Research data supporting "Amorphous materials as a frontier challenge for universal interatomic potentials"

This repository supports the manuscript:

Natascia L. Fragapane and Volker L. Deringer, "Amorphous materials as a frontier challenge for universal interatomic potentials" (in preparation).

AM26 dataset

This repository currently contains the AM26 benchmark dataset used to evaluate pre-trained MLIPs for amorphous materials.

The benchmark comprises 5 amorphous systems, together with reference density-functional theory (DFT) energies and forces, spanning a diverse range of chemistries and bonding environments.

The AM26 dataset is available in the data directory.

Structures included in the AM26 dataset were created using melt-quench simulations for the following systems:

a-C: Melt-quenched structures (1.5–3.5 g cm⁻³) generated using the C-GAP-17 model.

a-SiO₂: Melt-quenched structures (quench rates 10¹¹–10¹⁴ K s⁻¹) generated using the SiO_x-ACE-24 model.

a-LiPS: Melt-quenched structures spanning compositions along the Li₂S–P₂S₅ tie-line, generated using the model of N. L. Fragapane, V. L. Deringer, J. Chem. Theory Comput. 2026, 22, 3646-3659 [link].

Structures were taken from previous work for the following systems:

a-Si: Filtered subset of the a-Si-24 dataset from L. A. M. Rosset, D. A. Drabold, and V. L. Deringer, Nat. Commun. 2025, 16, 2360 [link].

a-GST: Combination of AIMD configurations from GST-GAP-22 and the external validation set spanning compositions along the GeTe–Sb₂Te₃ tie-line: see Y. Zhou, W. Zhang, E. Ma, V. L. Deringer, Nat. Electron. 2023, 6, 746-754 [link].

Analysis

We provide the scripts used to reproduce the benchmark error analysis and leaderboard presented in Fig. 2 of the manuscript in the scripts directory.

These scripts compute prediction errors with respect to the DFT reference data, evaluate the corresponding relative root-mean-square error (RelRMSE) scores, and generate the summary figures used for the benchmark analysis.

Repository status

This repository will be expanded as additional data and analysis associated with the manuscript become publicly available.