CsPbBr_3 database for force field training underlying the publication: Tuning Einstein Oscillator Frequencies of Cation Rattlers: A Molecular Dynamics Study of the Lattice Thermal Conductivity of CsPbBr3 
Jonathan Lahnsteiner, Max Rang and Menno Bokdam (December 2023)

These structures have been selected with the on-the-fly Machine-Learning Force Fields method as implemented in VASP 6.3:
Jinnouchi R., Lahnsteiner J., Karsai F., Kresse G., Bokdam M., "Phase transitions of hybrid perovskites simulated by machine-learning force fields trained on the fly with Bayesian inference", Phys. Rev. Lett. 122, 225701, (2019)

The structures have been automatically selected during a heating run from 50 to 500 K and the corresponding coordinates, energies, forces and stresses are stored in the ML_AB file. The ML force field generated by regression as implemented in VASP is stored in ML_FF. This file is readable for VASP 6.3 and higher. 

The ML_AB file can be used to generate new force fields with the method of the users preference. The open-source FPdataViewer software (https://github.com/dynamicsolids/FPdataViewer) can be used to read-in the ML_AB file. It also contains connection to the Atomic Simulation Environment with which descriptors can be generated. 

The data set is includes the ML_AB, ML_FF files for the following configurations:
    CsPbBr3 with light Cs masses (m=m_Cs/10)  
    CsPbBr3 with normal Cs masses (m=m_Cs)
    CsPbBr3 with heavy Cs masses (m=m_Cs*10)
Eventhough the DFT energy does not depend on the atomic mass, the structural phase space sampled during on-the-fly training can be different.

FPdataviewer factsheets
A high level overview of the ML_AB databases has been generated using the open-source FPdataViewer software. Each pdf file contains statistics related to the structures, energies and forces stored in the databases. The factsheet can be used to get a quick overview of the data stored in the database.



Typical INCAR file for MLFF production run in VASP 6.3:
################################
### Parameters for VASP      ###
################################
SYSTEM = FAPbI3
PREC = FAST
METAGGA = SCAN ; ALGO = Fast
SIGMA = 0.01 ; ISMEAR = 0
ENCUT =  287
NCORE = 8 ; KPAR = 4
EDIFF = 1E-4 ; NELMIN = 6
ISYM = 0
LREAL = A
IBRION = 0
ISIF = 0
MDALGO=3               
LANGEVIN_GAMMA = 1.0 1.0 1.0 1.0 1.0  # friction coef. for atomic DoFs for each species
PMASS=100              # mass for lattice DoFs
LATTICE_CONSTRAINTS = T T T   # fix x&y, release z lattice dynamics
PSTRESS=0.001          # P is set at 0.001 KB.
POTIM = 5.0
SMASS = 0
TEBEG = 500.00D0
TEEND = 500.00D0
NSW = 200000
NBLOCK = 100
NWRITE = 0
NELM = 100
ISPIN = 1
INIWAV = 1
IWAVPR = 1
ISTART = 0
LWAVE = .FALSE.
LCHARG = .FALSE.
################################
### MACHINE-LEARNING         ###
################################
### General parameters       ###
ML_FF_LMLFF = .TRUE.  # Set as LMLFF_FF = .TRUE., when some machine-learning force field calculations are executed. Default is .FALSE.
ML_FF_ISTART = 2     # Parameter to control restarting calculations. Set 0, 1 or 2.
                      # When you execute the training from the scratch, this needs to be set to 0. 
                      # When you restart the training reading the previous ab initio data (ABCAR file), this needs to be set to 1.
                      # When you execute MD using only MLFF, this needs to be set to 2 with LMLONLY_FF=.FALSE.
ML_IALGO_LINREG = 3




