we apply data-driven techniques to discover materials that work in the real world. through novel machine learning approaches and close collaboration with experimental partners, we bridge the gap between computational predictions and practical applications.
our group is funded by the carl-zeiss foundation as the “czs research group polymers in energy applications”. we are part of the friedrich-schiller-university jena and the helmholz institute for polymers in energy applications jena.
other major funding sources are merck and intel (awases research center), openphilanthropy, the helmholtz foundation model initiative (sol-ai project), the german research foundation (dfg, via the graduate school “coin”), and the carl zeiss foundation (via other projects such as nano@liver).
research
we develop actionable, data-driven solutions for real-world materials design challenges across multiple scales. our machine learning models serve as navigation systems for the chemical space, created in partnership with experimental collaborators.
we believe computational work without open code is mere advertisement.
leveraging tacit knowledge
while abundant chemical data exists in publications, most remains unutilized. laboratory experiments often proceed without optimal integration of existing scientific knowledge - sometimes even disconnected from previous work within the same group.
machine learning techniques, especially large language models, help us unlock and access this valuable information. they capture subtle, tacit aspects of chemistry that conventional machine learning approaches - limited to “idealized” structural representations - cannot grasp.
better inductive biases and representations
we know many things about the world and take basic physical and chemical laws for granted. yet most of our models remain unaware of these fundamental principles.
to enhance model robustness and predictive power, we develop novel ways to incorporate relevant inductive biases. much of this work focuses on model inputs - crafting representations of molecules and materials that faithfully carry more relevant information. this includes developing representations that bridge length scales.
learning beyond conventional objectives
most models are trained by minimizing the mismatch between predictions and ground truth. while this approach might yield good predictions that correlate with “ground truth” in specific cases, it doesn’t guarantee that models are right for the right reasons.
to address this challenge, we incorporate more than quantitative error feedback into our model training. we leverage principles from human-computer interaction to ensure our models learn in more comprehensive ways.