ecumetric

This repository contains the ecumetric model and sketch-map projector which were used in MAD-CORE.

The ecumetric model (ecumetric_model.pt) provides descriptors used to construct the MAD-CORE hierarchy. The model uses the SPACE architecture of metatrain. Its per-atom feature output is pooled into one 512-dimensional vector per structure, with concatenation of the mean and the standard deviation over atoms. Distances between these vectors, after scaling, define the ecumetric metric.

The ecumetric explorer model (ecumetric_explorer.pt) maps the ecumetric descriptors to eight-dimensional sketch-map coordinates for exploration and visualization.

For re-exporting, see reproduce/README.md.

Exploring a dataset with chemiscope

import ase.io
import chemiscope

structures = ase.io.read("structures.xyz", ":")
chemiscope.explore(structures, featurizer="ecumetric-1.0")

Citation

• MAD-CORE: Sofiia Chorna, Cesare Malosso, Jonathan Schmidt, Michele Ceriotti, Hierarchical universal open structure dataset for atomistic modeling, https://arxiv.org/abs/2610.21771
• SPACE architecture: https://arxiv.org/abs/2610.10830
• sketch-map: https://doi.org/10.1073/pnas.1108486108