Machine Learning Force Field Model Based On First-principles Calculations

Machine learning force field(ML-FF)based one first principle calculations has a great potential to study the properties of materials on a larger in both spatial and temporal scales.At present,linear regression(LR)and neural network(NN)are two of the most commonly used machine learning models.The NN has the advantage of training complex and nonlinear mapping functions in a relatively accurate and flexible manner.The LR on the other hand,has its advantages to simplify the training.When the system deviates from the original training set,the LR model is more stable.Besides,the molecular dynamics(MD)simulation cost in the LR model is relatively low.By contrast,although the NN model is more accurate and flexible in the training complex and non-linear mapping functions,the computation is costly.Therefore,it is of great interest to compare NN and LR models for different research systems.Herein,a comparative study of the LR and NN force fields has been performed for the copper system.It is found that,for systems close to the equilibrium,such as the bulk Cu,the LR and NN model almost give similar accuracy results.However,for systems far from the equilibrium,NN is more accurate compared with the LR model.Taking the copper-carbon system as a prototype,we have developed a new ML-FF model for binary material systems.The calculation accuracy of the newly developed model is much higher than that of the LR model for complex systems,which is comparable to that of the NN model.Importantly,its calculation speed is much faster than the NN model.We therefore conclude that the new ML model is universal and especially suitable for systems with large data volumes and complex functions.This work provides certain theoretical guidance for selecting machine learning models for different systems.