A February 2026 study of zirconium–copper metallic glass finds useful structural information extending roughly five angstroms, or half a nanometre, from an atom, around its second-neighbour shell. Two different machine-learning approaches converge on that scale when predicting the energy associated with a sample’s atomic arrangement.[1]
Metallic glasses are alloys whose atoms lack a repeating crystal lattice, yet form local clusters. In a 2024 experiment on two zirconium alloys, researchers combined seven measurement techniques. X-ray absorption constrained bonds around particular elements; electron diffraction probed individual clusters and their connections. Changing the electron probe’s size helped estimate ordered domains averaging 0.55 and 0.9 nanometres in the two glasses. These measurements describe how clusters assemble across space.[2]
Measuring how much context helps
Using 10,000 stress-free simulated samples, each containing 2,000 atoms, one model counted recurring atomic neighbourhoods and predicted total configurational energy from their frequencies. Its errors on 2,000 held-out samples were smallest with a five-angstrom radius.[1]
A vision transformer instead read three-dimensional grids of atomic positions. Its accuracy levelled off when information could be exchanged within cubes eight angstroms across. The authors compared cube and sphere volumes to express that as an equivalent radius near five angstroms.[1]
A palladium–silicon glass favoured a shorter neighbourhood. The useful outcome is a way to locate the scale carrying energy information for a particular glass chemistry.[1]
Sources & further reading
- Dual machine learning pinpoints the Radius of Informative Structural Environments in metallic glasses npj Computational Materials · 14 Feb 2026 · Accessed 20 Oct 2026
- Comprehensive characterization of the structure of Zr-based metallic glasses Scientific Reports · 28 Feb 2024 · Accessed 20 Oct 2026