About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Modeling, AI Applications, and Method Development in Nuclear Materials
|
| Presentation Title |
Benchmarking and Refining Machine-Learned Interatomic Potentials for Radiation Damage and Tritium Diffusion in LiAlO2 |
| Author(s) |
Ankit Roy, Ram Devanathan, Giridhar Nandipati, Andrew Casella, David Senor, Ayoub Soulami |
| On-Site Speaker (Planned) |
Ankit Roy |
| Abstract Scope |
Accurate interatomic potentials are essential for molecular dynamics (MD) simulations of radiation damage and gas diffusion in nuclear ceramics like LiAlO2 under extreme conditions relevant to tritium production. We benchmark six machine-learned interatomic potentials (MLIPs)—moment tensor potential(MTP), Gaussian approximation potential, deep potential for MD (DeepMD), atomic cluster expansion(ACE), message-passing ACE(MACE), pretrained), and MACE(trained from scratch), all trained on a shared density functional theory dataset including tritium. Evaluation criteria include energy accuracy, density, thermal equilibration, threshold displacement energy, tritium diffusivity, and computational cost. MTP achieves the best balance of accuracy and efficiency, while ACE and MACE models show unphysically high displacement energies, and deep potential underestimates them due to excessive repulsion. All models overestimate tritium diffusion, though pretrained MACE remains physically reasonable up to 500 K. Only MTP outperforms empirical potentials computationally. Building on these results, we apply active learning to iteratively refine MTP, yielding continuous improvement in tritium diffusion accuracy. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, Nuclear Materials, Ceramics |