**Project Heads**

Claudia Draxl, Sebastian Pokutta, Daniel Walter, Andrea Walther

**Project Members**

Arsen Hnatiuk

**Project Duration**

01.01.2024 − 31.12.2025

**Located at**

HU Berlin

Perovskite materials exhibit exceptional properties for a variety of applications such as electrocatalysis, ferroelectrics and, in particular, photovoltaics. As a consequence, the identification of new Perovskites through the assessment of key material parameters paves the way, e.g., for the next generation of solar cells and thus presents a driving force behind the turnaround of energy policies. These parameters of interest are, for example, given by the band-gap of the Perovskite or the effective mass of its carriers, both of which are tied to the light-harvesting capabilities of the material. Through the advent of big data, a starting point for the search of novel materials is given by the screening of vast collections such as the FAIR repository maintained by FAIRmat, i.e., a consortium of the German Research-Data Infrastructure NFDI, or the archive of the NOMAD Center of Excellence. However, available data on band-gaps is still quite scarce since its acquisition heavily relies on time-consuming experiments or numerical simulations using ab initio methods. These infer the band-gap based on the laws of quantum physics and rely, at their core, on the resolution of some approximation to the underlying Schrödinger equation. Depending on the practical realization, this creates a hierarchy of numerical models with a trade-off between inexpensive but inaccurate methods, e.g., generalized gradient approximation (GGA xc), and very accurate methods, e.g., GGA xc+GW, which, however, come at a significantly higher computational cost. The discrepancy between these time-consuming state-of-the-art methods and the practical demand for a fast-screening procedure serves as a founding stone for the current proposal. More in detail, it motivates the development of a data-driven high-throughput mapping, which, rapidly and accurately, predicts the band-gap and effective masses for a vast amount of candidate materials from cheaply available features. Recent AI-driven advances towards the fast assessment of material parameters rely on the decomposition of the parameter prediction into two simpler substeps. First, a computationally inexpensive feature-acquisition step which takes the “natural” features of the Perovskite, i.e., the position of its nuclei, their charges and number of electrons, and maps them to a space F of atomic and structural properties. Second, an explicit descriptor which describes the material properties as a function of the newly acquired features.

This project aims at the development of a data-driven approach for the computation of such a descriptor. For this purpose, a combination of sparse, nonsmooth regression and active learning is considered.

**Project Webpages
**

**Related Publications
**

**Related Pictures
**