Machine Learning Interatomic Potential Accelerates Large-scale Simulations Of High Thermal Conductivity Material Boron Arsenide

In recent years,both theoretical and experimental confirmation of very high thermal conductivity in III-V semiconductor boron arsenide(BAs)at room temperature has made it an excellent material for improving the thermal management efficiency of systems such as chips and lithium-ion batteries.Despite the highly symmetrical and simple structure of cubic boron arsenide(c-BAs),the exact value of its thermal conductivity is still not fully agreed from a theoretical and experimental point of view and lack of studies in high temperature.Machine learning based on artificial intelligence,especially machine learning interatomic potential that combine the accuracy of the first-principles calculations and the efficiency of classical interatomic potential,provide a new method to accurately and quickly predict the thermal conductivity of c-BAs.In this thesis,a new method for predicting thermal conductivity is proposed,that is,the thermal conductivity of c-BAs is accurately predicted by using machine learning interatomic potential,combined with lattice dynamics,Peierls-Boltzmann transport equation and molecular dynamics.The main studies include:The applicability scenarios of different training strategies for the interatomic potential are explored with the help of machine learning,with accuracy and reliability as evaluation metrics.The passive learning strategy relies on manual collection of atomic configurations to complete the training process,which is simple and easy to operate,and is suitable for scenarios where some accuracy can be sacrificed for computational speed.Active learning strategies rely on the D-optimality criterion and the Maxvol algorithm extrapolation to automatically select atomic configurations to complete the training process,which is a complex and time-consuming process.The accuracy is nearly 80%higher than that of passive learning,which is comparable to the first principle calculation results.It is suitable for scenes that require high accuracy and reliability.Machine learning interatomic potential is applied to calculate the interatomic force constants.The computational speed about 5 orders of magnitude higher than first-principles calculations,and the dispersion relation of phonons is accurately predicted.The accurate atomic trajectory is obtained by machine learning interatomic potential,and the phonon renormalization is carried out.Combined with the three phonon and four phonon scattering coupling theory,the phonon transport properties of c-BAs are explored.It is found that temperature-induced phonon renormalization increases the probability of phonon scattering by reducing the frequency of high-frequency phonons,so as to enhance the intensity of phonon scattering.In addition,the anharmonicity phonon scattering enhances with increasing temperature.The combined effect reduces the thermal conductivity of c-BAs.The thermal conductivity of c-BAs is accurately predicted based on machine learning interatomic potential by using equilibrium molecular dynamics and non-equilibrium molecular dynamics methods that naturally consider the high-order phonons scattering.Machine learning potential obtained based on different training strategies,passive learning is not applicable to long time scale molecular dynamics simulations,and active learning can accurately predict the thermal conductivity of c-BAs.The results show that the predicted values of the two methods are consistent in the high temperature scenario,and the predicted values at 1100k are 84.82 ±2.11 and 95 Wm-1K-1 respectively.The calculation result of thermal conductivity at room temperature is 1457± 50.98 Wm-1K-1,which is in good agreement with the current experimental value of 900-1300 Wm-1K-1.It is also found that the high-order phonons scattering significantly reduces the thermal conductivity of c-BAs by about 45%at 900-1200K,and the effect on the thermal conductivity of c-BAs is almost negligible at 300-800K.