IQTC Seminar: Prof. Harald Oberhofer
Computing and Predicting Properties of Energy Materials
Friday 19th of September at 11:00h at the Sala d’Actes de Química Física, Prof. Harald Oberhofer from Universität Bayreuth will deliver the lecture “Computing and Predicting Properties of Energy Materials”.
With the advent of ever more powerful computers and more accurate, yet efficient
algorithms, computational science has by now been widely accepted as a
valuable and equal contribution to both pure theory and experiment.
Traditionally, computation thereby played the roles of elucidating microscopic
properties and mechanisms of given systems and reaction pathways, leading to
numerous breakthroughs not otherwise possible. The choice of system or
reaction, thereby was mostly guided by experiment or the intuition and
experience of the researcher. Recently, though, modern data science approaches
such as data-mining and explainable machine learning allowed computation to
take on a role of pro-active exploration and design, supplementing the
traditional roles of computational materials science.
In my presentation I will outline some of my research regarding method
development and application in this field. First, I will present our recent
extension of the famous Newns Anderson model, converting it from a qualitative
tool aimed at gaining a rough understanding of adsorption and charge transfer
problems, to a fully convergent and quantitative method that can be used to
predict experimental results. Second, I will show how the use of explainable
machine learning models not only allows us to predict properties of materials,
in this case the elasticity of layered perovskites, but also to extract the
materials’ characteristics leading to these properties. Thus, we show how such
an approach can yield useful structure-function relationships.
Thereby, all my work is based on a bottom up approach, where specifically
developed methods on the electronic structure level, aided by machine learned
datapoint selection and machine learned or classical force-fields inform larger
scale effective, embedded or kinetic models, which in turn yield the
descriptors necessary for data-driven and machine learning based analyses of
more general problems.