ITR: Data Mining of Quantum Mechanical Calculations for Predicting Materials Structure

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

This award was made on a 'small' category proposal submitted in response to the ITR solicitation, NSF-02-168. It supports computational research and education on using data-mining techniques on data obtained using ab-initio methods to predict crystal structures of new materials. Ab-initio methods are becoming ubiquitous tools for physicists, chemists, and materials scientists. These methods allow scientists to evaluate and pre-screen new materials "in silico", rather than through time-consuming experimentation. A current limitation of ab-initio computation is that the method does not accumulate experience or knowledge (except for an increase in the skills of the scientist). For example, when calculating how the stability of alloys changes as a function of temperature and composition, each new system is treated independently of results one may have obtained previously on other systems. The goal of this work is explore a radically different approach, which uses data mining methods to inform new ab-initio investigations with knowledge obtained from results already collected on other systems. The objective of this research is to demonstrate quantifiable knowledge extraction from a large number of ab-initio calculations, and to use this knowledge in the prediction of crystal structure. The ab-initio calculations will be carried out using accurate and well-established techniques of density functional theory. Knowledge extraction techniques will be borrowed from the burgeoning world of data mining. These techniques have found growing applications in industry, e-commerce, and the social, chemical, and biological sciences. The research will initially focus around linear techniques, for example, multivariate regression methods like Principal Component Analysis and Partial Least Squares, and then include non-linear approaches using neural networks and clustering algorithms, as well as make use of more established techniques, such the cluster expansion. The web will facilitate this work by making it possible to gather data from the entire ab-initio community, and by providing a central public resource where data and data mining tools can be tested and stored. New developments gained from this research, and the ab-initio database that will be created, will be integrated with the teaching activities of the PI in computational materials modeling. Use in the classroom and computational laboratory will particularly assist students in learning the relationships between structure and energetics of materials. The PI's approach may have substantial impact on materials research and design. The successful completion of this research project may lead to a reliable method for determining the stable structure of a material, and to the creation of a public database of ab-initio calculated energies for a large number of crystal structures and alloys that can be queried by ab-initio practitioners, experimentalist researchers, students, and materials educators. %%% This award was made on a 'small' category proposal submitted in response to the ITR solicitation, NSF-02-168. It supports computational research and education on using data-mining techniques on data obtained using ab-initio methods to predict crystal structures of new materials. Predicting crystal structures for new alloys knowing only the identity of the constituent atoms is a long standing and fundamental challenge in materials science, and a major impediment to effective first-principles materials design. The objective of this research is to demonstrate quantifiable knowledge extraction from a large number of density-functional-theory based computations, and to use this knowledge in the prediction of crystal structure. Knowledge extraction techniques that have been applied in industry, e-commerce, and the social, chemical, and biological sciences, will be borrowed from the world of data mining. The web will facilitate this work by making it possible to gather data from the community, and by providing a central public resource where data and data mining tools can be tested and stored. New developments gained from this research, and the database that will be created, will be integrated with the teaching activities of the PI in computational materials modeling. Use in the classroom and computational laboratory will particularly assist students in learning the relationships between structure and energetics of materials. The PI's approach may have substantial impact on materials research and design. This project may lead to a reliable method for determining the stable structure of a material, and to the creation of a public database of calculated energies for a large number of crystal structures and alloys that can be queried by theoretical materials scientists, experimentalist researchers, students, and educators. ***