sxie22 / uf3-archived Goto Github PK

Candidates for elastic constant calculations are Elastic (used for the manuscript) and iprPy, offered by NIST.

Phonon Spectra may be generated using PhonoLammps + LAMMPS (used for the manuscript) or directly with Phonopy using our calculator for forces.

Currently, all three classes must be instantiated sequentially. It should be straightforward to implement a TransformerMixin that accepts all relevant arguments and creates handlers as attributes.

Currently to construct feature space from vasp calculation only 'vasprun*' files can be used. Would be wonderful if there was an option to also parse 'OUTCAR*' files instead of the vasprun files.

Curently, trained models are serialized with JSON, including knot sequences and b-spline coefficients per pair type. They can be loaded into a UFCalculator to compute energies, forces, etc.

To use the potential with LAMMPS, we export a tabulated potential, consisting of energy and force values at every knot point. This table is used by LAMMPS to load the B-Spline, which can be resampled on a finer grid using LAMMPS' pair_write for further scrutiny.

I am currently testing whether this process is sufficiently lossless. By construction, the B-Spline must pass through the control points at the knots. It should also be uniquely defined by the position of knots and the B-spline coefficients. Regenerating the coefficients from the tabulated potential is trivial, but may have noise associated with the precision at which the tabulated points were written.

If this noise is small enough, the JSON serialization becomes redundant. If not, we may consider distributing the B-spline coefficients alongside the tabulated points (possibly in subheader comments).