ALIGNN

Overview

ALIGNN implements the Atomistic Line Graph Neural Network for learning properties from atomistic structures. Its architecture combines convolutions on a bond graph and its line graph to represent pair and triplet interactions. The repository includes predictive models and training tools, rather than a materials database: JARVIS-DFT and other datasets supply training targets. Within a materials research workflow, ALIGNN can provide structure-based property estimates or support development of a predictor for a user-supplied dataset.

Property-model training accepts structure files in POSCAR, .cif, .xyz or .pdb format, an id_prop.csv mapping filenames to targets, and a configuration file defining training settings. Documented outputs include single-property regression, multi-output regression for properties or density-of-states targets, and binary classification. Pretrained models can make predictions directly; the JARVIS-ALIGNN web application specifically accepts POSCAR input and predicts formation energy, total energy per atom and bandgap using JARVIS-DFT-trained models.

ALIGNN-FF extends the repository to atomistic force fields, with energy, force and stress training data supplied through id_prop.json; training energies must be expressed per atom. The README describes models covering combinations of 89 elements, fine-tuning, an ASE calculator, and examples for relaxation, energy–volume curves, phonons and molecular dynamics. These capabilities should not be treated as evidence of accuracy for every target system. The source identifies ALIGNN-FF as actively developed, limits the classification script to binary tasks, notes DGL installation issues, and states that multi-GPU training has not been thoroughly tested.

Key Features

• Combines edge-gated convolutions on atomistic bond and line graphs to model pair and triplet interactions.
• Supports single-output and multi-output property regression, including electron and phonon density-of-states examples, plus binary classification.
• Provides configurable training from structure files and target mappings, with adjustable dataset splits and an option to preserve data order.
• Offers pretrained property prediction through `pretrained.py` and POSCAR-based property prediction through the JARVIS-ALIGNN web application.
• Supports ALIGNN-FF training and fine-tuning from energy, force and stress data, with energy supplied per atom.
• Includes an ALIGNN-FF ASE calculator and examples for structure relaxation, energy–volume curves, phonons and melt-quench molecular dynamics.

Use Cases

• Intended evaluation: screen candidate crystal structures with a pretrained formation-energy or bandgap predictor, then compare estimates with suitable reference calculations before selecting materials.
• Intended evaluation: train a domain-specific property model from labeled structures, using held-out data to assess whether its predictions transfer to the intended materials family.
• Intended evaluation: test ALIGNN-FF for structural relaxation or energy–volume analysis on representative systems and compare energies, forces and relaxed structures with reference results.
• Intended evaluation: explore multi-output learning for electron or phonon density-of-states targets using the documented training examples.

How to Use

1. Read the repository README and choose between pretrained property prediction, custom property training and ALIGNN-FF. These workflows require different targets and configurations.
2. Follow the README’s installation section for your environment. It provides conda, repository and pip routes and flags DGL installation difficulties; consult the linked DGL instructions when selecting dependencies.
3. For property training, collect supported structure files, create id_prop.csv, and adapt the supplied configuration example. Set target definitions and train/validation/test splits deliberately.
4. For pretrained inference, follow the documented pretrained.py example, or try POSCAR input in the JARVIS-ALIGNN app. Confirm which property and training dataset the selected model represents.
5. For force-field work, study the Silicon training notebook, prepare id_prop.json with per-atom energies, and evaluate predictions against reference data before extending to relaxation or dynamics.