Exact expressions for structure selection in cluster expansions

Abstract

The cluster expansion has proven to be a valuable tool in materials science to predict properties of configurationally ordered and disordered structures but the generation of cluster expansions can be computationally expensive. In recent years there have been efforts to make the generation of cluster expansions more efficient by selecting training structures in a way that minimizes approximate expressions for the variance of the predicted property values. We demonstrate that in many cases, these approximations are not necessary and exact expressions for the variance of the predicted property values may be derived. To illustrate this result, we present examples based on common applications of the cluster expansion such as bulk binary alloys. In addition we extend these structure selection techniques to Bayesian cluster expansions. These results should enable researchers to better analyze the quality of existing training sets and to select training structures that yield cluster expansions with lower prediction error.

Compressive sensing as a paradigm for building physics models

“…Furthermore, CE is a stringent test case for compressive sensing because a significant amount of effort has been expended developing advanced model building techniques, which have been implemented in sophisticated general-purpose computer codes. [15][16][17][21][22][23][24][25] …”

Approximate solutions to the cluster variation free energies by the variable basis cluster expansion

“…Based on this equivalence, it has been noted that, in carrying out cluster expansions, it is generally more convenient to use the SDG basis introduced in Ref. [2] than the concentration-dependent basis [6]. However, while the equivalence argument is valid for the full cluster expansion, for both finite clusters and infinite lattices, it breaks down if the expansion is truncated at any level.…”

Ab initiodetermination of structure-property relationships in alloy nanoparticles

“…The remaining 129 structures were low-energy structures at different compositions predicted by the cluster expansion. Due to the importance of the end compositions on the quality of the cluster expansion, 21 the pure Au and pure Pd nanoparticles were included twice in the training set in a way that ensured that during cross-validation the cluster expansion was always fit to a set of structures that included both pure Au and pure Pd.…”