Daniel Karls successfully defended his Ph.D. thesis at the University of Minnesota on the "Transferability of Empirical Potentials and the Knowledgebase of Interatomic Models (KIM)." Dr. Karls'' dissertation provides an introduction to interatomic models from a new perspective and a discussion of machine learning methods for assessing their transferability, i.e. the ability of a model to describe behavior outside its training set.'
From the abstract:
"Empirical potentials have proven to be an indispensable tool in understanding complex material behavior at the atomic scale due to their unrivaled computational efficiency. However, as they are currently used in the materials community, the realization of their full utility is stifled by a number of implementational difficulties. An emerging project specifically aimed to address these problems is the Knowledgebase of Interatomic Models (KIM). The primary purpose of KIM is to serve as an open-source, publically accessible repository of standardized implementations of empirical potentials (Models), simulation codes which use them to compute material properties (Tests), and first-principles/experimental data corresponding to these properties (Reference Data). Aside from eliminating the redundant expenditure of scientific resources and the irreproducibility of results computed using empirical potentials, a unique benefit offered by KIM is the ability to gain a further understanding of a Model's transferability, i.e. its ability to make accurate predictions for material properties which it was not fitted to reproduce. In the present work, we begin by surveying the various classes of mathematical representations of atomic environments which are used to define empirical potentials. We then proceed to offer a broad characterization of empirical potentials in the context of machine learning which reveals three distinct categories with which any potential may be associated. Combining one of the aforementioned representations of atomic environments with a suitable regression technique, we define the Regression Algorithm for Transferability Estimation (RATE), which permits a quantitative estimation of the transferability of an arbitrary potential. Finally, we demonstrate the application of RATE to a specific training set consisting of bulk structures, clusters, surfaces, and nanostructures of silicon. A specific analysis of the underlying quantities inferred by RATE which are used to characterize transferability is provided."
The full dissertation is here.