About this Abstract |
| Meeting |
MS&T25: Materials Science & Technology
|
| Symposium
|
High-Entropy Materials: Solid Solutions, Intermetallics, Ceramics, Functional Materials and Beyond VI
|
| Presentation Title |
A Methodological Study on Data Representation for Machine Learning Modelling of Thermal Conductivity of Rare-Earth Oxides |
| Author(s) |
Amiya Chowdhury, Acacio Rincon Romero, Halar Memon, Eduardo Aguilar-Bejarano, Tanvir Hussain, Grazziela Figueredo |
| On-Site Speaker (Planned) |
Amiya Chowdhury |
| Abstract Scope |
Quantitative structure-activity relationship (QSAR) modeling accelerates property prediction and materials discovery in thermal barrier coatings (TBC), where experimental cycles can take years. The effectiveness of machine learning in QSAR depends heavily on data quality and material representation. Traditional hand-crafted descriptors are constrained by crystal structure, limiting dataset diversity. In contrast, graph neural networks, such as the Crystal Graph Convolutional Neural Network (CGCNN), encode atomic positions and bonds allowing for a more detailed and diverse dataset. This study compares CGCNN with Random Forest (RF) and Gaussian Process (GP) models trained on literature-based descriptors, using a dataset of high-entropy, rare-earth pyrochlore oxides. Two data augmentation methods are tested to address the limited dataset size, including one unique to graph-based models. Results show that CGCNN substantially outperforms RF and GP, showcasing the promise of graph-based representations for improved QSAR modeling in TBC research. |