This repository contains datasets generated using the Physics-Informed Generative Model for Extrapolating Beyond Known Motifs (PIGEN) and its conditioned variants. Each dataset is provided as a CSV file in which each row corresponds to a single generative-model structure (deduplicated by composition). Columns include:

cif – CIF string representation of the generated structure

composition – chemical formula of the generated candidate

SPP – structural validity score

composition_novelty – novelty of the composition with respect to training data

compactness – geometric compactness metric

Iconf – structural complexity measure K per atom

total_entropy – local-environment diversity (MLED)

Additional conditioning-specific metrics depending on the model variant

The repository includes datasets from PIGEN models conditioned on single or multiple targets, such as compactness, MLED, hull energy, diversity, and complexity.


Files related to Figure 5

For the compositions LiNb₇N₈ and Rb₂Er₄O₇, two files are provided for each system:

Generated structures – raw PIGEN-generated candidates in CIF format

Optimised structures – globally optimised crystal structures obtained via crystal structure prediction workflows (CIF format)

These examples correspond to the structures highlighted in Figure 5 of the associated manuscript.