaling are performed on the obtained data features.

[0027] It should be noted that preprocessing the characteristics of the data to be tested can screen out invalid data, ensure the quality and consistency of the overall data, and construct an initial data set with material composition and material process as input and material property characteristics as output.

[0028] S3, build a model group consisting of multiple multi-input single-output generalized regression neural networks in parallel and train them to characterize the implicit relationship between material composition, process and its performance;

[0029] Specifically, the leave-one-out cross-validation method is used to divide the data set into N training sets and validation sets (N is the number of samples), and the GRNN parallel model group is trained and validated.

[0030] Furthermore, building a model group consisting of multiple multi-input single-output generalized regression neural networks in parallel includes the following steps:

[0031] S31, establish the network structure of GRNN;

[0032] Specifically, the number of neurons in the input layer, pattern layer, summation layer and output layer is selected.

[0033] In this embodiment, the input layer is the data item attribute value, that is, the input vector X=[x1,x2,...x n ], where the number of neurons is the dimension of the input vector in the learning sample.

[0034] The number of neurons in the pattern layer is usually equal to the number of samples in the training data set. Each neuron uses the radial basis function as the transfer function, calculates the Euclidean distance between the input vector and each sample, and uses the distance as the input of the next summation layer:

[0035]

[0036] In the formula, X i represents the sample corresponding to the i-th neuron, and σ is the smoothing factor.

[0037] The output S of the first neuron node of the summation layer D Represents the sum of the distance vectors transmitted by all mode layers, and other nodes S j It represents the weighted average of the weight value after Gaussian function transformation of the mode layer transmission distance and its corresponding output.

[0038] The output layer is the last layer of the GRNN network. The number of its nodes is equal to the dimension of the output data. Each feature represents:

[0039]

[0040] S32, completing the division of the sample set;

[0041] Specifically, N sample data are divided into N parts, one of which is taken as the test set each time, and the remaining N-1 parts are taken as the training set.

[0042] It should be noted that due to the long experimental cycle and the difficulty in data acquisition, there are fewer trainable data samples. Therefore, the leave-one-out cross-validation method is used to verify the model to reduce data bias, prevent model overfitting, and evaluate the generalization ability of the model.

[0043] S33, performing network adaptive training and learning;

[0044] Specifically, the cross-validation method in step S32 is used to optimize the smoothing factor σ. For each cross-validation, the spread value is gradually increased from 0.1 to 5, the step size is 0.1, the neural network model is trained, and the mean square error (MSE) of the predicted output and the verified output is used as the evaluation index of σ. The global optimal hyperparameters and the optimal sample division method are continuously updated and stored according to the minimum mean square error, and the current optimal GRNN network model is established.

[0045] S34, optimize the model to improve the performance and generalization ability of the model;

[0046] Specifically, the model is further optimized by adjusting parameters such as network structure, learning rate, and number of iterations to improve the performance and generalization ability of the model.

[0047] S35, predicting output performance based on the provided material composition and process input parameters;

[0048] Specifically, the constructed GRNN model is called to predict the output performance of the provided material composition and process input parameters.

[0049] S4, establishing a reverse prediction model combining a model group and a genetic algorithm;

[0050] Specifically, the Euclidean distance between the target performance and the GRNN predicted performance is designed as the fitness function, and the optimal components and processes that meet the target performance are searched through genetic algorithm screening, crossover and mutation operations, and a component-performance inverse prediction model combining a neural network model group and a genetic algorithm is established.

[0051] Furthermore, establishing a reverse prediction model combining a model group and a genetic algorithm includes the following steps:

[0052] S41, initializing the parameters of the genetic algorithm;

[0053] Specifically, initialize the maximum number of evolutions T, population size S, and crossover probability P of the genetic algorithm. c and mutation probability P m And other parameters;

[0054] S42, establish initial space;

[0055] Specifically, tent chaotic mapping is used to randomly generate S uniform individuals as the initial solution space.

[0056] In this embodiment, the principle of tent mapping to generate uniform random numbers is as follows:

[0057]

[0058] Wherein, ε=0.499; j=2,3,...S; x1 is a random number between (0,1).

[0059] The V+1-dimensional data before the individual is the component ratio, and its generation principle is:

[0060]

[0061] Where is the chaotic mapping random number.

[0062] The individual remaining dimension data are process parameters, and their generation principle is:

[0063] pop(l)=L l +(U l -L l )·x

[0064] Where, L l is the minimum boundary of the l-th dimension feature; U l is the maximum boundary of the l-th dimension feature.

[0065] The individuals in the population are represented as chromosome structure data through spherical coordinate transformation and encoding processing, the fitness values ​​of all individuals are calculated, and the average fitness and the best fitness of the population and its individuals are recorded.

[0066] S43, using roulette wheel method to select population;

[0067] Specifically, the probability of an individual being selected is related to its fitness, that is, The smaller the fitness value, the closer it is to the optimal solution that satisfies the conditions, and the greater the probability that the individual will be selected. a , the ath individual is selected, otherwise a new selection is made.

[0068] S44, set the crossover probability, randomly select two individuals from the population, and then randomly select the position for crossover to complete the single-point crossover operation;

[0069] Specifically, set the crossover probability P c is 0.3, two individuals are randomly selected from the population, and then the position for crossover is randomly selected to complete the single-point crossover operation, so that the excellent genes of the parent generation are inherited to the offspring, thereby generating new excellent individuals.

[0070] S45, setting the mutation probability, and randomly selecting an individual so that it undergoes a single-point mutation under the condition that the mutation probability is satisfied;

[0071] Specifically, set the mutation probability P m is 0.05, and an individual is randomly selected to undergo single-point mutation under the condition that the mutation probability is satisfied. The mutation operation can prevent the genetic algorithm from falling into the local optimal solution during the optimization process.

[0072] S46, continuously optimizing and updating the optimal solution and its fitness value until the maximum number of iterations is met;

[0073] Specifically, the optimal solution and its fitness value are continuously optimized and updated until the maximum number of iterations T is met, then the update is stopped and the optimal solution is output.

[0074] S5, compare and filter out the available prediction data and incorporate them into the total data set;

[0075] Specifically, the available prediction data are compared and screened out and incorporated into the total data set, and S2-S4 are repeated to establish a complete database.

[0076] In this embodiment, when the available prediction data are screened out, the acceptable error range of each feature is preset, and the prediction data that meets the requirements are compared and screened out and incorporated into the total data set.

[0077] It can be seen from the above embodiments that the generalized neural network GRNN used in the present invention has the characteristics of strong nonlinear mapping ability and fast learning speed. By adopting the leave-one-out cross-validation method for adaptive learning of hyperparameters, the prediction accuracy of the model is improved, so that it can achieve better prediction results when the sample data is small. The network model can be used to conduct small sample and high-precision material composition performance prediction research for the problem of insufficient learning samples caused by long experimental time and difficulty in data acquisition in the material design process.

[0078] In addition, based on the independent characteristics of various material properties, a parallel model group is established. This model group consists of multiple multi-input and single-output GRNN dimensionality reduction fitting prediction models, which can improve the accuracy of performance prediction. At the same time, a multidimensional initial population including parameters such as composition and process is generated through a genetic algorithm. The Euclidean distance between the target performance and the GRNN predicted performance is used as the fitness function to search for the optimal components and processes that meet the target performance. A reverse prediction model combining a neural network model group and a genetic algorithm is established, which can greatly reduce the cost of experimental trial and error and improve the efficiency of material design.

[0079] The present invention is not limited to the embodiments described in the specific implementation manner. Those skilled in the art may derive other implementation manners based on the technical solution of the present invention, which also fall within the technical innovation scope of the present invention.

Claims

1. A small sample material reverse prediction method, characterized in that: Includes steps: Obtain data characteristics of composite materials; Preprocess the data features; Build and train a model group consisting of multiple multi-input single-output generalized regression neural networks in parallel to characterize the implicit relationship between material composition, process and its performance; Establish a reverse prediction model combining model swarm and genetic algorithm; The available prediction data were screened out and incorporated into the total data set.

2. A small sample material reverse prediction method as claimed in claim 1, characterized in that: The data characteristics include material composition, material process and material properties.

3. A small sample material reverse prediction method as claimed in claim 2, characterized in that: The material composition is the weight ratio of each type of material to be added, the material process includes time, temperature, and pressure, and the material properties include elongation at break and hardness.

4. A small sample material reverse prediction method as claimed in claim 1, characterized in that: The preprocessing of the data features includes data cleaning, missing value processing, feature selection, and feature scaling of the obtained data features.

5. A small sample material reverse prediction method as claimed in claim 1, characterized in that: The method of building a model group consisting of multiple multi-input single-output generalized regression neural networks in parallel includes the following steps: Establish the network structure of GRNN; Complete the division of sample sets; Conduct network adaptive training and learning; Optimize the model to improve its performance and generalization ability; Output properties are predicted for the provided material composition and process input parameters.

6. A small sample material reverse prediction method as claimed in claim 5, characterized in that: When the sample set is divided, the leave-one-out cross-validation method is used to validate the model.

7. A small sample material reverse prediction method as claimed in claim 1, characterized in that: The method of establishing a reverse prediction model combining a model group and a genetic algorithm comprises the following steps: Initialize the parameters of the genetic algorithm; Establish the initial space; Use the roulette wheel method to select the population; Set the crossover probability, randomly select two individuals from the population, and then randomly select the position for crossover to complete the single-point crossover operation; Set the mutation probability and randomly select an individual to make a single-point mutation under the condition that the mutation probability is met; Continuously optimize and update the optimal solution and its fitness value until the maximum number of iterations is met.

8. A small sample material reverse prediction method as claimed in claim 1, characterized in that: When establishing the reverse prediction model combining the model group and the genetic algorithm, the Euclidean distance between the target performance and the GRNN prediction performance is used as the fitness function.

9. A small sample material reverse prediction method as claimed in claim 7, characterized in that: The specific method of establishing the initial space is: using tent chaotic mapping to randomly generate individuals with a uniform population size as the initial solution space.

10. A small sample material reverse prediction method as claimed in claim 1, characterized in that: When the available prediction data are screened out by comparison and incorporated into the total data set, the prediction data that meets the requirements are screened out by comparison and incorporated into the total data set by presetting the acceptable error range of each feature.

Images