Lecture is hosted by Penn State Professor Zi-Kui Liu, Corning Faculty Fellowship in Materials Science and Engineering
"Computational Tools for High Temperature Materials: From GNNs to Entropy as Information"
Qi-Jun Hong, Assistant Professor, Materials Science and Engineering Arizona State University
Abstract: We present a unified toolkit for predicting high-temperature materials properties that couples physics-based graph neural networks, the Materials Property Prediction (MAPP) framework, with rigorous first-principles simulation, the SLUSCHI package. On the machine-learning side, MAPP represents chemical formulas as graphs and learns formula-to-property mappings with uncertainty awareness, enabling rapid screening across broad chemical spaces. On the ab initio side, SLUSCHI orchestrates density-functional molecular dynamics and free-energy workflows, with current support for VASP, to compute melting points, disordering entropies, and phase stability.
A key advance is an information-based route that treats thermodynamic entropy as information content obtained from a single MD trajectory. We partition entropy into configurational and vibrational parts by (i) evaluating š¯‘†conf as the information entropy of local neighbor environments and (ii) extracting š¯‘†vib from phonon densities of states via velocityā€“autocorrelation, explicitly avoiding double counting while including all relevant states. In parallel, we introduce a parameter-free lossy compression of atomic trajectories whose compressed description length realizes the equivalence between information (Shannon) entropy and thermodynamic (Gibbs) entropy in this setting. Together, these ideas deliver accurate transition temperatures (e.g., melting) with large cost reductions relative to conventional multi-trajectory methods and quantitative agreement with gold-standard benchmarks.
The seminar will showcase end-to-end case studiesā€”such as large-scale melting predictions for thousands of minerals and identification of extreme high-melting compoundsā€”highlighting how MAPP and SLUSCHI interoperate to accelerate property prediction, phase-diagram construction, and CALPHAD modeling.
Bio: Qi-Jun Hong is an assistant professor of materials science and engineering at Arizona State University. Hong received his doctoral degree in chemistry from the California Institute of Technology in 2015. Before joining ASU, Hong was a post-doctoral research associate at Brown University and then an applied scientist in machine learning at Amazon.com, Inc.