Pillars

This pillar forms the theoretical and methodological basis for the activities of RAIS² in research and teaching. It deals with the technologies that make the development of AI methods possible. By networking with the other pillars, new methods are to be developed in direct exchange with potential users.

The focus in the field of artificial intelligence technology at UBT is on optimization methods and the theory of neural networks and kernel-based methods (Schiela, Rambau, Grüne, Wendland), machine learning methods (Christmann and Birke), machine learning for dynamic systems (Koltai), data science and machine learning for large amounts of data (Jablonski, Martens, Mayer), robotics (Henrich), and interactive artificial intelligence (Buschek).

UBT has a strong focus on artificial intelligence in the life sciences. This ranges from the development and application of artificial intelligence methods in protein design (Höcker), microscopy and simulation (Guthe, Müller), sequence analysis (Schmidt), structure prediction and RNA biochemistry (Hennig and Kuhn) to applications in cell and molecular biology (Henkel-Oberländer and Cavalcanti), molecular biophotonics (Krauss), and food analysis (Schwarzinger). Methodological focuses include, for example, large language models in protein design, CNN and U-NET based methods in microscopy, deep neural networks in sequence analysis, and approximate Bayesian computation in the simulation of biological systems.

Artificial intelligence methods are currently revolutionizing all areas of the life sciences, and UBT is making a significant contribution to this. The aim of this pillar is to pool expertise in the identification and development of artificial intelligence methods to answer scientific questions and generate knowledge in the life sciences. This is achieved by creating synergies through collaboration and exchanging methods and expertise between researchers in the fields of AI and the life sciences. To this end, we work with the Bayreuth Center for Molecular Biosciences, the North Bavarian Center for NMR Spectroscopy, the Bayreuth Center for Ecology and Environmental Research and the Medizincampus Oberfranken.

Artificial intelligence and machine learning methods have accelerated and improved materials research in recent years. A particularly interesting aspect in this context is that the data is intrinsically multiscale: The properties of a material depend on its chemical composition (i.e. on the atomic scale), but also on its meso- and macroscopic structure. Accordingly, AI models for materials have very specific requirements and boundary conditions, which often leads to new methodological developments. UBT is broadly positioned here. For example, Bayesian ML methods are used for the discovery of new materials (Ruckdäschel, Margraf, Künneth) or the analysis of impedance spectra of electroactive materials (Ciucci). Neural networks can be used to analyze heat conduction in colloidal materials (Oberhofer, Retsch). On the mesoscale, neural differential equations can be used to predict the formation of boundary layers in batteries (Röder). In polymer and protein research, so-called large language or generative AI (GenAI) models are developed to predict new materials and properties (Künneth, Höcker). Finally, atomistic ML models are used to accelerate quantum mechanical calculations for materials by several orders of magnitude (Oberhofer, Margraf).

The University of Bayreuth offers unique points of contact between the rapidly developing AI and experimental multiscale analysis methods. This infrastructure, which is visible through the Keylabs of the Bavarian Polymer Institute (e.g. Retsch, Gröschel, Schenk), the North Bavarian NMR Center (Senker) and the BayBatt Battery Cell Technology Center (Ciucci, Bianchini), is expected to generate clear synergies. The interdisciplinarity practiced on the UBT campus is a unique selling point with national and international visibility, linking AI in materials science and experimental materials research. This research philosophy is supported by the SFB MultiTrans.

The volume of data is also continuously increasing in business and industry, which correspondingly increased  demand for methods to analyze data efficiently and use artificial intelligence effectively. This pillar therefore brings expertise in the modeling, storage and analysis of structured and unstructured data and the efficient recognition of patterns in data (Agnes Koschmider), the understanding of statistical methods and techniques to detect anomalies in data (Daniel Baier), human-centric AI based on principles of fairness and transparency (Niklas Kühl), the organizational aspects of AI projects (Maximilian Röglinger, Anna-Maria Oberländer) and the application of (generative) AI (Niklas Kühl, Agnes Koschmider) in industry and society.

Innovative solutions for the production and recycling of technical products (remanufacturing) are also being developed to close critical data gaps (smart data acquisition), achieve intelligent value creation using existing data volumes and to efficiently use data-based knowledge (Frank Döpper). The aim is to generate usable knowledge content (lean data) from large and unstructured amounts of data (big data) and make it available for use (Agnes Koschmider, Frank Döpper). The structure and scope of data on the manufacture, composition, use and waste generation of products are also relevant for AI models and are used for research into global material cycles and the recovery of valuable secondary raw materials (Christoph Helbig).

Close cooperation with the Institute for Entrepreneurship and Innovation at UBT promotes innovation and technology transfer in the field of AI in particular (Rebecca Preller, Matthias Baum).

Modern AI systems are increasingly permeating our everyday lives. From spam filters in our emails and facial recognition on our smartphones to medical diagnostics, driver assistance systems and automated reasoning in court. Against this backdrop, we must always ask ourselves what impact the widespread use of AI will have on modern society. Key questions in this context are: How can we ensure that existing norms and values are upheld while benefiting from the use of cutting-edge technology? (Lena Kästner, Johanna Thoma, Mirco Schönfeld) What demands do we as a society place on modern AI systems (Lena Kästner) and under what circumstances may they be used in sensitive areas such as law enforcement (Christian Rückert)? How do we want to shape collaboration between humans and AI in the future (Niklas Kühl, Daniel Buschek)? How can AI be used to analyze and control personal educational paths (Torsten Eymann)? What must effective regulation of AI achieve and how should ethical decisions be formulated and legitimized during development (Christian Rückert, Johanna Thoma, Ruth Janal)? How should the use of generative AI be assessed with regard to data protection regulations (Christoph Krönke, Agnes Koschmider)? How can AI be used for a modern and human-centered public administration (Anna Oberländer)? How can the transparency of AI systems be guaranteed despite existing trade secrets (Ruth Janal)? And how can different requirements for AI from different areas and interest groups be reconciled (Lena Kästner, Anna Oberländer, Niklas Kühl)?

There are already a number of projects and initiatives at the UBT addressing these questions, which are further expanded within RAIS². These include the interdisciplinary working group FATE@UBT (Fairness Accountability Transparency and Ethics in AI) and participation in the Policy Working Group as part of the AI Alliance.

Like other empirically oriented disciplines, environmental sciences and ecology are increasingly making use of powerful artificial intelligence and machine learning tools. The aim is to use the increasing wealth of data to gain knowledge and make predictions and to develop solutions for current problems such as climate change, environmental pollution, habitat loss and species extinction. At UBT, AI tools are used in particular for monitoring, predicting and understanding complex processes and patterns in ecosystems. For example, AI-based recognition methods enable efficient monitoring of environmental processes (Jentsch, Steinbauer) and species and individual characteristics (Higgins, Mair, Laforsch, Löder, Schott) from image, audio and video sources. Machine learning is also used extensively in the field of predicting species distributions (Hülsmann) and environmental risks (Mair) as well as in the downscaling of environmental data sets (Higgins).

This pillar benefits from the established structures and networks in the field of environmental research at the University of Bayreuth and cooperates closely with the Bayreuth Center for Ecology and Environmental Research BayCEER and the BayCEER Keylabs.