Δ-model correction of foundation model based on the model's own understanding.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Universal interatomic potentials, or foundation models, offer broad applicability but may require adjustments for specific material subclasses.
• Existing models like CHGNet, while powerful, can exhibit limitations when encountering data scarcity in training sets, particularly for novel or underrepresented atomic environments.

Purpose of the Study:

• To develop and validate a Δ-learning augmentation scheme for refining universal interatomic potentials.
• To address the limitations of foundation models in accurately describing specific material systems, such as ultra-thin oxide films and metal-sulfur interfaces.

Main Methods:

• Implementation of a Gaussian Process Regression (GPR) based Δ-model for residual corrections.
• Exploration of various aggregation strategies (global, species-separated, atomic) for representation vectors within the Δ-model.
• Application and evaluation of the augmented CHGNet model on copper oxide and sulfur adatom systems.

Main Results:

• The augmented CHGNet model accurately predicted the energetics of the "8" Cu oxide, outperforming previous density functional theory-based predictions.
• The Δ-model effectively corrected errors in CHGNet's description of sulfur adatom overlayers on Cu(111), Ag(111), and Au(111).
• The need for corrections was linked to the scarcity of metal-sulfur environments in the training data, causing overreliance on sulfur-sulfur interactions.

Conclusions:

• Δ-learning provides an efficient method for augmenting universal interatomic potentials, enhancing their predictive accuracy for specific material subclasses.
• The GPR-based Δ-model, utilizing the inherent representation of foundation models, offers a generalizable approach to address data limitations.
• This work highlights the necessity of augmentation schemes for robust application of universal potential models across diverse materials.