| dc.description.abstract | In materials science, the traditional methodological framework is the
identification of the composition-processing-structure-property causal pathways
that link hierarchical structure to properties. However, all the properties of
materials can be derived ultimately from structure and bonding, and so the
properties of a material are interrelated to varying degrees.
The work presented in this thesis, employed artificial neural networks (ANNs) to
explore the correlations of different material properties with several examples in
different fields. Those including 1) to verify and quantify known correlations
between physical parameters and solid solubility of alloy systems, which were
first discovered by Hume-Rothery in the 1930s. 2) To explore unknown crossproperty
correlations without investigating complicated structure-property
relationships, which is exemplified by i) predicting structural stability of
perovskites from bond-valence based tolerance factors tBV, and predicting
formability of perovskites by using A-O and B-O bond distances; ii) correlating
polarizability with other properties, such as first ionization potential, melting
point, heat of vaporization and specific heat capacity. 3) In the process of
discovering unanticipated relationships between combination of properties of
materials, ANNs were also found to be useful for highlighting unusual data
points in handbooks, tables and databases that deserve to have their veracity
inspected. By applying this method, massive errors in handbooks were found,
and a systematic, intelligent and potentially automatic method to detect errors in
handbooks is thus developed.
Through presenting these four distinct examples from three aspects of ANN
capability, different ways that ANNs can contribute to progress in materials
science has been explored. These approaches are novel and deserve to be pursued
as part of the newer methodologies that are beginning to underpin material
research. | en_US |
