Statistical learning for alloy design from electronic structure calculations

The objective of this thesis is to explore how statistical learning methods can contribute to the interpretation and efficacy of electronic structure calculations. This study develops new applications of statistical learning and data mining methods to both semi-empirical and density functional theory (DFT) calculations. Each of these classes of electronic structure calculations serves as templates for different data driven discovery strategies for materials science applications. In our study of semi-empirical methods, we take advantage of the ability of data mining methods to quantitatively assess high dimensional parameterization schemes. The impact of this work includes the development of accelerated computational schemes for developing reduced order models. Another application is the use of these informatics based techniques to serve as a means for estimating parameters when data for such calculations are not available.;Using density of states (DOS) spectra derived from DFT calculations we have demonstrated the classification power of singular value decomposition methods to accurately develop structural and stoichiometric classifications of compounds. Building on this work we have extended this analytical strategy to apply the predictive capacity of informatics methods to develop a new and far more robust modeling approach for DOS spectra, addressing an issue that has gone relatively unchallenged over two decades. By exploring a diverse array of materials systems (metals, ceramics, different crystal structures) this work has laid the foundations for expanding the linkages between statistical learning and statistical thermodynamics. The results of this work provide exciting new opportunities in computational based design of materials that have not been explored before.