While Density Functional Theory (DFT) provides the necessary accuracy to describe realistic systems, its computational cost limits its applicability to medium-sized systems and short timescales. Machine Learning Interatomic Potentials (MLIPs) offer an attractive alternative, enabling large-scale simulations with near ab initio accuracy [1, 2]. However, their reliability critically depends on their ability to generalize across structurally diverse and out-of-distribution environments. This limitation becomes particularly relevant in scenarios involving non-equilibrium processes and complex structural transformations, where local atomic environments can vary significantly beyond those represented in the training data. The recent experimental observation of an atomically thin nanowire of the ionic ceramic ZrO2 under tensile stress, revealed by high-resolution transmission electron microscopy (HRTEM) [3], provides a compelling example of such a challenge. This process involves a dynamic transformation pathway spanning multiple structural regimes, from bulk-like environments to highly undercoordinated nanostructures. Capturing these mechanisms requires models capable of maintaining accuracy across a wide range of coordination environments and dimensionalities. In this work, we use the simulation of ZrO2 nanowire formation as a testbed to systematically assess and guide the transferability of MLIPs. We evaluate model performance across multiple structural regimes relevant to the transformation process, including bulk, surfaces, intermediate neck configurations, and atomically thin wires. Three modeling strategies are compared: (i) zero-shot application of universal MLIPs, (ii) fine-tuning of pretrained models using system-specific data, and (iii) training specialized models from scratch. Our results show that transferability, rather than raw accuracy on bulk-like structures, is the key factor enabling reliable simulation of complex structural transformations. Universal models exhibit limited performance when extrapolated to undercoordinated and highly strained configurations, while fine-tuned models significantly improve predictive accuracy but remain sensitive to the diversity of the training data. Models trained from scratch achieve high accuracy within specific domains but lack generalization across geometries. By explicitly linking simulation performance to structural diversity, this work provides practical guidelines for selecting and adapting MLIP strategies in complex atomistic simulations. More broadly, we establish a simulation-driven framework to evaluate and improve MLIP generalization, enabling more reliable modeling of non-equilibrium processes and low-dimensional ceramic nanostructures.
References:
[1] Volker L. Deringer, Miguel A. Caro, and Gábor Csányi. Machine learning interatomic potentials as emerging tools for materials science. Advanced Materials, 31(46), September 2019.
[2] Bruno Focassio, Luis Paulo M. Freitas, and Gabriel R. Schleder. Performance assessment of universal machine learning interatomic potentials: Challenges and directions for materials’ surfaces. ACS Applied Materials & Interfaces, 17(9):13111–13121, July 2024.
[3] Bruno Focassio, Tanna E.R. Fiuza, Jefferson Bettini, Gabriel R. Schleder, Murillo H.M. Rodrigues, João B.S. Junior, Edson R. Leite, Adalberto Fazzio, and Rodrigo B. Capaz. Stability and rupture of an ultrathin ionic wire. Physical Review Letters, 129(4), July 2022.
Comissão Organizadora
Pedro Alves da Silva Autreto
Comissão Científica
Ferramenta completa de submissão, avaliação e publicação de anais para eventos científicos.
