NNP/CG-MM:

NNP/CG-MM is a method by which an all-atom neural network potential (NNP) can be systematically embedded into a bottom-up coarse-grained (CG) environment. A custom trained or foundational NNP can be used, and the multiscale coarse-graining (MS-CG) method (or force-matching) can be used for constructing the CG environment.

This repository contains all the essential code and files required to generate the input atomistic trajectories, train the NNPs, construct the NNP-CG and CG-CG interactions, and conduct NNP/CG-MM simulations.

Corresponding paper: J. Chem. Theory Comput. 2026

Key steps:

1. All-atom molecular dynamics simulations to generate reference trajectories (alternatively, QM/MM or AIMD trajectories can also be used)

2. Mapping at the CG resolution and constructing the NNP-CG coupling terms and CG-CG interactions using multiscale coarse-graining (MS-CG) force-matching (FM): OpenMSCG

3. Training a Neural Network Potential (NNP): DeePMD (alternatively, architectures such as MACE and Allegro can also be used)

• Other models used in this work (DNN@MB-pol, MACE-OFF, MACE-POLAR-1)

5. NNP/CG-MM simulations: LAMMPS patched with DeePMD and PLUMED

• LAMMPS: main classical MD engine
• PLUMED: for enhanced sampling

Additional analysis codes and the trajectories can be obtained upon reasonable request.