GitHub

Fitting a Delta-ML model on data from MD

This example demonstrates how to fit a Δ-ML model on top of an existing TB parameterization for GaAs. While all Hamiltonian components (TB, ML, SOC) can in principle be optimized simultaneously, it is significantly more stable to fit them separately.

To this end, we will use eigenvalue data (provided by eigenval.h5) from an MD trajectory (provided by structures.h5) computed with VASP. A POSCAR file (e.g., initial structure of the MD run) is also required to provide atom species information and reference positions. Inputs file can be found here.

Supercell block

The supercell block tells Hamster to analyze multiple structures. The keywords poscar and xdatcar specify the reference POSCAR file and the trajectory file (XDATCAR or *.h5).

The selection of structures is controlled by:

  • nconf: Number of structures to sample.
  • nconf_min: Minimum trajectory index (optional, default = 1).
  • nconf_max: Maximum trajectory index (required).

ML block

The ML model uses a Gaussian kernel to interpolate between a set of descriptor vectors of Hamiltonian matrix elements. In practice, it is recommended to perform hyperparameter optimization for the ML model.

Sampling of descriptor vectors is controlled by:

  • npoints: Number of descriptor vectors to sample.
  • ncluster: Number of clusters used in k-means clustering.

Kernel model parameters can be modified via:

  • sim_params: Kernel width.
  • env_scale: Scaling factor applied only to the environment value.
  • apply_distortion: If enabled, all values in the descriptor vector are updated according to distorted atomic positions.

Optimization

Both train_mode and val_mode should be set to md. The optimization can be parallelized over the number of structures using MPI. Set update_tb to false so that only the ML parameters are updated. Using a small learning rate (e.g., 0.01) is recommended to improve optimization stability.

Output

The ML model parameters are written to a file ml_params.dat. The training and validation losses are found in the output file hamster_out.h5.