lammps/src/MLIAP/README.md

This package provides a general interface to families of machine-learning interatomic potentials (MLIAPs). This interface consists of a mliap pair style and a mliap compute. The mliap pair style is used when running simulations with energies and forces calculated by an MLIAP. The interface allows separate definitions of the interatomic potential functional form (model) and the geometric quantities that characterize the atomic positions (descriptor). By defining model and descriptor separately, it is possible to use many different models with a given descriptor, or many different descriptors with a given model. The pair_style supports the following models:

• linear,
• quadratic,
• nn (neural networks)
• mliappy (general Python interface to things like PyTorch).

It currently supports only one class of descriptors, sna, the SNAP descriptors. It is straightforward to add new descriptor and model styles.

The mliap compute style provides gradients of the energy, force, and stress tensor w.r.t. model parameters. These are useful when training MLIAPs to match target data. Any model or descriptor that has been implemented for the mliap pair style can also be accessed by the mliap compute. In addition to the energy, force, and stress gradients, w.r.t. each model parameter, the compute also calculates the energy, force, and stress contributions from a user-specified reference potential.

Generating the model files from the third-party packages

• To train the linear and quadratic models with the SNAP descriptors, see the examples in FitSNAP.
• To train the nn model with the SNAP descriptors, check the examples in PyXtal_FF.

Building MLIAP with Python support

The mliappy energy model requires that the MLIAP package be compiled with Python support enabled. This extension, written by Nick Lubbers (LANL), provides a coupling to PyTorch and other Python modules. This should be automatically enabled by default if the prerequisite software is installed. It can be enforced during CMake configuration by setting the variable MLIAP_ENABLE_PYTHON=yes or for conventional build by adding -DMLIAP_PYTHON to the LMP_INC variable in your makefile and running the cythonize script on the .pyx file(s) copied to the src folder.

This requires to also install the PYTHON package and have the cython software installed. During configuration/compilation the cythonize script will be used to convert the provided .pyx file(s) to C++ code. Please do not run the cythonize script in the src/MLIAP folder. If you have done so by accident, you need to delete the generated .cpp and .h file(s) in the src/MLIAP folder or there may be problems during compilation.

More information on building LAMMPS with this package is here.