abfml 2026.2.3

A problem-oriented package for rapidly creating, screening, and optimizing new machine learning force fields

ABFML

1. Introduction

ABFML is a software platform developed on the popular deep learning framework PyTorch. It provides extensions and support for third-party packages like LAMMPS and ASE. The primary goal of ABFML is to offer researchers a user-friendly platform for the rapid, simple, and efficient development, fitting, and application of new machine learning force field models. By leveraging PyTorch's powerful tensor operations, ABFML simplifies the complex process of developing force field models and accelerates their application in molecular simulations.

2. Installation

2.1 Installing ABFML

To get started with ABFML, follow these steps to set up the environment and install the necessary components:

# 1. Create a new conda environment:
   conda create -n abfml python=3.11

# 2. Activate the conda environment:
   conda activate abfml
#   This command activates the `abfml` environment, making it the current working environment.

# 3. Install abfml:
#   Option 1: Install from source (local directory). Download ABFML and Navigate to the ABFML directory. 
   cd path/ABFML
   pip install .
#   This installs ABFML from the local source directory.

#   Option 2: Install from PyPI
   pip install abfml
#   This installs ABFML directly from the Python Package Index (PyPI).

2.2 Installing ABFML-LAMMPS

In general for force field development users, this step can be skipped for now. It would be more efficient to apply it to large-scale molecular dynamics when all aspects of the force field are relevantly calculated/checked by the more convenient ASEs To integrate ABFML with LAMMPS, follow these steps:

# 1. Download LAMMPS:
#   Obtain the LAMMPS package suitable for your system from the [LAMMPS website](https://lammps.sandia.gov/). This will be the molecular dynamics engine that ABFML will interact with.

# 2. Download Libtorch:
#   Libtorch is the PyTorch C++ API. Unless it can be guaranteed that the system is "new" enough, please use the Pre-cxx11 ABI version.

# 3. Copy ABFML files to LAMMPS:  
   cp abfml/lmp/ABFML lammps/src
   cp abfml/lmp/Makefile.mpi lammps/src/MAKE
#   These commands copy the necessary ABFML files into the LAMMPS source directory.

# 4. Modify the Makefile:
#   Edit `Makefile.mpi` in the `lammps/src/MAKE` directory to match your system's configuration, specifically setting the paths to your compilers and libraries.  
   CCFLAGS += -I/PATH/libtorch/include/
   CCFLAGS += -I/PATH/libtorch/include/torch/csrc/api/include/
   CCFLAGS += -I/PATH/libtorch
   LINKFLAGS += -L/PATH/libtorch/lib/ -ltorch -lc10 -ltorch_cpu
   
# 5. Compile LAMMPS with ABFML:
#   Navigate to the `src` directory of LAMMPS and install: 
   export LD_LIBRARY_PATH=path/libtorch/lib:$LD_LIBRARY_PATH
   cd src
   make yes-abfml
   make mpi

# 6. Run lammps:  
   pair_style abfml model.pt
   pair coeff * * 29 30
  
#   Use the element number to represent the element type

Note: Ensure your GCC compiler is version 9.0 or above, as older versions may not support.

3. Build a new model

Build your own model input in the input file and specify the path to the model's py file, This is an example of a simple LJ potential, which can be found in example/new-model。

{
  "model_setting": {
          "name": "user_defined",
          "model_path": "D:/Work/PyCharm/ABFML/example/LJ/LJ-model.py",
          "field_name": "LJ",
          "normal_name": "LJNormal",
          "descriptor" : {
                         "epsilon": 1.0,
                         "sigma": 1.5
          }
  }
}

Then you can use the relevant modules for training, validation and simulation

4. Run

To run an example using ABFML, follow these steps:

# 1. Activate the ABFML environment:
   conda activate abfml

# 2. Navigate to the example directory:
   cd ABFML/example/DP-sea
  
# 3. Quick check for model correctness
   abfml check -m model.pt -d float32
   
# 4. train:
   abfml train input.json
#   This command starts the training process using the configurations specified in `input.json`.

# 5. valid:
   abfml valid -m 'model.pt' -f "../data/test.extxyz" -n 10

Use the ASE calculator to perform quick calculations on the model created.

from abfml.calculate import ABFML
from ase.build import bulk

calc = ABFML('model.pt')
structure = bulk('Cu', a=3.62)
structure.calc = calc
energy = structure.get_potential_energy()
force = structure.get_forces()

Cite

If you use ABFML in your research, please cite the following paper:

1. Xingze Geng, Jianing Gu, Gaowu Qin, Lin-Wang Wang, Xiangying Meng; ABFML: A problem-oriented package for rapidly creating, screening, and optimizing new machine learning force fields. J. Chem. Phys. 7 February 2025; 162 (5): 052502. https://doi.org/10.1063/5.0247559