Regensburg 2025 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
HL: Fachverband Halbleiterphysik
HL 3: Focus Session: Machine Learning of semiconductor properties and spectra
HL 3.2: Hauptvortrag
Montag, 17. März 2025, 10:00–10:30, H17
Generative Models on the Rise - Which one shall I pick for my Inverse Design Problem? — •Hanna Türk1,2, Elisabetta Landini2, Christian Kunkel2, Patricia König2, Christoph Scheurer2, Karsten Reuter2, and Johannes Margraf2,3 — 1EPFL, Lausanne, Switzerland — 2Fritz-Haber-Institut der MPG, Berlin, Germany — 3Universität Bayreuth, Bayreuth, Germany
The pursuit of novel materials through computational discovery appears endless due to the vast space of potential structures and compositions. For inorganic materials, this complexity is heightened by the combinatorial possibilities presented by the periodic table, where even a single-crystal structure can theoretically exhibit millions of compositions.
Recently, generative machine learning models have emerged as method for direct exploration of the material design space. Here, we evaluate the efficacy of various conditioned deep generative models, including reinforcement learning, variational autoencoders, and generative adversarial networks, in the prototypical task of designing Elpasolite compositions with low formation energies. Utilizing the fully enumerated space of 2 million main-group Elpasolites, we rigorously assess the precision, coverage, and diversity of the generated materials. Furthermore, we develop a hyperparameter selection scheme tailored for generative models in chemical composition space. Finally, we demonstrate the power of these machine learning models on a realistic application.
[1] Chem. Mater. 2022, 34, 9455-9467.
Keywords: Generative Models; Inverse Design; Machine Learning