Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 7: Emerging Topics in Chemical and Polymer Physics, New Instruments and Methods I
CPP 7.5: Talk
Monday, March 9, 2026, 12:30–12:45, ZEU/0255
Property-guided diffusion modeling for efficient exploration of chemical spaces — •Leonardo Medrano Sandonas1, Michael Hanna1, Julian Cremer2, and Gianaurelio Cuniberti1 — 1TUD Dresden University of Technology, Germany. — 2Pfizer Worldwide R&D, Germany.
The rational in silico design of chemical compounds requires a deep understanding of both structure-property and property-property relationships across chemical compound space, as well as efficient methodologies for defining inverse property-to-structure mappings. In this presentation, I will discuss our recent efforts to leverage the "freedom of design" concept [Chem. Sci. 14, 10702 (2023)] in the chemical space of drug-like molecules to develop an efficient generative AI framework capable of designing molecular compounds with targeted quantum-mechanical (QM) properties. To this end, we have implemented a property-guided active learning approach that optimizes the performance of equivariant diffusion models within each property space. Generated molecules and their associated QM properties are validated through exhaustive DFT calculations at the PBE0+MBD level using the FHI-aims code. Our findings reveal a consistent improvement in property accuracy when guiding the diffusion model to explore less populated regions of the target property space. This performance is further enhanced by incorporating molecular building blocks into the initial training set. We expect our work to advance the development of sustainable generative AI frameworks for identifying molecules tailored to specific chemical processes.
Keywords: Generative modeling; Molecular design; Quantum mechanics; Property prediction; Chemical space