Parts | Days | Selection | Search | Updates | Downloads | Help
KFM: Fachverband Kristalline Festkörper und deren Mikrostruktur
KFM 31: Perovskite and Photovoltaics III (joint session HL/KFM)
KFM 31.1: Talk
Friday, March 22, 2024, 09:30–09:45, EW 203
Data-efficient machine learning for perovksite alloys — •Jarno Laakso, Henrietta Homm, and Patrick Rinke — Department of Applied Physics, Aalto University, Espoo, Finland
Perovskite solar cells are highly efficient, but their commercialization has been hindered by toxicity and lack of stability. Compositional engineering can mitigate these problems [1], but the complexity of the perovskite materials space makes the search for an optimal solar cell material challenging. We previously demonstrated how machine learning (ML) can accelerate property predictions for the perovskite alloy CsPb(Cl/Br)3 [2]. However, the extensive density functional theory (DFT) calculations required for model training prevent applications to more complex materials. Here, we facilitate model training with a data-efficient scheme, validated on CsPb(Cl/Br)3 data and extended to the ternary alloy CsSn(Cl/Br/I)3.
Our approach employs clustering to build a compact but diverse initial data set of atomic structures. We then apply a two-stage active learning approach to first improve the robustness of the ML-based structure relaxations and then fine-tune the accuracy near equilibrium structures. Tests for CsPb(Cl/Br)3 reveal that our scheme reduces the number of required DFT calculations during model training by up to 50%. The fitted model for CsSn(Cl/Br/I)3 is robust, with all ML-based structure relaxations converging in our tests. The relaxations are also highly accurate, having an average error of 0.5 meV/atom.
[1] iScience 23, 101359 (2020). [2] Phys. Rev. Mater. 6, 113801 (2022).
Keywords: machine learning; perovskites; alloys