Augmentation of Universal Potentials for Broad Applications.

Area of Science:

• Computational Materials Science
• Surface Science
• Machine Learning in Chemistry

Background:

• Density Functional Theory (DFT) offers high accuracy for materials simulations but is computationally expensive.
• Universal potentials, like CHGNet, aim to reduce the computational cost of DFT-level calculations.
• Pretrained universal potentials may not generalize well to systems outside their training data.

Purpose of the Study:

• To evaluate the performance of pretrained universal potentials for systems beyond their training data.
• To improve the accuracy of universal potentials through fine-tuning and Δ-learning approaches.
• To investigate experimentally observed defects in the Ag(111)-O surface reconstruction.

Main Methods:

• Utilized the pretrained CHGNet universal potential.
• Applied fine-tuning and Δ-learning techniques to augment potential performance.
• Investigated defect formation mechanisms in the Ag(111)-O surface reconstruction.

Main Results:

• Pretrained CHGNet showed potential for out-of-the-box success but also significant failures in predicting ground state configurations.
• Fine-tuning and Δ-learning approaches successfully augmented the performance of universal potentials for specific cluster and surface systems.
• The study successfully investigated and explained experimentally observed defects in the Ag(111)-O surface reconstruction.

Conclusions:

• Universal potentials can be effectively improved for specific applications through targeted fine-tuning.
• The enhanced universal potentials provide a cost-effective method for studying complex surface phenomena.
• This work elucidates the mechanics behind experimentally observed defects in the Ag(111)-O surface reconstruction.