It will take place on Tuesday, May 18, 2021, 3 pm (CEST) on Zoom:
Finite temperature properties with first principles accuracy, is machine learning the way to go?
Accurate predictions of phase transition temperatures have always been a dream of materials physicists. Using first principles methods calculations are usually extremely time-consuming and challenging, whereas force fields without extensive and careful tuning tend to provide inaccurate answers. Machine-learned force fields are an obvious solution to this dilemma but training them can be a time-consuming and laborious process.
In this talk, I demonstrate that training on the fly yields highly accurate machine-learned force fields (MLFF) that meet the challenges of predicting finite temperature properties with an accuracy close to the original first-principles method. Our machine learning approach is based on Bayesian regression and uses a combination of radial and angular features computed locally for each atom. The Bayesian regression not only provides predictions for the energies, forces, and stress tensor, but also predicts the uncertainty of these predictions. If the uncertainties exceed a certain threshold, first principles calculations are performed “on the fly”, the structure is added to the training data set, and the MLFF is refined "on the fly". Training is performed simply by heating (or cooling) all phases of interest. Typically, an accurate force field can be obtained in few days and the training requires no special intervention or expertise from the user.
The accuracy of the approach is demonstrated for several materials. For metallic zirconium, our simulations successfully reproduce the first
order displacive martensitic phase transition from hcp to bcc Zr . For Zr, we also show that the MLFF reproduces phonon dispersions and elastic properties with excellent precision. Zirconia (ZrO2) constitutes a more challenging test, with two phase transitions from monoclinic to tetragonal to cubic. Again, the MLFF yields excellent predictions for both transition temperatures . Moreover, we are able to predict the thermal conductivity in very good agreement with experiment. Melting temperatures of Al, Si, Ge, Sn and MgO are predicted using ML-FF trained using various density functionals . In this case, we show that the differences between different density functionals are far larger than the errors introduced by ML. Finally, we address the phase transitions in hybrid perovskites – a class of materials promising for thin film solar cells. Specifically, we calculate the phase transition temperatures of MAPbO3 and several other organic perovskites and find again very good agreement with experiment .
Apart of demonstrating that on the fly MLFFs provide excellent predictions on par with the original density functional, we also show that diverse methods are required to calculated phase transition temperatures: these include slow heating and cooling (Zr), thermodynamic integration (ZrO2), interface pinning (melting temperatures) as well as umbrella sampling (MAPbI3) and free energy perturbation theory.
 P. Liu, C. Verdi, F. Karsai, and G. Kresse, submitted
 C. Verdi, F. Karsai, P. Liu, R. Jinnouchi, and G. Kresse, submitted
 R. Jinnouchi, F. Karsai, G. Kresse, Phys. Rev. B 100, 014105 (2019).
 R. Jinnouchi, J. Lahnsteiner, F. Karsai, G. Kresse, and M. Bokdam,
Phys. Rev. Lett. 122, 225701 (2019).
About the speaker
Professor Georg Kresse received his doctoral degree from the Vienna University of Technology in 1993. After his habilitation at the Vienna University of Technology in 2001, he was offered a full professorship by both the University of Oxford and the University of Vienna in 2006. In 2007 he accepted the chair for Computational Quantum Mechanics in Vienna. Since 2011 Kresse is a full member of the Austrian Academy of Sciences and since 2012 of the International Academy of Quantum Molecular Sciences. He is the recipient of several awards, including the 2003 "START Grant" of the Austrian Science Fund (FWF), the "Hellmann Preis" of the Internationale Working group for Theoretical Chemistry, and the Kardinal-Innitzer-Preis in 2016.
His main research interests are Theoretical Solid State Physics, Surface Sciences and Computational Materials Physics. His work on ab initio density functional theory has shaped the application of density functional theory in materials sciences worldwide. Georg Kresse is the main author and developer of the computer program "VASP" (Vienna ab initio simulation package), which is the most widely used program for quantum mechanical simulations of solids and their surfaces. The three publications on the algorithms implemented in VASP have been cited between 40.000 and 65.000 times each and belong to the 100 most cited research articles ever published.
His current work focuses on the precise description of electron interactions in solids and real materials and encompasses modern perturbative many-body theory, quantum Monte Carlo methods, and machine learning. Georg Kresse is the author of more than 400 research articles. With a Web of Sciences h-index of over 105 he is among the most cited physicists worldwide.
Low-volume newsletters, targeted to the scientific and industrial communities.Subscribe to our newsletter