psiflow – modular and scalable active learning for interatomic potentials

Recently, we launched psiflow v2.0.0. Psiflow is a modular and scalable library for developing interatomic potentials. It uses Parsl to interface popular trainable interaction potentials with quantum chemistry software, and is designed to support computational workflows on hundreds or thousands of nodes. Psiflow is designed as an end-to-end framework; it can orchestrate all computational components between an initial atomic structure and the final trained potential. Using a variety of active learning approaches, the system's phase space is efficiently explored without requiring ab initio molecular dynamics.

Its features include:

  • active learning algorithms with enhanced sampling using PLUMED
  • Weights & Biases logging for easy monitoring and analysis
  • periodic (CP2K) and nonperiodic (NWChem) systems
  • massively parallel execution on thousands of nodes
  • efficient GPU molecular dynamics using OpenMM
  • supports the latest equivariant potentials such as MACE and NequIP

In addition, psiflow features a modular and concise API in which arbitrarily large workflows can be formulated, which can then be executed on a large number of different execution resources, including clouds (e.g. Amazon Web Services, Google Cloud), clusters (e.g. SLURM, Torque/PBS, HTCondor) and even container orchestration systems (e.g. Kubernetes). Visit the psiflow documentation for more details.