Hierarchical universal open structure dataset for atomistic modeling
Abstract
Atomic-scale simulations have been transformed by machine-learning models that predict the properties of atomistic structures with the accuracy of quantum calculations, but at a small fraction of the cost. These models have increased dramatically in generality, thanks to the availability of synthetic datasets of ever-increasing size and scope. The proliferation of datasets, with properties often computed using different levels of electronic-structure approximation, is problematic, as there is often overlap between different datasets, and strong correlations between structures obtained with short molecular dynamics trajectories or geometry optimization. To address these problems, we introduce a protocol to stratify a dataset into a hierarchy of nested subsets, subsampling from the coarsest to the finest resolution based on an "ecumetric" built to reflect the latent space representation of several popular universal machine-learning interatomic potentials. We then use this approach to combine most of the open structural datasets currently available, consolidating them into MAD-COSMO, a well-organized collection of almost 1.7 billion structures, whose 224 most diverse members form the MAD-CORE stratified coreset, which can be easily used for all sorts of tasks, from exploratory experiments to the training of foundation models at scale. We envision that this collection will guide and help coordinate future efforts, focusing them on the most relevant and/or undersampled regions of chemical and structural space, and that it will be further extended to incorporate the outcomes of the collective work of this community.
Related
- QUASAR: A Universal Autonomous System for Atomistic Simulation and a Benchmark of Its Capabilities
- A Lightweight Universal Machine-Learning Interatomic Potential via Knowledge Distillation for Scalable Atomistic Simulations
- Assigning the Low-Energy Excitons in Far-Red Photosystem II through Atomistic Modeling
- From sparse quantum-computing data to atomistic simulation with universal machine-learning interatomic potentials
- Quantum-accurate atomistic modeling of enzyme catalysis using a machine learned potential
BibTeX
@misc{chorna2026hierarchical,
title = {{Hierarchical universal open structure dataset for atomistic modeling}},
author = {Sofiia Chorna and Cesare Malosso and Jonathan Schmidt and Michele Ceriotti},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.21771},
eprint = {2610.21771},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.21771},
}
RIS
TY - UNPB
TI - Hierarchical universal open structure dataset for atomistic modeling
AU - Sofiia Chorna
AU - Cesare Malosso
AU - Jonathan Schmidt
AU - Michele Ceriotti
PY - 2026
DA - 2026/10/15
PB - arXiv
DO - 10.48550/arxiv.2610.21771
UR - https://arxiv.org/abs/2610.21771
AB - Atomic-scale simulations have been transformed by machine-learning models that predict the properties of atomistic structures with the accuracy of quantum calculations, but at a small fraction of the cost. These models have increased dramatically in generality, thanks to the availability of synthetic datasets of ever-increasing size and scope. The proliferation of datasets, with properties often computed using different levels of electronic-structure approximation, is problematic, as there is often overlap between different datasets, and strong correlations between structures obtained with short molecular dynamics trajectories or geometry optimization. To address these problems, we introduce a protocol to stratify a dataset into a hierarchy of nested subsets, subsampling from the coarsest to the finest resolution based on an "ecumetric" built to reflect the latent space representation of several popular universal machine-learning interatomic potentials. We then use this approach to combine most of the open structural datasets currently available, consolidating them into MAD-COSMO, a well-organized collection of almost 1.7 billion structures, whose 2^24 most diverse members form the MAD-CORE stratified coreset, which can be easily used for all sorts of tasks, from exploratory experiments to the training of foundation models at scale. We envision that this collection will guide and help coordinate future efforts, focusing them on the most relevant and/or undersampled regions of chemical and structural space, and that it will be further extended to incorporate the outcomes of the collective work of this community.
ER -