# A06 – Ab Initio Accuracy at Large Scales by Machine Learning

**PI(s): Jörg Neugebauer, Erik Bitzek, (MPIE)**

**SFB researchers: Marvin Poul (MPIE)**

The aim of project A06 is to compute Gibbs Free Energies at defects and to construct defect phase diagrams (DPD) for systems that are too large to be treated by Density Functional Theory (DFT). Traditionally these systems are treated with classic interatomic potentials, which however often lack the accuracy to make quantitative predictions. In this project we fit machine learning interatomic potentials (MLIP) to retain the accuracy of DFT, but at larger scale.

The construction of DPDs starts with static calculations of the segregation energy which we then augment step-by-step with entropy and pressure terms to arrive at the Gibbs Free Energy as a function of chemical potential. In practice the chemical complexity is difficult to fully capture even with highly efficient MLIPs due to the sheer number of possible segregation configurations at extended defects. We develop additional surrogate models on top of MLIPs that allow to probe the configurational space in a directed manner and only calculate the segregation energy directly with MLIPs that are predicted by the surrogate model to contribute to the DPD.,This approach allows to efficiently estimate the DPD of a defect even with hundred thousand to millions of individual segregation configurations. The project collaborates closely with A02 and B01 (grain boundaries) and with C05 regarding DPDs.

**Publications: **

- [1] M. Poul, L. Huber, E. Bitzek, and J. Neugebauer, (2022) Systematic Atomic Structure Datasets for Machine Learning Potentials: Application to Defects in Magnesium. arXiv preprint arXiv:2207.04009. arxiv-export-lb.library.cornell.edu/abs/2207.04009v3