Finding the right atomic structure is one of the most important and difficult problems in computational materials science. Because material properties depend strongly on how atoms are arranged, even small errors in structure prediction can affect how researchers understand stability, reactivity, and function.

A new study introduces NARA, a computational framework that combines a firefly algorithm with uncertainty-aware machine learning to make atomic structure prediction more robust. The method is designed to handle cases in which conventional optimization can be misled by limitations in the underlying theoretical model.

Unlike standard approaches that quickly converge on the “single best”, NARA is designed to navigate towards a diverse search for multiple distinct structures. This is critical because a material’s properties often depend not only on the global ground state, but also on metastable structures accessible under realistic conditions. To balance broad, efficient exploration with computational reliability, NARA utilizes uncertainty estimates to strategically deploy expensive ab initio calculations only when necessary.

In tests on copper surface oxides, NARA achieved higher efficiency than the widely used basin-hopping algorithm. In tests on gold nanoclusters, the framework recovered both planar and non-planar competing structures in a single run, which is important because their relative stability can depend sensitively on the theoretical treatment.

The results suggest that combining multimodal optimization with uncertainty-aware machine learning technique can improve both the efficiency and robustness of atomic structure searches. This could support faster and more reliable computational discovery of new materials.

The work titled “Robust global optimization of atomic structures via a learning loss-informed on-the-fly firefly algorithm”, was published on AI Agent (published on March 31, 2026).
DOI：10.20517/aiagent.2025.13