Accelerated Chemical Reaction Optimization using Multi-Task Learning
Event details
| Date | 14.11.2023 |
| Hour | 14:00 › 15:00 |
| Speaker | Kobi Felton is a chemical engineer interested in solving problems at the intersection of chemical engineering, chemistry and software. He holds a Bachelor of Science in Chemical Engineering from North Carolina State University and a MPhil Research and PhD in Chemical Engineering from the University of Cambridge, where he was a recipient of the Marshall Scholarship and the Cambridge-Marshall PhD Scholarship. His current projects include: Optimization of distillation control systems, Self-optimization of reactions using bayesian algorithms, Design of experiments for time-series datasets, and Designing descriptors for chemical reactions. |
| Location | Online |
| Category | Conferences - Seminars |
| Event Language | English |
Reaction optimization is a standard challenge in process development that is often time- and labor-intensive. Recent work has shown that Bayesian optimization is an effective method for identifying optimal conditions, yet it can still be experimentally and materially expensive for reactions with categorical selections (e.g., catalyst, base). Ideally, it would be possible to utilize the vast amount of data available in electronic lab notebooks to speed up optimization. In this talk, we demonstrate that multi-task Bayesian optimization can accelerate reaction optimization by leveraging historical data. We first conduct simulations to investigate the effectiveness of multi-task Bayesian optimization, showing that it can reduce the number of experiments required to find optimal conditions by up to 80%. Furthermore, we demonstrate the effectiveness of multi-task Bayesian optimization on an experimental case study of several related C-H activation reactions run in an automated flow chemistry reactor. Finally, we study several mechanisms by which multi-task Bayesian optimization can fail and propose mitigations.
Practical information
- General public
- Free
Organizer
- Andres M Bran, Rebecca Neeser, Yannick Calvino, Philippe Schwaller
Contact
- Andres M Bran, Rebecca Neeser, Yannick Calvino, Philippe Schwaller