Machine learning-assisted prediction of optical properties of solids

• Applicant: Professor Dr. Erich Runge
• Subject Area: Theoretical Condensed Matter Physics
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 537033066

While in many areas of materials research the paradigm of machine learning (ML) is very successful, this does not yet apply to the design of materials for optical applications, despite its great technical importance. The main reason for the low ML efforts concerning optical properties are immense difficulties in the generation of sufficiently large and robust data sets of optical properties. This is especially true for the numerical determination of the frequency-dependent dielectric function [epsilon (omega)] as “the” central optical property. For its numerical determination, a hierarchy of more and more reliable, but also more and more expensive methods exist. One could speak of a "Jacob's ladder of the calculation of optical properties". This ranges from the simple approximation of independent particles via the linear response version of the time-dependent density functional theory (DFT) to the solution of the Bethe-Salpeter equation. Based on extensive calculations on all levels of Jacob's ladder, predictors are to be developed that predict for large material classes, based on the results of one level, whether it is necessary or not to climb the next, significantly more resource-intensive level. Rules of thumb and heuristics existing in the literature and the community, for which material classes which hierarchy levels are necessary and where the weak points of different methods are, will be evaluated and quantified in this project. Equally and in parallel to the DFT-based work packages, ML methods and especially neural networks (NN) will be used to develop ever better ML-based and hybrid predictors based on the growing project database of optical properties calculated at different levels. A particular challenge is to find ML-suitable encodings of the dielectric function. Both subprojects complement each other, since, for example, according to so-called active learning, such systems are calculated on exactly those Jacobian ladder levels with respect to which the ML-assisted predictions are particularly uncertain. The final goal is, besides a deeper understanding of the calculation of optical properties and the further developments of ML techniques, a very large database of optical properties. Regarding the actual computed systems, a collaboration with one or more of the large databases (e.g. NOMAD, Materials Project) is envisaged. In addition, ab initio-trained NNs are to create a much larger database of "only" predicted optical properties. This should be searchable in the future for the realization of suitable optical functionalities. Materials that according to the predictions should have specific desired properties could then be validated by the user through synthesis or "expensive" more precise calculations and ideally be used technically.

• DFG Programme: Research Grants