Skip to content

alff Documentation

alff


ALFF: Frameworks for Active Learning Graph-based Force Fields and Material Properties Calculations.

Developed and maintained by C.Thang Nguyen

ALFF is designed to automatically generate graph-based force fields through iterative active learning cycles (ML training → MD exploration → DFT labeling). The workflow is fully modular and highly customizable, enabling the entire process to run end-to-end without any user intervention. Key features of ALFF include:

  • Automatic generation of necessary scripts for ML training, MD simulations, and DFT calculations.
  • Submission of jobs to remote clusters, support almost connection protocols (e.g., Local, SSH, Cloud APIs,...), job schedulers (e.g., SLURM, SGE, PBS, TORQUE,...), and heterogeneous computing resources.
  • Automated monitoring of job status and retrieval of results upon completion.
  • Parsing of results and execution of active learning iterations automatically.
  • Candidate-selection criteria based on energy, forces, (and/or) stress.
  • Easily configure sampling spaces, including (and/or) temperatures, stresses, enhanced samplings, van der Waals correction.
  • Support multiple MD/DFT calculators.
  • Support several leading graph-based MLIP models, and easy to implement any type of other MLIP models.

A distinctive capability of ALFF is its support for distributing workloads across multiple remote clusters. Instead of relying on a single cluster with long queue times, ALFF can submit, monitor, and handle results across several heterogeneous computing infractructures "asynchronously", dramatically accelerating the active learning workflow. Imagine being able to combine computing resources from completely different sources - Google Cloud, Amazon Web Services, national supercomputing centers, local campus clusters, and more - and have them work together on the same workflow? That is exactly what ALFF can do.

The modular design allows ALFF to easily extend/implement new functionalities/workflows. It also includes built-in workflows to automatically compute phonon dispersion, Potential energy surface (PES), elastic constants tensor, and more to be added.

The only task required from users is to provide a configuration file - then fasten the seatbelt and enjoy the ride.