Requests for technical support from the VASP group should be posted in the VASP-forum.

Category:Machine Learning

From Vaspwiki
Jump to navigation Jump to search

This section describes the methodology used for force-field generation using machine learning. One first should checkout the theoretical background Machine learning force field: Theory together with the basic description how to run the machine learning calculations Machine learning force field calculations: Basics. Then gain some hands-on experience with the following tutorial: Liquid Si - MLFF.


Key arguments for using machine learning force fields in VASP:

  • On-the-fly training is directly integrated in VASP
  • Training data is collected and force field is generated while MD is running
  • Sophisticated sampling and training algorithm based on Bayesian learning
  • Drastically reduces computational cost of MD simulations
  • Easy to use, ideally only a single extra INCAR tag: ML_LMLFF
  • Yet fully configurable: 50+ tags


  • Besides the usual input files (INCAR, POSCAR, etc.) the machine learning force field method may require the following input files, depending on the mode of operation (see ML_ISTART):
    • ML_AB Ab initio data used to create the training data.
    • ML_FF File containing force field parameters.


  • The machine learning force field method generates the following output files:
    • ML_LOGFILE Main output file for the machine learning force field method.
    • ML_ABN New abinitio data (used as ML_AB in the next run).
    • ML_REG Output file summarizing regression results.
    • ML_HIS Output file summarizing the histogram data.
    • ML_FFN File containing new force field parameters (used as ML_FF in the next run).
    • ML_EATOM Output file containing local atomic energies
    • ML_HEAT Output file including local heat flux

All INCAR tags belonging to the machine learning force field method start with the prefix ML_.

Theoretical Background

How to


  • Basic tutorial to learn how to perform on-the-fly learning and how to control the accuracy of the force field: Liquid Si - MLFF.