Host institution: Technische Universitat Darmstadt (TUDa), Germany*
Co-supervisor: Grenoble INP Institut d'ingénierie et de management de l'Université Grenoble Alpes (G-INP), France
Partner enterprise: Matgenix, Belgium
Abstract of the project:
For various applications, the atomic-scale dynamics of the phenomena is critical with respect to the performance. Simulation of such phenomena can thus be a very valuable tool. However, a compromise often needs to be found between the accuracy that can be achieved and the time- and length-scales of the simulation. On the one hand, semi-empirical interatomic potentials can be used for rather long and large simulations, but they lack accuracy in sampling a wide chemical space. On the other hand, ab initio molecular dynamics (AIMD) simulations can always achieve the required accuracy but are very demanding, and as such they can only be applied to rather small numbers of atoms for relatively short simulation time frames (~100 ps). Recently, machine-learned potentials (MLPs) parametrized by ab initio data have been employed to optimize the efficiency-versus-accuracy trade-off:[1] i.e., expanding the time and length scales accessible with AIMD while retaining a similar level of accuracy. In particular, “learn on-the-fly” (LOTF) approaches[2] are really promising. These rely on active learning to automatically determine when the MLP needs to be retrained with more data points to preserve the accuracy and avoid highly non-physical atomic dynamics. The present project aims to develop an automatic approach to develop MLPs which can then be used to simulate relevant problems such as (i) the oxygen evolution reaction at WO3-based catalyst for water splitting (PhD6), (ii) the diffusion of Na-ion in MXenes as positive electrodes in batteries (PhD7) and (iii) the deposition of thin films with ABX3 structure A=Ca, Sr, Ba, B= Ti, Zr,Hf and X=S, Se because they should provide interesting properties as light absorbers for photovoltaic systems with improved stability (PhD8). At TUDa, the PhD will mainly learn how to (i) use AiiDa to automatically generate new reference data,[3] (ii) retrieve already existing reference data from open databases, and (iii) build accurate MLPs from the gathered data. At G-INP, the PhD will mainly learn about the LOTF approach. The research applied to the specific projects (PhD6, 7 and 8) will be performed both at TUDa and G-INP. The industrial partner Matgenix will guide software development.
Supervisors:
TUDa: H. Zhang
G-INP: N. Jakse
Partner Enterprise:
Matgenix
D. Waroquiers
* If you have worked/lived/resided in this country for more than 12 months during the 3 years before the start of this PhD (typically: Sept 1st 2022 to August 31st 2025), you are not eligible for this position.
1 P. Friedrich et al., Nat. Mater. 2021, 20, 750.
2 C. Wang et al., Chem. Mater. 2020, 32, 3741.
3 K. Mathew et al., Comput. Mater. Sci. 2017, 139, 140
2 C. Wang et al., Chem. Mater. 2020, 32, 3741.
3 K. Mathew et al., Comput. Mater. Sci. 2017, 139, 140
