Publication date: 15th December 2025
Emerging photovoltaics, such as organic or perovskite based photovoltaics, can contribute to a more sustainable society. However, marketability of emerging PV still requires improvements of materials and processes towards higher efficiency, better stability, and lower cost. Using domain knowledge driven “trial and error” experimentation, improvements occur only slowly. Artificial Intelligence driven (AI) workflows can replace the trial and error search by systematic sampling of chemical and processing space, which is expected to dramatically increase the speed of discovery of record breaking materials or smart processes. However, AI driven workflows typical require large training datasets, a no-go for experimental science.
In this presentation, I show how the incorporation of physics knowledge allows building of workflows that can learn from less than a hundred different individual experiments. These workflows can be deployed in single labs or small consortia and should thus allow a general acceleration of innovation in our field. Essentially, we use quantum chemical simulations to incorporate quantities into the training dataset that serve as proxies for the properties of interest.
We have used this workflow successfully to discover new organic hole transport molecules for perovskite solar cells[1], to find organic semiconductors with strongly selective light absorption [2], and to identify radiation hard organic semiconductors for space applications and radiation detectors. Finally, I will discuss multi-objective optimizations which are essential for marketability, where several requirements must be matched at the same time.
We acknowledge the Bavarian Initiative “Solar Technologies go Hybrid” (SolTech), as well as funding from Deutsche Forschungsgemeinschaft (project BR-4031/22-1)
