SKM 2023 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 6: New Methods: Experiments and Theory
O 6.9: Talk
Monday, March 27, 2023, 12:30–12:45, GER 39
Designing Covalent Organic Frameworks Through Active Machine Learning — •Yuxuan Yao1,2, Christian Kunkel3, Karsten Reuter3, and Harald Oberhofer2 — 1Chair for Theoretical Chemistry, Technische Universität München — 2Chair for Theoretical Physics VII, University of Bayreuth — 3Fritz-Haber-Institut der Max-Planck-Gesellschaft
Covalent organic frameworks (COFs) are a class of materials, that are formed by molecular building blocks (BBs) connected with covalent bonds. They have found application in many fields such as catalysis, or optoelectronics. It is well known that their design space is far too large to sample one by one because numerous available BBs. We modify an earlier active machine learning (AML) approach that explores the massive available BBs space through the use of surrogate models for charge injection and transport descriptors. In this method, the Gaussian Process Regression(GPR) and AML are combined to train the molecular space. This way we ensure that only promising molecules that are very different from already explored ones have their descriptors evaluated by Density Functional based Tight Binding (DFTB) calculations. Specifically, we modify molecular generation rules in order to produce three-fold rotationally symmetric candidates molecules for use in hexagonal COFs. In the future this approach can be generalized for any other symmetries, to potentially even allow for 3-dimensional network generation. We gauge the performance of our AML by evaluating the success ratio, which is the ratio of promising candidates to all generated molecules.