Skip to Main content Skip to Navigation
Journal articles

Towards better efficiency of interatomic linear Machine Learning potentials

Abstract : Interatomic machine learning potentials have achieved maturity and became worthwhile alternative to conventional interatomic potentials. In this work we profile some characteristics of linear machine learning methods. Being numerically fast and easy to implement, these methods offer many advantages and appear to be very attractive for large length and time scale calculations. However, we emphasize that in order to be accurate on some target properties these methods eventually yield overfitting. This feature is rather independent of training database and descriptor accuracy. At the same time, the major weakness of these potentials, i.e., lower accuracy with respect to the kernel potentials, proves to be their strength: within the confidence limits of the potential fitting, one can rely on less accurate but faster descriptors in order to boost the numerical efficiency. Here, we propose a hybrid type of atomic descriptor that combines the original forms of radial and spectral descriptors. Flexibility in choice of mixing proportions between the two descriptors ensures a user defined control over accuracy / numerical efficiency of the resulting hybrid descriptor form. The performance and features of the above linear machine learning potentials are investigated for the interatomic interactions in metals of primary importance for fusion and fission applications, Fe and W. The suggested hybrid approach opens many avenues in the field of linear machine learning potentials that up to now are preferentially coupled with more robust and computationally expensive spectral descriptors.
Complete list of metadatas
Contributor : Contributeur Map Cea <>
Submitted on : Friday, January 17, 2020 - 11:00:14 AM
Last modification on : Thursday, June 25, 2020 - 2:54:03 PM



Alexandra Goryaeva, Jean-Bernanrd Maillet, Mihai Cosmin Marinica. Towards better efficiency of interatomic linear Machine Learning potentials. Computational Materials Science, Elsevier, 2019, 166, pp.200-209. ⟨10.1016/j.commatsci.2019.04.043⟩. ⟨cea-02443478⟩



Record views