Identifying Split Vacancy Defects with Machine-Learned Foundation Models and Electrostatics -- APS March 2025 Contributed Talk

'Identifying Split Vacancy Defects with Machine-Learned Foundation Models and Electrostatics' talk at APS March 2025 in Anaheim.

Presentation Recording: https://youtu.be/n1Vy0m_o-L0

Google Scholar for References: https://scholar.google.com/citations?user=P-7ICrQAAAAJ&hl=en
Website: https://seankavanagh.com

Transcript

1. Evaluating equivariant neural network force fields for point defects Seán R. Kavanagh, Environmental Fellow @ Harvard Host: Prof Boris Kozinsky
2. Evaluating equivariant neural network force fields for point defects Seán R. Kavanagh, Environmental Fellow @ Harvard Host: Prof Boris Kozinsky
3. Session MAR-C49; 260C
4. Defects: Flash Introduction
5. Defects: Flash Introduction “Vacancy”
6. Defects: Flash Introduction
7. Defects: Flash Introduction “Substitution”
8. Defects: Flash Introduction
9. Defects: Flash Introduction “Interstitial”
10. 10 Quantum Defects Transparent Conducting Materials Transistors Thermoelectrics Photovoltaics & Photocatalysts Also LEDs, nuclear cladding, photo/electro-catalysis, radiation detectors, gas sensing, space materials… Wolfowicz et al. Nat Rev Mater 2021 Batteries Defects: Flash Introduction
11. 22 - (+ finite-size and chemical potential corrections) Defect supercell Bulk supercell = ΔH d Defect Calculation Workflow !! = #exp −Δ) *" +
12. Standard defect supercell relaxation Exploring the Energy Landscape Mosquera-Lois & Kavanagh*, Matter 2021
13. Exploring the Energy Landscape Mosquera-Lois & Kavanagh* et al. npj Comp Mater 2023 Mosquera-Lois & Kavanagh* et al. J. Open Source Software 2022 Defect structure-searching strategies: • ShakeNBreak: • Chemically-guided bond distortions & constrained rattling. • Accuracy rate >95% with known metastabilities. • Significant defect reconstructions identified in >30 studies • CMA-ES: Arrigoni & Madsen npj Comput Mater 2021 • DASP: Huang et al. J Semiconductors 2022 • AIRSS: Morris Phys Rev B 2008
14. Case Study: Split Vacancies 27 VX → Xi + 2 VX (‘Split-vacancy’) Varley et al J. Phys.: Condens. Matter 2011 Kononov et al. J. Phys.: Condens. Matter 2023 Li, Kavanagh et al. Chem Mater 2024 Kavanagh* arXiv (JPhys Energy Emerging Leaders) 2025 VGa in Ga2O3
15. Case Study: Split Vacancies 28 VX → Xi + 2 VX (‘Split-vacancy’) Varley et al J. Phys.: Condens. Matter 2011 Kononov et al. J. Phys.: Condens. Matter 2023 Li, Kavanagh et al. Chem Mater 2024 Kavanagh* arXiv (JPhys Energy Emerging Leaders) 2025 Displacement vs Distance to Defect
16. Case Study: Split Vacancies = Lower energy split vacancy VGa in Ga2O3; ΔE ~ 1 eV Electrostatic energy model: (i.e. Ewald summation)
17. Case Study: Split Vacancies 32 Kavanagh* arXiv (JPhys Energy Emerging Leaders) 2025 1 Fowler et al. J Appl Phys 2024 (c.f. ~10 previously- known cases)1 Kumagai et al. Phys Rev Mater 2021
18. “Exhaustive” = All ML-predicted split vacancies with energies within 0.35 eV of simple vacancy are candidates >2 orders of magnitude reduction of DFT calculations (“discovery acceleration factor”; DAF) Split Vacancies: Electrostatics & MLIPs Kavanagh* arXiv (JPhys Energy Emerging Leaders) 2025 MLIPs (Prevalence ~ 10%)
19. “Exhaustive” = All ML-predicted split vacancies with energies within 0.35 eV of simple vacancy are candidates >2 orders of magnitude reduction of DFT calculations (“discovery acceleration factor”; DAF) Split Vacancies: Electrostatics & MLIPs Kavanagh* arXiv (JPhys Energy Emerging Leaders) 2025 MLIPs (Prevalence ~ 10%) Wednesday, 260B Session MAR-L50
20. Split Vacancies: Electrostatics & MLIPs 35 Kavanagh* arXiv (JPhys Energy Emerging Leaders) 2025 (c.f. ~10 previously- known cases)1 1 Fowler et al. J Appl Phys 2024 Test Set: All known inorganic crystalline solids (ICSD), and predicted metastable materials on the Materials Project database.
21. Split Vacancies: Electrostatics & MLIPs 36 Kavanagh* arXiv (JPhys Energy Emerging Leaders) 2025 (c.f. ~10 previously- known cases)1 1 Fowler et al. J Appl Phys 2024
22. Split Vacancies: Conclusions • Far more common than known (~10 previously-known, now >10,000). • ~10% prevalence rate for cation vacancies in oxides/nitrides • Particularly common in ionic compounds with strong electrostatic driving forces. • Playing a crucial role in ion migration and doping. • Demonstration of the utility of foundation machine-learning potentials – with important caveats… Prevalence doped.readthedocs.io shakenbreak.readthedocs.io ShakeNBreak w/M3GNet: Mosquera-Lois, Kavanagh, Ganose, Walsh. npj Comput Mater 2024 ✉ skavanagh@seas.harvard.edu
23. Session MAR-C49; 260C ✉ skavanagh@seas.harvard.edu
24. doped: Calculating Defects in Solids Kavanagh* et al. J Open Source Softw 2024
25. Conclusion doped.readthedocs.io shakenbreak.readthedocs.io Prof. Boris Kozinsky Fellowship (Harvard): Thank you for listening!