| About this Abstract |
| Meeting |
2024 TMS Annual Meeting & Exhibition
|
| Symposium
|
AI/Data Informatics: Computational Model Development, Verification, Validation, and Uncertainty Quantification
|
| Presentation Title |
Harnessing Equivariant Neural Networks for High-throughput Screening of Novel Superconductors |
| Author(s) |
Jason Gibson, Ajinkya Hire, Oscar Barrera, Philip Dee, Benjamin Geisler, Peter Hirscheld, Richard Hennig |
| On-Site Speaker (Planned) |
Jason Gibson |
| Abstract Scope |
A significant barrier in high-throughput screening of superconductors lies in the prohibitive computational cost of determining the Eliashberg spectral function (α2F). Although previous studies have employed machine learning models to bypass costly calculations, the lack of a sufficiently large and accurate database has hindered advances in α2F prediction. To address this, we computed 1,400 accurate α2F's, creating a robust database on which we trained variants of equivariant neural networks. These models, while offering varying levels of accuracy and cost-efficiency, incorporate computed properties into the node embeddings. Our most efficient model, requiring only structural information, yields a moderate Pearson correlation coefficient on predicted α2F-derived properties λ, ωlog, and ω2. While, our most accurate model, which necessitates moderately expensive computations, obtains a high Pearson correlation coefficient on the test set. In the final phase of our study, we undertook a high-throughput screening of the MaterialsProject database to identify potential high Tc superconductors. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Modeling and Simulation, |