Hybrid Modeling and Theory-Driven Data Science

05.08.2020 | Sprache: english

Research and technology trends in Big Data and AI

We are in the process of preparing our fourth COMET application and are looking for partners to jointly explore the potential of these key technologies:
Learn more about the advantages and opportunities of a COMET partnership and secure your access to funded top-level research.

Target Group

Researchers from the manufacturing or pharmaceutical industry


Data-driven models have gone everywhere: Backed by their successes in comparison to recent technological fields such as image processing or speech recognition and powered by ongoing digitization, data-driven models have been taken up in all disciplines. In some of these disciplines, a large amount of domain knowledge is available in the form of (physical) theories – for example, in manufacturing or pharmaceutical industries, this domain knowledge may come in the shape of differential equations representing the behavior of materials or the reactions within a process describing the system under study.

Rather than suggesting data-driven models as a substitute for this theoretical knowledge, the present talk will give an insight into how this knowledge can be incorporated into the data-driven models. The combined – hybrid – model will generally have a larger consistency with the physical world and, due to the use of data, higher accuracy than its theoretical counterpart. We will illustrate the various ways to combine theoretical with data-driven models at the hand of several success stories achieved at the Know-Center.

In the second part of the session, we ask for your input: What is the type of knowledge available in your profession? And for what problems do you currently foresee data-driven models as promising approaches? What are the major roadblocks that you see ahead? Allocating speaker time to all attendees will open the forum for discussing how theory-guided data science can be put to good use in your respective field.

After the event you will know:

The manifold possibilities of incorporating theoretical knowledge into data-driven models, e.g.,

  • How deep learning can solve differential equations

  • How domain knowledge can help in feature engineering

  • How a data-driven approach can improve approximate physical models

Furthermore, you will learn about

  • Previous success stories in theory-driven data science and hybrid modeling

  • The future research direction of Know-Center in this field



Bernhard Geiger

Bernhard Geiger

Knowledge Discovery