Andreas Grüneis: Reaching Chemical Accuracy in ab initio Simulations of Complex Materials

Andreas Grüneis: Reaching Chemical Accuracy in ab initio Simulations of Complex Materials

The discussion starts at 29:02.

Complex oxides belong to a class of materials that represent a major challenge for the most widely-used ab initio methods employed in atomistic computer simulations, such as density functional theory. The long-term objective of our work is to achieve an ab initio prediction of a wide range of material properties of complex oxides with unprecedented accuracy. We plan to reach this goal by combining existing highly accurate quantum chemical wavefunction-based methods for solids with novel machine learning techniques. We aim to investigate ground state properties of various crystal phases, defective structures, and surfaces with adsorbates. These systems will be selected in close collaboration with the experimental groups of Ulrike Diebold and Gareth Parkinson. Two different machine-learning methods will be explored in this project that are both based on kernel ridge regression. On the one hand, we will explore Delta-machine learning techniques based on descriptors derived from the atomic structure. On the other hand, we will investigate novel wavefunction representations to learn highly accurate ab initio data directly from different levels of many-electron theory. This combined approach holds the promise to keep the required training set size at an absolute minimum and take advantage of a diverse set of machine learning tools explored in collaboration with several TACO subprojects.

The talk was given at the TACO Kick-off Meeting in Vienna on September 28, 2021.