 |
 |
 |
Researchers at the Technical University of Denmark have assembled SoLiD26, a dataset of 15.4 million calculated atomic structures for machine-learning models of solid-liquid interfaces. In the team's model test, fine-tuning a pretrained interatomic-potential model on the data cut its mean absolute force-prediction error by about 76% on a held-out set of large structures. The work is a September 23 arXiv preprint; the dataset is available by request while a public release is prepared.
Machine-learned interatomic potentials (MLIPs) estimate the energy and forces acting on atoms much faster than repeatedly solving the underlying quantum-mechanical calculations. Their training sets often cover molecules and bulk crystals, but less often the changing boundary where a solid meets a liquid. That interface matters in electrochemistry, catalysis and corrosion, where atoms in the surface, solvent and dissolved ions interact.
SoLiD26 combines calculations from several earlier materials studies at DTU Energy. Most structures came from ab initio molecular-dynamics simulations, which model how atoms move over time. The collection has 15,379,823 structures, up to 576 atoms each, covering 15 elements. Each record contains atomic positions, simulation-cell information, energy and atomic forces. The authors filtered out inconsistent calculations, duplicate structures and outliers; the paper describes a curated dataset rather than 15 million new simulations run for this release.
Fine-tuning lowered force error on larger structures
The researchers compared MACE models trained from scratch or fine-tuned on SoLiD26 with a general pretrained MACE model. They reserved every structure larger than 200 atoms for testing: 272,889 structures in all. Against the density-functional-theory reference forces, the fine-tuned model's mean absolute error was 0.0118 electronvolts per angstrom, down from 0.0501 for the baseline. The test set contained only six elements—hydrogen, carbon, oxygen, potassium, platinum and gold—although the full dataset spans 15.
The comparison is against calculated reference forces, not measured battery performance, corrosion rates or catalyst output; the authors also say they did not systematically optimize the model hyperparameters.
The dataset is not publicly downloadable yet
Uncompressed, the 16 dataset files total about 100 gigabytes. The paper says researchers can request them from the authors while a public release is being prepared. The team's MIT-licensed code covers the data-ingestion and preparation pipeline, but it is not the dataset itself.
A public release would let other groups test the benchmark, while tests on more of the dataset's 15 elements would show whether the gain extends beyond the six-element test set.
Image credit: Example structures from Jonas Busk et al., Figure 1 of SoLiD26, version 1, licensed under CC BY 4.0. These are atomistic data visualizations, not experimental photographs.
Sources
- SoLiD26: A First Principles Solid-Liquid Interface Dataset for Machine-learned Interatomic Potentials, arXiv, September 23, 2026.
- SoLiD26 data-ingestion and preparation pipeline, Technical University of Denmark CAPeX research team.