A Set of Moment Tensor Potentials for Zirconium with Increasing Complexity.

Area of Science:

• Materials Science
• Computational Chemistry
• Physics

Background:

• Machine learning force fields (MLFFs) offer high fidelity and continuous improvement for atomistic simulations.
• Developing accurate MLFFs requires extensive quantum-mechanical (QM) data and careful model construction.

Purpose of the Study:

• To propose a hybrid small-cell approach combining offline and active learning for systematic MLFF development.
• To construct and evaluate MLFFs with increasing complexity using the moment tensor potential formalism.
• To assess the predictive accuracy of MLFFs for diverse material properties of Zr.

Main Methods:

• Implemented a hybrid small-cell approach integrating offline and active learning strategies.
• Systematically expanded a QM database to train MLFFs with incrementally increasing model complexity.
• Employed the moment tensor potential formalism for MLFF construction.
• Quantitatively evaluated structural, elastic, thermal, and defect properties of Zr.

Main Results:

• Model complexity positively correlates with prediction accuracy for MLFFs.
• MLFFs accurately predicted properties of unseen configurations not present in the training data.
• High-complexity MLFFs (1513 parameters) showed good agreement with DFT benchmarks, though subtle features were sometimes obscured by noise.

Conclusions:

• The hybrid MLFF approach enables systematic database expansion and model complexity increase.
• MLFFs demonstrate strong predictive capabilities for various material properties, even for out-of-sample data.
• While high complexity improves accuracy, careful consideration of training data and potential noise is necessary for capturing fine physical details.