Prediction Of Second-phase Work Function And Corrosion Potential Of Biological Magnesium Alloy Based On Machine Learning

Magnesium alloys have become promising new biomedical materials due to their good mechanical properties,degradation properties and biocompatibility.However,magnesium alloy implants may lose their supporting function prematurely due to rapid degradation in physiological environment.Therefore,studying the corrosion degradation properties of magnesium alloys is one of the hot issues of biomedical magnesium alloys.It is well known that the most typical corrosion type of biological magnesium alloys is electrochemical corrosion.The research methods of electrochemical corrosion mainly include traditional experimental analysis method and the first-principles calculation method based on density functional theory（DFT）.The traditional trial-and-error method is inefficient and often consumes a lot of experimental materials and cycles.The first-principles calculation method based on DFT is usually only for a specific system,and there are problems of large amount of calculation and long period of time consumption for complex systems.Machine learning can effectively use big data generated from computational simulations and material experiments to predict the properties of materials.In this paper,the method of machine learning was adopted to study the biological magnesium alloys.For the electrochemical corrosion problem,the calculation data of electronic work function that can reflect the galvanic corrosion trend of the second phase and the magnesium matrix,and experimental data of corrosion potential that can reflect the corrosion resistance of alloys were collected respectively as the targets.The machine learning models were established to realize the accurate predictions of electronic work function and corrosion potential.The main results are as follows:The electronic work function of different second phases calculated based on the first principles of DFT were collected as the target,and 19 features including the number of element atoms,properties of alloying elements,optimized cell parameters,formation energy and surface energy were also collected,with a total of 150 rows and 20 columns of data.Pearson correlation coefficient（Pearson）between each feature was calculated for correlation analysis,and four groups of features are obtained as strong correlation features:the atomic weight and period of alloying elements;the electronegativity,radius of alloying elements,and the ratio of element atomic radius to magnesium atomic radius;cell parametersαandβ;cell parameterγand different crystal plane indices.The feature importance analysis was carried out by calculating the Pearson and Spearman correlation coefficients（Spearman）between each feature and the target.It was found that the group,electronegativity and radius of alloying elements are important features of the work function.Six models including Multiple Linear Regression（MLR）,Ridge Regression（RR）,Support Vector Regression（SVR）,Random Forest（RF）,Gradient Boosting Regression Tree（GBRT）and e Xtreme Gradient Boosting Tree（XGBoost）were established to train and predict work function data.The accuracy of each model was evaluated by Coefficient of Determination（R²）,Mean Absolute Error（MAE）,Root Mean Square Error（RMSE）.The R² of the SVR model on the training set is 0.731,and the R² on the testing set is 0.654,which has the highest prediction accuracy.The work function of different crystal planes of the Mg₇Zn₃ phase calculated by the first-principles based on DFT was compared with the predicted value of the SVR model.The error between the calculated and predicted value is small,and the absolute error range is between plus and minus 0.2 e V,which proves that the model has good generalization ability.The electrochemical corrosion potentials of different alloys were collected as the target,and 15 features including chemical composition,processing technology,corrosive medium and concentration were also collected,with a total of 330 rows and16 columns of data.Through feature correlation analysis,it was found that features are independent of each other,and there are no strong correlation feature groups.Through feature importance analysis,the content of Mg,Mn,Al alloying elements and the concentration of corrosive medium are the important features of the corrosion potential.MLR,RR,SVR,RF,GBRT and XGBoost models were established to train and predict corrosion potential data.Through model evaluation,the R² of the RF model on the training set is 0.945,and the R² on the testing set is 0.702,which has the highest prediction accuracy.The electrochemical corrosion potentials of as-cast Mg,Mg-5Zn,Mg-3Nd,Mg-5Nd,Mg-30Nd,Mg-0.5Zn-0.5Nd-1.5Sc and as-extruded Mg-3Nd and Mg-2Zn-0.46Y-0.5Nd alloys in 3.5%and 0.9%Na Cl solution were measured,respectively.Compared with the predicted value of the RF model,the error range between the predicted and measured value is between-0.05 and 0.1 V,the error is small,which verifies that RF model has a good generalization ability for corrosion potential prediction.