Pretrained machine-learned interatomic potentials (MLIPs) lack tailored accuracy for specific material systems like molybdenum-oxygen (Mo-O) interactions. Herein, we develop a fine-tuned MLIP by optimizing the DPA-2 model with around one thousand data points from density functional theory (DFT) calculations to study O adsorption on Mo surfaces. The fine-tuned model achieves DFT-level accuracy in predicting key properties of Mo-O systems, including lattice constants, elastic moduli, and formation energies of bulk Mo and Mo oxides $({\mathrm{MoO}}_2$, ${\mathrm{MoO}}_3)$, as well as surface energies and O adsorption energies on Mo low-index surfaces across various coverages and sites. In contrast, the zero-shot DPA-2 model without fine-tuning exhibits substantial errors in complex nonbulk configurations, with errors in predicted energies, atomic forces, and virial tensors that are up to two orders of magnitude larger than those of the fine-tuned model when benchmarked against DFT results. Fine-tuning substantially reduces deviations from DFT, demonstrating that universal MLIPs can be effectively enhanced with small specialized datasets at low computational cost. This work provides a reliable tool for Mo-O surface simulations and offers insights for fine-tuning MLIPs in targeted material research.

Close

PrevNext

View figure in article

• Density functional theory
• First-principles calculations
• Machine learning