Carcará is a high-performance Python framework designed for atomistic simulations powered by on-the-fly machine learning interatomic potentials (OTF-MLIP). It streamlines the integration of first-principles accuracy with the efficiency of classical force fields, enabling the automated development of robust potentials during the simulation process.
Key Features
On-the-Fly Training: Automate the training cycle during simulations, reducing the need for manual dataset curation and ensuring the potential is accurate for the relevant phase space.
Diverse Configuration Sampling: Generate new training structures through multiple pathways:
Random Displacements: Introduce displacements in atomic positions and lattices randomly using “Normal” or “Uniform” distributions.
Molecular Dynamics (MD): Structural exploration via various ensembles, such as NVT or NPT.
Minimum Energy Paths (MEP): Sampling through diffusion paths and transition states, using methods like Nudged Elastic Band (NEB), getting configurations along reaction coordinates for rare event sampling.
Active Learning & Uncertainty Quantification: Utilize Machine Learning Committees (ensembles) to identify configurations with high model disagreement, targeting high-uncertainty regions for further training.
Scalable Sample Generation: Efficiently produce large-scale datasets for ML training through continuous random displacements or extended MD trajectories.
Core Focus
The hallmark of Carcará is its optimization for continuous, autonomous model evolution. Whether you are running long-scale molecular dynamics or generating vast structural libraries via random displacements, Carcará ensures that your machine learning model adapts to the chemical environment in real-time, significantly lowering the barrier for complex materials modeling.