About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Hume-Rothery Symposium: Data-Driven Materials Discovery and Phase Stability
|
| Presentation Title |
Elemental Features for Data-Driven Materials Properties Prediction |
| Author(s) |
Dane Morgan |
| On-Site Speaker (Planned) |
Dane Morgan |
| Abstract Scope |
This talk shares some of our work that builds on Chris Wolverton's seminal Magpie elemental descriptors approach. This approach provides a general framework for developing machine learning model descriptors for any compound based on its composition. I will discuss how composition-based elemental features can be accurate and enable rapid and broad development of property models. I will also show that they are often as good as hand-engineered descriptors. Finally, I will share how we can use calibrated ensemble and kernel density approaches to get robust uncertainty and domain-of-applicability estimates for elemental feature based models. Applications will include a range of materials properties and integration with CALPHAD modeling to allow for greater generalization than traditional approaches. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Computational Materials Science & Engineering, Modeling and Simulation |