About this Abstract |
| Meeting |
MS&T25: Materials Science & Technology
|
| Symposium
|
High-Entropy Materials: Solid Solutions, Intermetallics, Ceramics, Functional Materials and Beyond VI
|
| Presentation Title |
Completely bypassing DFT calculations via graph neural networks for vacancy formation energies in high entropy alloys |
| Author(s) |
Nathan Linton, Parampreet Singh, Dilpuneet S. Aidhy |
| On-Site Speaker (Planned) |
Nathan Linton |
| Abstract Scope |
While density functional theory (DFT) calculations are the conventional means to calculate vacancy formation energies (VFEs), their applicability in high entropy alloys (HEAs) is highly restricted due to large compositional and statistical space. We propose a machine learning framework that completely bypasses DFT calculations. This is achieved by training a fine-tuned CHGNet model as an input to graph neural network models to predict VFEs. Specifically, the model is trained on Bader charges as an extra descriptor to the VFE model that significantly improves results over descriptors based on the periodic table alone. Our model predicts atomic displacements, atomic volume and Bader charges with high accuracy, all of which collectively result in accurate VFE predictions. While the model is trained on FCC Ni-Cu-Au-Pd, minor fine tuning subsequently enables accurate predictions in other structures such as Ni-Co-Cr indicating its transferability. |