About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Modeling, AI Applications, and Method Development in Nuclear Materials
|
| Presentation Title |
Exploring the Extremes: Entropy-Driven Data Generation for Universal Atomistic Simulation |
| Author(s) |
Danny Perez |
| On-Site Speaker (Planned) |
Danny Perez |
| Abstract Scope |
Developing interatomic potentials that remain reliable across broad chemical, structural, and thermodynamic conditions requires training data that extend beyond the familiar neighborhoods of equilibrium materials. In this talk, I will present a chemistry-agnostic data-generation strategy based on information-entropy maximization. The method maximizes diversity in a physically motivated feature space. The result is a DFT-labeled database of nearly 1.7 million structures spanning nearly the entire periodic table. This entropy-maximized sampling philosophy provides a broad physical prior for machine-learning interatomic potentials. By deliberately populating high-diversity regions of configuration space—including, SMAX supports the construction
of potentials that are ultra-robust by design. When used to train Graph Atomic Cluster Expansion (GRACE) potentials, the resulting models exhibit state-of-the-art transferability in demanding atomistic settings while producing state-of-the-art predictions of a broad range of properties. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, Machine Learning, Modeling and Simulation |