Basics of machine learning force fields | VASP Lecture

Georg Kresse explains why and how force fields can be trained in VASP using machine learning on-the-fly. He also showcases some example applications based on research conducted in his group.

Also check out this tutorial on machine learning force fields: www.vasp.at/tutorials/latest/md/part2/

00:00:00 Introduction of the speaker
00:00:49 *Outline of the talk*
00:02:31 Definition, power and limitation of “ab initio”
00:05:31 Basic idea of machine learning force fields
00:07:34 Descriptors
00:11:35 Why these descriptors
00:12:53 Details on calculating descriptors
00:16:25 Key assumption: Treat the total energy as a sum of local energies
00:18:07 Definition of forces
00:19:57 How many equations will we have to fit
00:21:54 *Wrap up one*
00:24:20 Linear regression
00:27:26 Kernel trick, reference atoms
00:31:33 *Wrap up two*
00:33:46 On-the-fly learning
00:38:20 **Wrap up three**, design matrix
00:41:40 singular value decomposition after training, regularization
00:46:30 INCAR tags
00:52:18 Examples: solid silicon, CO on Rh surface,
00:55:24 How to train a force field
00:57:49 Refitting your data base
01:01:22 Zirconia: ZrO2
01:10:11 Lattice thermal conductivity
01:13:15 Simple hip to bcc Zr phase transition
01:17:07 LiF in H2O: salvation energies
01:18:46 Delta machine learning
01:19:03 Acknowledgement
01:19:56 Summary
01:21:21 A word of warning
01:23:13 Q&A
01:23:20 What is the ground truth that the machine learning methods are learning?
01:24:13 How to choose Langevin parameters?
01:27:23 Is there an initial guess for the force fields?