SPARC (Smart Potential with Atomistic Rare Events and Continuous Learning) :is a Python package built around the ASE library.
SPARC is a modular workflow for training machine learning interatomic potentials (MLIPs) for reactive/nonreactive system.
It supports various first-principle calculators, and ML arcitechture together with accelerated sampling of configurational space, which enables efficient simulation and on-the-fly model improvement.
Scientific Overview
Constructing a reactive and transfarable ML potential is challanging, as it requires high quality training data which includes rare events, high energy intermediates, and transition state. A ML model that can generate chemically accurate potential energy surface is important for studying rare events. SPARC automates this by pushing a trained ML model to access different configurations in the potential energy surface to predict beyond its knowledge. The idea behind SPARC is to use the active learning protocol together with advance sampling techniques that systematicallt identifies new configurations and trains a ML model on-the-fly. the workflow consist of four main steps, three main steps, ML model training using DeePMD-kit package, accelerated sampling driven PES exploration using PLUMED library, and new configuration labelling using first-principle calculations, which are executed iteratively in a loop until a reactive and stable MLIP is constructed. .
Key Features
SPARC provides the following core functionalities:
First-principles calculations and ab inito molecular dynamis simulation together with PLUMED plugin.
Machine-learning potential training with DeepMD-kit.
Machine-learning molecular dynamics (ML/MD) simulations using ASE MD engine for NVT ensamble.
Active learning using query-by-committee approach to label new configurations for continuous model improvement.
Potential energy surface exploration with integration of PLUMED plugin.
Visualisation of verious properties to monitor the workflow.
Use Cases
Here are some example use cases for SPARC:
Material property prediction: : Develop accurate interatomic potentials for materials property prediction and molecular dynamics simulation.
Chemical reaction pathways: : Uncover reaction mechanisms by finding new configurations and training a ML model on-the-fly.
High-throughput simulations: to efficiently explore large chemical spaces.
Installation
For detailed installation instructions, please refer to the Installation Guide.