Overview

What AGeDi is

AGeDi (Atomistic Generative Diffusion) is a framework for periodic atomistic structure generation using diffusion models.

Core capabilities

• Build graph data from ASE Atoms objects

• Train diffusion models on atomic coordinates and/or atom types

• Generate new structures from formulas, templates, or explicit defaults

• Use either a command line interface or a Python API

High-level package layout

• agedi.data

  • AtomsGraph: graph structure used by model/noisers

  • Dataset: Lightning DataModule for splitting and batching training data

• agedi.models

  • ScoreModel: combines representation + conditioning + score heads

  • agedi.models.schnetpack: PaiNN-based translator and heads

• agedi.diffusion

  • Agedi: LightningModule orchestrating loss, training, and sampling

  • noisers: forward/reverse diffusion components by variable type

  • distributions and sdes: priors, schedules, and stochastic dynamics

• agedi.functional

  • script-friendly entry points (create_*, train, train_from_atoms, sample)

• agedi.cli

  • agedi train / agedi sample / agedi inspect

Typical workflow

1. Load ASE structures.

2. Build/train model (CLI or Python API).

3. Inspect saved hyperparameters/checkpoints.

4. Sample new structures.

5. Export as ASE trajectory for downstream evaluation.