Incorporating AIMD and MLIP in the flow
Dear all,
I have a large unit-cell (200 atoms) for which I want to calculate the thermal conductivity. I am expecting to require a very large number of displacements, which is beyond my computational capability. To reduce complexity, I am thinking of using Ab Initio MD (AIMD) and Machine Learning Interatomic Potential (MLIP).
My questions are:
-
Is it possible to replace the reference data (i.e. the random displacements and forces computed by e.g. VASP) with data from AIMD and still have good accuracy? The output of AIMD is trajectories with random displacements and forces at a given temp, which could constitute training data for FCs.
-
Can I use an MLIP model (e.g. GAP, MTP, NNP, SNAP, ..) to predict the forces of the generated displacements. The MLIP itself will be trained on AIMD data of a smaller system. In this case, the hiPhive training models will be trained on forces predicted by another ML model. Can I get reasonable accuracy with this approach?
Thanks