Dresden 2017 – wissenschaftliches Programm
MM 68: Electronic Structure Theory: New Concepts and Developments in Density Functional Theory and Beyond - VII
MM 68.9: Vortrag
Donnerstag, 23. März 2017, 18:15–18:30, GER 38
Cross-validation in the cluster expansion method — •Axel Hübner, Santiago Rigamonti, and Claudia Draxl — Humboldt-Universität zu Berlin
The cluster expansion technique allows the construction of model Hamiltonians for an efficient evaluation of the total energy of alloys. This technique relies on a fit to a small set of ab-initio calculations for selected atomic configurations. Key aspects to maximize the predictive performance of the model are the selection of a set of basis functions, i.e. clusters, and of configurations. To achieve this, the cross-validation technique is typically used . In this work, an analytical formula for the calculation of the leave-many-out cross-validation score (CV) is derived. This formula exhibits numerical instabilities, whose analytical properties yield a criterion for structure selection in cluster expansions. Moreover, a relation between the noise in the data and the CV is outlined. This leads to a tool which allows us to estimate, for a given noise level, the size of the ab-initio data set upon which no improvements of the model are obtained. These results are exemplified for a cluster expansion of the thermoelectric clathrate alloy Ba8AlxSi46−x, calculated with the CELL package .
 A. van d. Walle et al., Journal of Phase Equilibria 23 (2002), Aug., Nr. 4
 M. Troppenz et al., submitted (2016); S. Rigamonti et al., in preparation.