MATH+ Project AA2-7
Sparse Deep Neuronal Networks for the Design of Solar Energy Materials
The design of new materials for solar cells still relies heavily on very time-consuming material screenings. A simulation-based approach is capable of classifying compounds of Perovskite type with respect to thermodynamic stability in the formation and further properties using density function theory [1]. Likewise, the relationship between key performance indicators can be assessed from experimental data [2]. A major challenge for the fast and reliable design of new solar energy materials is the fact that there are still significant discrepancies between the compound properties predicted by simulation and the experimental data for numerous Perovskite type materials. This motivates the development of new mathematical techniques for improved machine learning approaches targeted within this project.
Project partner:
Thomas Kühne, Dynamics of Condensed Matter, Universität Paderborn
Current activities:
- Andrea Walther and Franz Bethke together with R. Varadarajan, S. Jost, H. Mirhosseini, T. Kühne und C. Plessl submitted an article on the predictions of formation energies with machine learning.
- Andrea Walther and Franz Bethke submitted an article on semismooth conjugate gradient methods.
Past activities:
- The project was represented at the GAMM annual meeting in Aachen via a presentation on "The Semismooth Conjugate Graident Method" given by Franz Bethke. Among the conceptual algorithm the talk showed first theoretical results and numerical applications in image denoising.
- Franz Bethke participated in the ALOP Workshop on Algorithmic Optimization and Data Science in Trier.
- Franz Bethke participated in the Software Engineering 2022. This year the conference hosted multiple workshops and a session tract dedicated towards research software engineering! The workshop "Testing Neural Applications" by Markus Goetz and Peter Steinbach and the talk "Fuzzing Computational Material Science Code Parsers" by Sebastian Müller and Jan Arne Sparka represent strong relations with AA2-7.
- Sebastian Jost completed his bachelor thesis "Comparison of architectures and parameters for artificial neural networks". His results could improve our the existing premilinary studies on training a neural network for predicting the stability of chemical compositions. His code can found here.
- Poster presentation with the topic "Combining the ADMM and Active Signature Methods for the Training of Neural Networks with Nonsmooth Activation Functions" at COMinDS YoungResearchers' Seminar by Franz Bethke on Mai 28. 2021
- Kickoff meeting with all PIs and project partners in March 2021
- Franz Bethke visits group of Thomas Kühne in August 2020
References:
[1] R. Raghupathy, H. Wiebeler, T. Kühne, C. Felser, and H. Mirhosseini. Database screening of ternary chalcogenides for p-type transparent conductors. Chemistry of Materials, 30(19):6794–6800, 2018.
[2] E. Unger, L. Kegelmann, K. Suchan, D. Sörell, L. Korte, and S. Albrecht. Roadmap and roadblocks for the band gap tunability of metal halide perovskites. Journal of Materials Chemistry A, 5:11401–11409, 2017.
