Active Δ-learning with universal potentials for global structure optimization.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Universal machine learning interatomic potentials (uMLIPs) offer broad applicability but may require further data for out-of-sample refinement.
• Global optimization of material structures is crucial for discovering novel properties and functionalities.

Purpose of the Study:

• To develop and evaluate an active learning scheme for improving uMLIPs in global optimization tasks.
• To compare the performance of this active learning approach against various global optimization algorithms and foundation models.

Main Methods:

• Implemented an active learning strategy augmenting foundation models with a Δ-model using Gaussian process regression (GPR) and SOAP descriptors.
• Evaluated the scheme using silver-sulfur clusters and surface reconstructions on Ag(111) and Ag(100).
• Compared performance against random structure search, basin hopping, GOFEE, and replica exchange (REX).

Main Results:

• Active learning with GPR-based Δ-models demonstrated robustness in identifying global minima.
• Replica exchange (REX) proved most efficient in terms of total CPU time.
• CHGNet, MACE-MP0, and MACE-MPA were compared as foundation models.

Conclusions:

• Active learning combined with GPR-based Δ-models is a reliable method for enhancing uMLIPs in global structure prediction.
• While active learning excels at accuracy, REX offers superior computational efficiency for global optimization tasks.