Skip to content

Latest commit

 

History

History
123 lines (69 loc) · 3.86 KB

MTP.md

File metadata and controls

123 lines (69 loc) · 3.86 KB
Raw

MTP Info

Training Cycle

Train

Input Files

curr.mtp train.cfg

Output Files

None

mlp Command

mlp train curr.mtp train.cfg --max-iter=500 --trained-opt-name=curr.mtp --curr-pot-name=curr.mtp --stress-weight=5e-4 --force-weight=5e-3

Purpose

Trains the potential in curr.mtp on the data in train.cfg. [GLWH Nov 26 2020] In other words, this is the fit. This is the part a potential is fitted to the energies, forces, and stresses in a group of vasp calculations.

Calc Grade

Input Files

curr.mtp train.cfg train.cfg temp.cfg

Output Files

temp.cfg state.als

mlp Command

mlp calc-grade curr.mtp train.cfg train.cfg temp.cfg --nbh-weight=0.0 --energy-weight=1.0

Purpose

Calculates the MV Grade of train.cfg and sets up the state.als file. [GLWH Nov 26 2020] The MV grade isn't needed. But the als file stores the active learning set.

Relax

Input Files

mlip.ini curr.mtp catalog.cfg

Output Files

relaxed.cfg_* unrelaxed.cfg_* selected.cfg_*

Note: The _* signifies that there will probably be multiple files, depending on the number of processes you are running the command with.
[GLWH Nov 26 2020]
For me the output files are B-preselected.cfg_#. And watch out! If you run this command twice in a row, without deleting B-preselected.cfg it appends to the file. Gotcha!

mlp Command

mlp relax mlip.ini --force-tolerance=1e-3 --stress-tolerace=1e-2 --max-step=0.03 --cfg-filename=catalog.cfg --save-relaxed=relaxed.cfg --save-unrelaxed=unrelaxed.cfg

Purpose

This is sort of the heart of the MTP algorithm. It attempts to relax the structures in catalog.cfg to an equilibrium state. If it fails to do so, within a certain tolerance of estimated error, the structures that break this tolerance are collected. [GLWH Nov 26 2020] Some (or all of these preselected structures will be added to the training set. This selection is done in the next step: select-add.

Select Add

Input Files

curr.mtp train.cfg selected.cfg state.als

Note: selected.cfg is a concatenation of all selected.cfg_* from the previous step

Output Files

diff.cfg active_set.cfg

mlp Command

mlp select-add curr.mtp train.cfg selected.cfg diff.cfg --select-threshold=3.0 --nbh-weight=0.0 --energy-weight=1.0 --als-filename=state.als --selected-filename=active_set.cfg

Purpose

This step selects files from selected.cfg to run in VASP. These will then be added to the training set and the process starts again.

Errors Encountered when Training MTP's

Potential mlp Errors:

train Step Hangs Forever

"This often happens when there are too few configurations in the database - there is a condition that the error on each iteration should decrease, but when there are too little configurations, severe overfitting occurs and the iterative algorithm cannot decrease the error due to round-off errors. Can be safely ignored and forgotten about when you expand the training database sufficiently large. Not easy to detect automatically, though." - Alexander Shapeev

MLIP: ERROR: Error reading .* file

This error occurs when you have an input in an mlp command but the file it refers to is not there. For example, if you see MLIP: ERROR: Error reading .mtp file, then your mtp file is missing.

Potential AutoGR Errors:

It is especially important to look out for AutoGR errors because if AutoGR fails, VASP will automatically generate KPOINTS for you, and most likely not at the same KPOINT density you want.

The point group doesn't match the niggli basis id

This one hopefully will be fixed soon in the AutoGR code.

Potential VASP Errors:

Internal error in subroutine SGRCON: Found some non-integer element in rotation matrix

This has an unknown cause, but could be caused by ... (maybe add some more info here that I don't know).