Machine Learning Compared with DFT: Streamlining Boron-Cluster Simulations

Machine Learning Compared with DFT: Streamlining Boron-Cluster Simulations

Boron Nanoclusters

Mr. Daniel Lowe

Dr. Bora Karasulu

Boron nanoclusters can form planar, cage-like, open and compact structures because electron-deficient multicentre bonding produces many low-energy arrangements. Identifying stable structures means the use of a large number of geometry optimisations, making direct screening with density functional theory computationally expensive. In this project, 125 AIRSS-generated candidates for each cluster size from B₁ to B₂₀₀ were independently optimised using the 7NET, MACE and PET-MAD machine-learning interatomic potentials, giving 75,000 optimisation attempts. The lowest-energy converged structure from each model and cluster size was selected, producing 600 structures for analysis. PET-MAD achieved the highest overall convergence at 99.83%, followed by 7NET at 99.44% and MACE at 96.92%. 7NET generally produced the most compact, highly coordinated structures, while PET-MAD gave the closest qualitative comparison with an independent VASP-DFT B₅₃ reference. These results show that MLIPs can efficiently screen boron nanoclusters, although their predictions remain model-dependent and require matched DFT calculations for rigorous validation.

Chemistry, Computational, Boron