Date of Award

1-21-2022

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Chemistry

First Advisor

Jerome Darsey

Second Advisor

Jeffery Moran

Abstract

Bulk materials made from transition elements and their alloys are substances used in many industrial applications and are composed of elements contained in the d- and f-blocks of the periodic table. Many of the useful properties of these materials are related to their unfilled d- and f-orbitals. It is supposed that computational models of orbitals describing two-atom homogeneous clusters of transition elements provide enough information to predict bulk chemical and physical properties of transition materials. This thesis reports the use of a simple (one hidden node) three layer feed-forward/back-propagating artificial neural network and orbital energies obtained from DFT-SCF/MO approximations to form highly predictive models of chemical and physical properties. This work identifies three highly predictive (q2 = 0.91) artificial neural network models, (3d1−9), (3d1−9, 4d1−7,9), and (3d1−9, 4d1−7,9, 5d2−9), that accurately correlate orbital energies of two-atom transition clusters with 1st ionization potentials. Many elements (Y, V, Cr, Ti, Zr, Nb, Tc, Ru, Mn, Rh, Ni, Ag, Cu, Co, Fe, Re, Pt, Ir, and Au) demonstrate a tendency to be predicted with a high degree of accuracy (±0.04% to ±0.98% error). However, these models do not accurately (q2 < 0.90) correlate melting point or electron affinity with orbital energies, even though there is a linear relationship (q2 = 0.50 and 0.47 for melting point and electron affinity, respectively). Results demonstrate that a simple neural network model can be useful for predicting bulk chemical and physical properties. Using orbital energies as the input variable simplifies calculations and allows the neural network to be quickly trained to recognize relationships which can be used to predict properties of new materials.

Download

Included in

Chemistry Commons

Share

COinS