| Time: | November 6, 2024, 2:00 p.m. – 4:00 p.m. |
|---|---|
| Download as iCal: |
|
The Special Interest Group Data Infrastructure provides a forum for interested working groups wishing to establish or further develop an RDM infrastructure at working group or institute level. We invite you to a monthly SIGDIUS seminar, to which we invite internal and external experts for presentations and discussions. SIGDIUS members will have the opportunity to exchange their experiences with concrete RDM infrastructures.
We cordially invite all interested parties to our next meeting on 6 November 2024 at 2:00 p.m. For participation, please send an e-mail to Juergen.Pleiss@itb.uni-stuttgart.de.
This seminar will be held as an online seminar with talks from:
Alessandro Butté
DataHow, Zürich
Optimal design of experiments for hybrid models
Most Biopharmaceutical Companies are currently considering and adopting digitalization as a tool to improve their efficiency and competitiveness. This is particularly evident in process development and in manufacturing process operation. Here digitalization is best implemented through the adoption of Digital Twins, which are centered on a representative model of the process. Hybrid models are ideal for this purpose. They combine mechanistic models, which include available consolidated process knowledge, which allows reducing the experimental effort needed for their calibration, and data driven models, which allow describing portions of the process that need to be learned as a flexible function of the specific characteristics of this process. The machine learning component is also enabling knowledge transfer, i.e., the promising capability of transferring process knowledge across different products, resins, or scales. In other words, by a proper and systematic use of process data, it is possible to further reduce the experimental effort to develop, scale-up, and validate new processes. In this context, relatively little attention is dedicated to the topic of optimal experimental designs when using hybrid models. While experimental design strategies like factorial designs are the state-of-the-art in the industry, they might result in highly suboptimal designs when used to develop processes and when using ML-based models. We will demonstrate that, through the use of Bayesian-based optimization, it is possible to progressively focus the experimental activities in the portion of the design space of the process parameters that it is mostly of interest for the process, thus simultaneously optimizing the process and increasing process knowledge close to the optimum. Several examples of industrial relevance from different processes will be discussed to demonstrate the advantages of Bayesian-based optimization and the role that adaptive models play in such procedure.
Alessandro Butté received his MSc. at Politecnico di Milano (Italy) and his Ph.D. at ETH Zurich (Switzerland) in Chemical Engineering. After a postdoc at the Georgia Institute of Technology (USA), he joined ETH Zurich as a senior scientist. In 2008, he joined Lonza as head of downstream technologies in the sectors of small molecules and peptides and, later, as project manager. He was also involved in the pilot program to introduce Quality by Design. In 2013, he rejoined ETH as a lecturer, and in 2017, he co-founded DataHow AG, where he is serving as CEO. He is the author of more than 90 papers in international peer-reviewed journals and several patents. In 2015 he completed an executive MBA at the University of St. Gallen.
Anett Seeland & Bernd Flemisch
TIK/SimTech, University of Stuttgart
Jupyter Services for Researchers at the University of Stuttgart
The talk will give an overview of current and future possibilities to get Jupyter Servers for researchers of the University of Stuttgart and their project partners. Thereby, the presentation will focus on services that are being developed as part of the NFDI4Ing and Jupyter4NFDI projects.
