Coating impact damage prediction method based on material genome

Abstract

The invention discloses a coating impact damage prediction method based on a material genome. The method comprises the following steps: (1) obtaining original data through an impact experiment and related tests; (2) taking the residual stress, the maximum shear stress, the shear stress depth and the porosity of the amorphous coatings with different thicknesses as material gene characteristic data, taking the crack length caused by impact damage as performance data, constructing a corresponding relation database, and obtaining an original data set of a machine learning model; (3) establishing eight machine learning models, and dividing a training set and a test set through 5-fold cross validation; (4) performing grid search on the training set to obtain an optimal hyper-parameter of the model, and further obtaining an optimal machine learning model; (5) evaluating the data in the test set to obtain a machine learning model with the highest accuracy; and (6) predicting impact damage according to the machine learning model with the highest accuracy. According to the method, impact damage prediction is realized based on machine learning, and the precision of performance prediction in coating service performance research is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of coating impact performance prediction, and in particular to a coating impact damage prediction method based on material genome. Background Art

[0002] Material genetic engineering technology integrates high-throughput experiments, high-throughput calculations and databases, and uses data mining technology to systematically find the quantitative relationship between material components / processes-organizational structures-performance, and then rationally design materials with the help of integrated computational simulations and key experiments, accelerate the entire process of "discovery-research and development-production-application" of materials, shorten the material research and development cycle, reduce the cost of material research and development, and trigger scientific and technological innovation and business model changes in the field of new materials. Faced with complex and nonlinear material relationships, conventional statistical analysis methods often find it difficult to construct an effective prediction model. Among them, a method of implementing artificial intelligence technology in the machine learning room has been successfully applied to the construction of material relationships, selecting important parameters that affect the target performance as input features, and establishing the optimal relationship model between input and output, and finally making efficient predictions of the composition, process, organization and performance of the material.

[0003] As a structural material, the impact performance of the coating is an important indicator for evaluating its safety, reliability and effectiveness during loading. Recently, researchers have introduced material genome engineering technology into the development of alloys and have achieved certain results. However, the quantitative relationship between the microstructure of the coating and the impact performance has not yet been established. Therefore, if we can develop a theoretical model of relevant impact failure behavior and establish a quantitative relationship between impact performance and process and structure, so as to achieve rapid prediction of impact damage, it will be of great significance both in the preparation of high-performance amorphous alloy coatings and in the research and development of engineering materials. Summary of the invention

[0004] The purpose of the present invention is to reduce the complexity of model construction in the process of coating damage research, establish a quantitative relationship between impact damage and process and structure, so as to achieve rapid prediction of impact damage, and propose a prediction method for coating impact damage based on material genome.

[0005] The technical solution implemented by the present invention is as follows: a coating impact damage prediction method based on material genome, the steps are as follows:

[0006] (1) Obtain original data through impact mechanical properties test, residual stress test, shear stress test and porosity test;

[0007] (2) The residual stress σ, maximum shear stress τ, shear stress depth γ, and porosity p of a series of amorphous coatings with different thicknesses are used as the input feature data of the material gene, and the crack length L caused by impact damage is used as the output damage data. A corresponding relationship database is constructed to obtain the original data set of the machine learning model;

[0008] (3) establishing eight machine learning models, and dividing the original data set into a training set and a test set by 5-fold cross validation;

[0009] (4) Train the machine learning models in the training set and obtain eight optimal machine learning models with optimal hyperparameters in the machine learning models through grid search;

[0010] (5) Evaluate the data in the test set using the eight best machine learning models under hyperparameters to obtain the machine learning model with the highest accuracy;

[0011] (6) Predict impact damage based on a highly accurate machine learning model.

[0012] The amorphous coatings of different thicknesses refer to amorphous coatings with a thickness of 200 to 800 μm obtained by changing different thermal spraying times.

[0013] The crack length L caused by the impact damage is obtained by performing a drop hammer impact test on the amorphous coating;

[0014] The residual stress σ of the coating is obtained by the curvature method;

[0015] The maximum shear stress τ and the shear stress depth γ are obtained by finite element simulation software;

[0016] The porosity p is obtained by an X-ray three-dimensional imaging system.

[0017] The machine learning model uses Anaconda3 software to perform model training using a machine learning algorithm.

[0018] The eight machine learning models include: ridge regression model (RG), support vector machine model (SVR), k nearest neighbor model (KNN), decision tree model (DT), extreme forest model (ETSR), random forest model (RF), ensemble model (BAG) and artificial neural network model (MLP).

[0019] The 5-fold cross-validation first randomly divides the data set into 5 non-overlapping subsets of the same size; then the first subset is used as the test set, and the remaining 4 subsets are used as training sets to train the model; return to the previous step, use the second subset as the test set, and the remaining 4 subsets as training sets to train the model, and perform this process in sequence until each subset has been used as a test set; the ratio of the training set to the test set is 4:1, and after combining with the subsequent 5-fold cross-validation, the test set will cover the entire sample space.

[0020] The optimal hyperparameters in the machine learning model are tuned using grid search (GridSearchCV).

[0021] The eight optimal machine learning models under the hyperparameters are used to evaluate the data in the test set, and the evaluation index is the goodness of fit R 2 , mean square error MSE, root mean square error RMSE and mean absolute percentage error MAE, the expressions of each coefficient are as follows:

[0022]

[0023] Where n is the number of samples; y i is the actual value of the sample; is the predicted value of the model; is the sample average; goodness of fit R 2 The value is [0,1], R 2 The closer it is to 1, the better the model fits, and the closer it is to 0, the worse the model fits. The root mean square error (MSE) is the mean of the sum of the squares of the errors between the corresponding points of the predicted value and the actual value. The closer the MSE is to 0, the more perfect the model is. The root mean square error (RMSE) is MSE with a square root added, and its order of magnitude is more intuitive. The closer the RMSE is to 0, the more consistent the predicted value is with the actual value, and the more perfect the model is. The average absolute error (MAE) reflects the actual situation of the predicted value error. The closer the MAE is to 0, the more perfect the model is.

[0024] The prediction method is applicable to various coatings.

[0025] The beneficial effect of the present invention is that the present invention proposes a method for predicting coating impact damage based on the material genome, establishes a quantitative relationship between impact damage and process and structure, thereby realizing the rapid prediction of coating impact damage. The method of the present invention is easy to implement and simple to operate, can be widely applied to different coatings, and can solve the problem of impact performance dispersion caused by impact tests. Based on the genetic engineering of materials, the present invention develops a method for rapidly predicting coating impact damage and a method for rapidly estimating interface shear strength, which opens up new ideas for how to optimize coating impact damage, reduces the complexity of model construction in the coating damage research process, and improves the accuracy of damage prediction in the coating service performance research process. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The present invention is a flowchart of the steps of the coating impact damage prediction method based on the material genome;

[0027] Figure 2 The heat map of the Pearson correlation coefficient distribution between the four features and damage;

[0028] Figure 3 Comparison of the evaluation indicators R2, MSE, RMSE and MAE of 8 machine learning models;

[0029] Among them, RG is the ridge regression model; SVR is the support vector machine model; KNN is the k nearest neighbor model; DT is the decision tree model; ETSR is the extreme forest model; RF is the random forest model; MLP is the artificial neural network model; BAG is the integrated model;

[0030] Figure 4 It is a schematic diagram of the artificial neural network model;

[0031] Figure 5 This is the impact cracking diagram of amorphous coating;

[0032] Figure 6 This is a diagram of impact cracking of crystal coating. DETAILED DESCRIPTION

[0033] The specific embodiments of the present invention are as follows Figure 6 shown.

[0034] like Figure 1 As shown, this embodiment is a coating impact damage prediction method based on material genome, the steps are as follows:

[0035] (1) The original data are obtained through impact mechanical properties test, residual stress test, shear stress test and porosity test.

[0036] (2) The residual stress σ, maximum shear stress τ, shear stress depth γ, and porosity p of a series of amorphous coatings with different thicknesses are used as the input feature data of the material gene, and the crack length L caused by impact damage is used as the output damage data. A corresponding relationship database is constructed to obtain the original data set of the machine learning model. The thermodynamic diagram between the four features and the damage data is obtained based on the Pearson correlation analysis, as shown in Figure 2 As shown in the figure, it can be seen that the crack length L has a strong correlation with these four characteristics. It is positively correlated with the residual stress at 0.76, and negatively correlated with the maximum shear stress, shear stress depth, and porosity, which are -0.58, -0.79, and -0.74, respectively.

[0037] (3) Eight machine learning models were established, including ridge regression model (RG), support vector machine model (SVR), k-nearest neighbor model (KNN), decision tree model (DT), extreme forest model (ETSR), random forest model (RF), artificial neural network model (MLP) and ensemble model (BAG). The original data sets of the machine learning models were divided into training set and test set through 5-fold cross validation.

[0038] (4) Train the machine learning models in the training set and obtain the optimal hyperparameters through grid search, thereby obtaining 8 optimal machine learning models under the optimal parameters.

[0039] (5) Evaluate the data in the test set using the optimal machine learning model to obtain the machine learning model with the highest accuracy, such as Figure 3 As shown, the artificial neural network model has the highest accuracy, and the goodness of fit of the training set Training is R 2 The highest is 0.9997, the lowest root mean square error MSE is 0.0016, the lowest root mean square error RMSE is 0.0398, and the lowest average absolute error MAE is 0.0274; the goodness of fit R 2 The highest is 0.9889, the lowest root mean square error MSE is 0.0729, the lowest root mean square error RMSE is 0.2699, and the lowest average absolute error MAE is 0.1847.

[0040] (6) Predict impact damage based on a highly accurate machine learning model.

[0041] like Figure 4As shown, the hyperparameters of the artificial neural network model are tuned as follows: the hidden layer has a single layer with 3 neurons hidden_layer_sizes=(3), the activation function activation='tanh', the gradient descent method solver='lbfgs', the penalty coefficient alpha=0.0001, the maximum number of iterations max_iter=91, and the random state random_state that determines the weight is 21.

[0042] From input layer to hidden layer:

[0043] where h j There are three neural units (j=1, 2, 3), x i is the input layer (x1=σ, x2=τ, x3=γ, x4=p), weight w ij is the effect of input on neuron output, obtained through training optimization, b j is the bias term, which is optimized together with the weight. Therefore, the hidden layers are h1=x1w 11 +x2w 21 +x3w 31 +x4w 41 +b1,h2=x1w 12 +x2w 22 +x3w 32 +x4w 42 +b2,h3=x1w 13 +x2w 23 +x3w 33 +x4w 43 +b3.

[0044] From hidden layer to output layer: in

[0045] Therefore, the crack length predicted by the artificial neural network model for impact damage is L = tanh(h1) + tanh(h2) + tanh(h3) = tanh(x1w 11 +x2w 21 +x3w 31 +x4w 41 +b1)+tanh(x1w 12 +x2w 22 +x3w 32 +x4w 42 +b2)+tanh(x1w 13 +x2w 23 +x3w 33 +x4w 43+b3). According to this model, the impact damage of amorphous coating and crystalline coating is predicted.

[0046] Embodiment 1:

[0047] This embodiment is to predict impact cracking of an amorphous coating.

[0048] Step 1: Perform drop hammer impact test on coatings with different thicknesses of 200-800μm, such as Figure 5 As shown, a database of coating residual stress, maximum shear stress, shear stress depth, porosity, and coating crack length is obtained.

[0049] Step 2: The artificial neural network machine learning model MLP was used to predict the impact damage of the amorphous coating using Anaconda3 software. It was found that the predicted crack lengths were 0, 0, 0, 6.251 μm, 7.281 μm, 5.840 μm, and 4.413 μm, respectively, which were consistent with the actual measurements (e.g. Figure 5 The crack lengths in Figures a to g are 0, 0, 0, 6.136μm, 7.331μm, 5.559μm, and 4.615μm, which are close.

[0050] Therefore, this model is suitable for crack length prediction of amorphous coatings in engineering applications.

[0051] Embodiment 2:

[0052] This embodiment is to predict impact cracking of crystalline coatings.

[0053] Step 1: Perform drop hammer impact test on coatings of different thicknesses, such as Figure 6 As shown, a database of coating residual stress, maximum shear stress, shear stress depth, porosity, and coating crack length is obtained.

[0054] Step 2: The artificial neural network machine learning model MLP was used to predict the impact damage of the amorphous coating using Anaconda3 software. It was found that the predicted crack lengths were 0, 0, 4.544 μm, 5.776 μm, 5.423 μm, 2.652 μm, and 2.176 μm, respectively, which were consistent with the actual measurements (e.g. Figure 6 The crack lengths of (a) to (g) are close: 0, 0, 4.268μm, 5.894μm, 5.276μm, 2.986μm, and 2.368μm.

[0055] Therefore, this model is suitable for crack length prediction of crystalline coatings in engineering applications.

[0056] The above embodiments are merely illustrative of the principles and performances of the present invention and are not exhaustive. People can also obtain other embodiments based on the present embodiments without creative work, and these embodiments all fall within the protection scope of the present invention.

Claims

1. A coating impact damage prediction method based on material genome, characterized in that: The prediction method steps are as follows: (1) Obtain original data through impact mechanical properties test, residual stress test, shear stress test and porosity test; (2) The residual stress σ, maximum shear stress τ, shear stress depth γ, and porosity p of a series of amorphous coatings with different thicknesses are used as the input feature data of the material gene, and the crack length L caused by impact damage is used as the output performance data to construct a corresponding relationship database and obtain the original data set of the machine learning model; (3) Eight machine learning models were established and the original data set was divided into training and test sets through 5-fold cross validation; (4) Train the machine learning models in the training set and obtain eight optimal machine learning models with optimal hyperparameters in the machine learning models through grid search; (5) Evaluate the data in the test set using the eight optimal machine learning models under hyperparameters to obtain the machine learning model with the highest accuracy; (6) Predict impact damage based on the machine learning model with the highest accuracy.

2. The coating impact damage prediction method based on material genome according to claim 1 is characterized in that: The amorphous coatings of different thicknesses refer to amorphous coatings with a thickness of 200 to 800 μm obtained by changing different thermal spraying times.

3. The coating impact damage prediction method based on material genome according to claim 1 is characterized in that: The crack length L caused by the impact damage is obtained by performing a drop hammer impact test on the amorphous coating; The residual stress σ of the coating is obtained by the curvature method; The maximum shear stress τ and the shear stress depth γ are obtained by finite element simulation software; The porosity p is obtained by an X-ray three-dimensional imaging system.

4. The coating impact damage prediction method based on material genome according to claim 1 is characterized in that: The machine learning model uses Anaconda3 software to perform model training using a machine learning algorithm.

5. The coating impact damage prediction method based on material genome according to claim 1 is characterized in that: The eight machine learning models include: linear ridge regression model, support vector machine model, k nearest neighbor model, decision tree model, extreme forest model, random forest model, ensemble model and artificial neural network model.

6. The coating impact damage prediction method based on material genome according to claim 1 is characterized in that: The 5-fold cross-validation first randomly divides the data set into 5 non-overlapping subsets of the same size; then the first subset is used as the test set, and the remaining 4 subsets are used as training sets to train the model; return to the previous step, use the second subset as the test set, and the remaining 4 subsets as training sets to train the model, and perform this process in sequence until each subset has been used as a test set; the ratio of the training set to the test set is 4:1, and after combining with the subsequent 5-fold cross-validation, the test set will cover the entire sample space.

7. The coating impact damage prediction method based on material genome according to claim 1 is characterized in that: The optimal hyperparameters in the machine learning model are tuned using grid search.

8. The coating impact damage prediction method based on material genome according to claim 1, characterized in that: The eight optimal machine learning models under the hyperparameters are used to evaluate the data in the test set, and the evaluation index is the goodness of fit R 2 , mean square error MSE, root mean square error RMSE and mean absolute percentage error MAE, the expressions of each coefficient are as follows: Where n is the number of samples; y i is the actual value of the sample; is the predicted value of the model; is the sample average; goodness of fit R 2 The value is [0,1], R 2 The closer it is to 1, the better the model fits, and the closer it is to 0, the worse the model fits. The root mean square error (MSE) is the mean of the sum of the squares of the errors between the corresponding points of the predicted value and the actual value. The closer the MSE is to 0, the more perfect the model is. The root mean square error (RMSE) is MSE with a square root added, and its order of magnitude is more intuitive. The closer the RMSE is to 0, the more consistent the predicted value is with the actual value, and the more perfect the model is. The average absolute error (MAE) reflects the actual situation of the predicted value error. The closer the MAE is to 0, the more perfect the model is.

9. The coating impact damage prediction method based on material genome according to claim 1, characterized in that: The prediction method is applicable to various coatings.

Images