DICE: software, source data and trained models

Research materials accompanying DICE, an active-learning distillation workflow for developing compact machine-learning interatomic potentials. DICE-software.zip: workflow implementation, user documentation and CPU examples. DICE-paper.zip: numerical source data for the main and supplementary figures and tables, analysis scripts and figure-reproduction code. DICE-models.zip: eleven trained study models, model identities, native species maps and runtime requirements. DICE-datasets.zip: labelled target DFT references, student training datasets and test sets for Al₂O₃, YSZ and MIL-120–CO₂, together with historical cross-phase student training data and the independent 120-frame liquid–amorphous test set. Extract the model and dataset archives into the same parent directory; they share research-assets/. The accompanying README files explain package contents, numerical verification, figure/table reconstruction and evaluation of the supplied models. Licences: DICE software and paper-reproduction source code: MIT. Author-owned data, figures and documentation: CC BY 4.0. Study model weights adapted from DPA4-Neo/DPA4C-Nano: CC BY-NC 4.0. Each archive includes licence scope and source attribution.