Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 83: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 4
O 83.2: Talk
Friday, September 9, 2022, 10:45–11:00, S054
Stacking the odds: Distribution-biased generative deep learning for molecular design — •Joe Gilkes1,2, Julia Westermayr1, Rhyan Barrett3, and Reinhard J. Maurer1 — 1Department of Chemistry, University of Warwick, UK — 2HetSys CDT, University of Warwick, UK — 3Warwick Mathematics Institute, UK
Organic electronics applications pose a number of often competing requirements on molecular design that are hard to satisfy by conventional synthesis. Devices such as organic light-emitting diodes (OLEDs) must exhibit closely aligned optoelectronic properties, yet their component molecules must be easily synthesisable and stable. The odds of finding suitable molecules when drawing random samples from chemical space are still too low for targeted design of candidate systems for OLED devices. We develop an automated molecular design approach based on iterative biasing of a generative deep learning model. In successive iterations, the output of this model is filtered with a deep learning surrogate model of electronic structure and then used to retrain the generative model with a bias. This enables us to create models that are progressively biased towards, e.g., higher ionisation potentials, or smaller fundamental gaps. We also demonstrate how we can bias towards multiple properties simultaneously by filtering our results with the SCScore model for synthetic complexity. This creates more synthetically viable molecules while still meeting optoelectronic requirements. Our approach efficiently creates novel molecules with tuned optoelectronic properties. Clustering analysis reveals trends in bonding patterns which can be utilised in molecular design.