RAPID ANALYSIS OF COMBINATORIAL LIBRARY DATA AND INTEGRATION WITH DATABASES: REGRESSION AND CLASSIFICATION ALGORITHMS FOR AUTOMATED LEARNING IN EXPERIMENTAL AND SIMULATED COMPOSITION SPREAD DATA

摘要

We are developing a variety of ways to rapidly analyze large amount of data amassed from combinatorial libraries and methods to integrate the experimental data with databases. A particular emphasis is placed on mapping the structural properties across libraries and quickly cross-referencing them with known crystallographic databases. The need for such a study is common among combinatorial investigation of virtually all topics. In particular, we are investigating the use of machine learning algorithms for material property prediction and material classification. These algorithms allow for data mining automation in analyzing large material databases such as the ICSD crystallographic database. The algorithms also provide a means for determining latent variables -primary predictors of material structure and properties that might otherwise be buried in the large datasets. Two powerful algorithms were investigated for their performance: Relevance Vector Machines for regression analysis and mean shift theory for material classification. Furthermore, a new correlation-length based algorithm was developed for material classification.