Introduction

Overview of PLIP

The Physical LassoLars Interaction Potential (PLIP) is a hybrid python/C++ package for generating machine learning interatomic potentials (MLIP) for atomistic simulations. A key issue related to the current MLIP methods that are often designated as “black boxes” is the lack of physical and chemical interpretability of the obtained potential. The PLIP method combines a physically motivated mathematical formulation for the potential PLIP descriptors and a constrained linear regression. The PLIP model can be deployed in LAMMPS to perform high-performance molecular dynamics simulations.

PLIP layout

../_images/PLIP.png

The process of generation of PLIP based machine learning interatomic potentials can be divided into three parts :

  • Generation of plip fingerprint (descriptor) : In this step, the structures of interest are first converted into a set of numerical descriptors known as PLIP fingerprint. These descriptors capture relevant information about the atomic environment, such as distances, angles, and coordination numbers, which are used to characterize the local atomic interactions within the material.

  • Optimzing the coefficients of the descriptors with LassoLars method: Once the PLIP fingerprints have been generated for the structures in the database, the next step is to perform LassoLars fitting to select the relevant coefficients.

  • Generating LAMMPS compatible potentials: In the final step, LAMMPS-compatible potentials are generated based on the fitted coefficients.