Generalized Machine Learning Potential Models for Elemental Nanoclusters

Abstract Nanoclusters occupy a unique size regime between isolated atoms and bulk materials. Their electronic, structural, and thermodynamic properties are governed by quantum effects and complex many-body interactions. Accurately modeling their potential energy surfaces (PES) is essential for understanding their behavior and applications in areas such as catalysis, nanoelectronics, and energy storage. In this work, we develop Gaussian Approximation Potential (GAP) models to describe a comprehensive 54 elemental nanoclusters across the periodic table. GAP, a machine learning interatomic potential (MLIP), learns the PES directly from ab-initio data and has proven effective in accurately modeling systems with low symmetry, structural diversity, and complex energetics. Our approach utilizes a diverse training and test dataset of over 170,000 nanocluster configurations obtained with targeted sampling strategies and high-fidelity density functional theory (DFT) calculations. The GAP models are rigorously benchmarked against DFT, demonstrating strong agreement in energy and force predictions, robust structural and dynamical performance across cluster sizes and chemistries, and accurate reproduction of phase and dynamic behavior. We further assess the generalization of the model in both the cluster and the bulk regimes via performance analysis of their structural properties. Despite the wide structural diversity in the dataset, the framework achieves high accuracy and transferability by combining structural weighting with a Bayesian training approach. Our work establishes a comprehensive testbed of MLIPs for low-dimensional systems across a wide chemical space spanning s-, p-, and d-block elements, offering a path toward universal potentials with ab-initio fidelity.