Combining Quantum Chemistry and Machine Learning for Global Structure Optimization

Abstract

In molecular chemistry and materials science the knowledge of the exact chemical composition of a material is crucial to understand its properties and to allow design of new functional materials. Atomistic modelling can help to analyze the structure-property relation and to determine structures of new unknown materials.

In this talk I will present a method that can be used to determine the structure of molecules and solids without any previous knowledge except the stoichiometry. These quantum chemical calculations are normally computationally extremely demanding, but by training a machine learning potential on-the-fly (by means of a Gaussian Process) the computational demands can be drastically reduced. Using not only the thermodynamic stability calculated using density functional theory, but additionally, experimental values or calculated properties, we are able to determine even meta-stable configurations beyond the thermodynamic minimum.


About the Speaker

Wilke Dononelli is Senior PostDoc at the HMI Group, BCCMS and MAPEX Center. After a bachelor in chemistry and mathematics followed by a master in chemistry, he received a PhD in theoretical chemistry (catalytic reactions at surfaces) at University Oldenburg. Then Wilke worked as a Postdoctoral researcher in theoretical physics (evolutionary algorithms and ML) at Aarhus University and now focuses on theoretical chemistry and materials science at University Bremen. Wilke uses and develops methods for materials modelling including atomistic simulation based on density functional theory, high level wave function quantum chemical methods combining them with evolutionary algorithms and machine learning.

Speaker:
Dr. Wilke Dononelli

---

Das Data Science Center wird gefördert vom: