MDP.jl, the Julia library of Molecular Dynamics (MD) potentials, is being developed to provide fast and accurate potentials for classical MD simulations on exascale supercomputers. Its goals include coupling empirical and machine learning (ML) potentials and quantifying uncertainties in trained ML potentials. This talk presents the latest developments in MDP.jl regarding descriptors and force field computation.
Molecular Dynamics (MD) simulations require a potential energy function to describe the force field governing the interaction among atoms. Force calculation often takes between 50% and 90% of the overall complexity, thus, determining an adequate potential is a crucial task. Classical MD simulations use Empirical Potentials (EP), such as Lennard-Jones or Tersoff, and Machine Learning Potentials (MLP), such as Neural Network Potentials (NNP) or Spectral Neighborhood Analysis Potentials (SNAP). Despite their success, some potentials present limitations when solving complex study cases as ultrahigh temperature ceramics in hypersonic flows. Particularly, existing EP, ReaxFF and COMB, do not produce satisfactory results since they require retraining, while MLP demand a significant amount of Density Functional Theory (DFT) data for the training process. Coupling both potentials is a promising approach for reducing the amount of training data, leveraging physics information coded in the EP. Currently, there are no open-source libraries that allow this coupling and at the same time offer other key features, such as quantifying uncertainties in the trained MLP or identifying near-optimal configurations to include in the training data. MDP.jl, the Julia library of MD potentials, is being developed to fill this gap, aiming to provide fast and accurate potentials for classical MD simulations on exascale supercomputers. In this talk, the latest developments in MDP.jl regarding the descriptors and force field calculation are presented.