About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
Deep Learning on DFT-Derived Charge Densities for Stacking Fault Energy Predictions in High Entropy Alloys |
| Author(s) |
Ali Azarmi, Pouya Abasmirzaie, Nathan Linton, Serveh Kamrava, Dilpuneet S Aidhy |
| On-Site Speaker (Planned) |
Ali Azarmi |
| Abstract Scope |
We use charge density, derived from density functional theory (DFT) calculations, to develop a hybrid CNN-MLP neural network architecture to predict stacking fault energies in face centered cubic (FCC) high entropy alloys (HEAs). The model integrates two parallel processing branches: a Convolutional Neural Network (CNN) to extract hierarchical spatial features from charge density maps, and a Multilayer Perceptron (MLP) to capture explicit composition-property relationships. The outputs are fused to learn complex interactions between spatial electronic structure and global compositional dependencies. The hybrid model achieves exceptional accuracy with an Rē of 0.983 and an RMSE of 10.09 mJ/mē on test data spanning binary, ternary, and quaternary Ni-Fe-Cr-Co alloys. Crucially, the model demonstrates a remarkably consistent normalized root mean squared error (NRMSE of 0.02) across all alloy complexities, indicating robust generalization from simple to complex systems. This work shows that charge density can be reliably used to predict SFEs in complex alloys. |
| Proceedings Inclusion? |
Planned: |