MedeA Training: On-the-fly Machine-Learning Forcefields with MedeA VASP

MedeA Training: On-the-fly Machine-Learning Forcefields with MedeA VASP

This UGM training session will demonstrate the recent machine learning force field (MLFF) capability using MedeA VASP. 


Traditionally, ab initio molecular dynamics simulations are used to very accurately determine the dynamic properties of materials, where the interactions of atoms and electrons are calculated fully quantum mechanically. However, these simulations are often limited to small simulation cell and simulation times. One way to hugely speed-up these calculations is to use machine learned force fields, which are parameterizations of the potential energy surface. The challenge with machine learning techniques, however, is selecting the proper and minimal training data. “On-the-fly” learning within VASP overcomes this challenge in the following way. A VASP-MD run is used for the learning. During the run of this MD calculation, ab initio data is picked out and added to the training data, and a forcefield is continuously built up. A judgement is made at each step, whether to make an ab initio calculation and possibly add data to the forcefield, or to use the forcefield for that step and skip learning at that step. The more accurate the forcefield gets, the less sampling is needed, and the more expensive ab initio steps are skipped.


In this training session, steps to generate MLFF for NiSi-Si system will be demonstrated. NiSi-Si interface is a technologically relevant material system for complementary metal oxide semiconductor (CMOS) device applications, for its tunable Schottky barrier height at the interface. Following steps will be outlined during the training: 


  1. Running a VASP-MD calculation for Ni, Si, and NiSi systems to generate on-the-fly forcefields 

  2. Judging the quality of the forcefield using Bayesian error and RMSE analysis 

  3. Comparing the MLFF predicted properties (volume-energy curve and thermal expansion coefficient) with DFT calculated values 

  4. Running a MD simulation of a large NiSi-Si interface using the generated MLFF for longer time scale.


Shubham Pandey
Support & Application Scientist

Computational materials enthusiast with expertise in density functional theory calculations and graph based neural networks.

Xiaoli Liu
Support & Application Scientist

Xiaoli Liu, PhD, is a support and application scientist at Materials Design