Machine learning enhanced empirical potentials for metals and alloys

Abstract

Interatomic potential (i.e. force-field) plays a vital role in atomistic simulation of materials. Empirical potentials like the embedded atom method (EAM) and its variant angular-dependent potential (ADP) have proven successful in many metals. In the past few years, machine learning has become a compelling approach for modeling interatomic interactions. Powered by big data and efficient optimizers, machine learning interatomic potentials can generally approximate to the accuracy of the first-principles calculations based on the quantum mechanics theory. In this works, we successfully developed a route to express EAM and ADP within machine learning framework in highly-vectorizable form and further incorporated several physical constraints into the training. As it is proved in this work, the performances of empirical potentials can be significantly boosted with few training data. For energy and force predictions, machine tuned EAM and ADP, can be almost as accurate as the computationally expensive spectral neighbor analysis potential (SNAP) on the fcc Ni, bcc Mo and Mo-Ni alloy systems. Machine learned EAM and ADP can also reproduce some key materials properties, such as elastic constants, melting temperatures and surface energies, close to the first-principles accuracy. Our results suggest a new and systematic route for developing machine learning interatomic potentials. All the new algorithms have been implemented in our program TensorAlloy. Program summaryProgram Title: TensorAlloyCPC Library link to program files:https://doi.org/10.17632/w8htd7vmwh.2Code Ocean capsule:https://codeocean.com/capsule/1671487Licensing provisions: LGPLProgramming language: Python 3.7Journal reference of previous version: Comput. Phys. Commun. 250 (2020) 107057, https://doi.org/10.1016/j.cpc.2019.107057Does the new version supersede the previous version?: YesReasons for the new version: This new version is a significant extension to the previous version. Now machine learning approaches and physical constraints can be used together to tune empirical potentials (for example the embedded atom method). Machine learning optimized empirical potentials can be almost as accurate as machine learning interaction potentials but run much faster.Nature of problem: Optimizing empirical potentials with machine learning approaches and physical constraints.Solution method: The TensorAlloy program is built upon TensorFlow and the virtual-atom approach. We successfully developed a route to express the embedded atom method and the angular-dependent potential within machine learning framework in highly-vectorizable form and further enhanced the potentials with physical constraints. Machine learning can significantly boost their performances with few training data.Additional comments including restrictions and unusual features: This program needs TensorFlow 1.14.*. Neither newer or older TensorFlow is supported.