Machine Learning Study Based On First-principles Of Solutes Properties In Metals

With the rapidly evolving computing power,people can obtain the desired parameters of materials properties used for materials development and modification through simulation calculation.First-principles(FP)method based on density functional theory(DFT)(DFT-FP)has become a common method of simulation calculation as its high-precision results.Although DFT-FP has greatly promoted the evolution of materials science,it is computationally intensive and time-consuming,requiring great computational cost.In recent years,machine learning(ML)methods have been gradually applied to materials science with the accumulation of data.ML methods have the potential to substantially alter and heighten the computational efficiency of properties parameters in materials science,thereby accelerating the development and progress of materials science.In this paper,we systematically studied the interaction between transition metal(TM)substitutional solutes and hydrogen/helium/carbon/nitrogen/oxygen interstitial solutes and the diffusion behavior of TM solutes in metals using DFT-FP and ML method based on support vector machine(SVM)(SVM-ML),exploring the application of ML in this field.The main contents and results are as follows:Firstly,we studied the interaction between TM solutes and oxygen in tungsten using DFT-FP.A database of binding energies of TM solutes-oxygen pairs and TM solutes-vacancy-oxygen complexes was established,which provides data support for subsequent SVM-ML study.The physical mechanism of TM solutes-oxygen interaction can be understood by the combination of elastic and electronic interactions,which arises from the elastic interaction for undersized TM solutes and the electronic interaction for oversized TM solutes.These analyses provide a physical basis for subsequent ML feature selection.Combining the DFT-FP results in this paper,previous calculation data of our group and literature data,we obtained a dataset of binding energies that characterizes the interaction between substitutional solutes and interstitial solutes in metals,and studied it using SVM-ML.According to the DFT-FP analysis results of physical mechanism,some parameters such as atomic radius,melting point,unpaired d electrons,Pauling electronegativity and solute concentration were selected as the basic feature set.Then,the fit,leave-one-out cross-validation(LOO CV)and extrapolation procedures for the dataset were performed.For LOO CV,the root mean square error(RMSE)is 0.102 eV and the squared correlation coefficient(r2)is 0.923.This indicates that the selected features can reflect the binding energy well,which confirms the DFT-FP analysis results of physical mechanism.Our work shows that SVM-ML can predict the binding energies of numerous new alloy systems with quite small computational cost,and these predictions provide valuable references for subsequent theoretical and experimental researches.Combining the previous data accumulation of the group and literature survey,we established a dataset of TM solutes diffusion in metals containing 240 diffusion barriers,and analyzed it using SVM-ML.By optimizing input features,we got the optimal combination of features,including ionic radius,bulk moduli,melting point and unpaired d electrons.We futher performed the fit,LOO CV and extrapolation procedures.The RMSE=0.122 eV and r2=0.939 from LOO CV,indicating that the SVM-ML predictions are in good agreement with the DFT-FP calculations based on the selected features.Furthermore,we predicted TM solutes diffusion barriers in 16 new metal hosts.The change rules of these predictions with atomic number are consistent with the theory prediction rules of DFT-FP in the literature.The above results demonstrate that SVM-ML can predict the parameters of materials properties with quite small computational cost,and is a promising method to heighten the computational efficiency and accelerate materials science researches.