Materials, Medicines and Model Tuning - Automating Experimental Design with Bayesian Optimisation

Topic: Materials, Medicines and Model Tuning - Automating Experimental Design with Bayesian Optimisation

When: 17 April at 12.00-13.15

Where: Online

Speaker: Carl Hvarfner, PhD Student, Computer Science, Lund University

Spoken language: English

Abstract

Materials science, pharmaceuticals, robotics and Machine Learning all contain resource-intensive design processes. The efficiency of experimental design has emerged as a critical bottleneck, which necessitates innovative approaches to streamline the discovery process. This seminar introduces Bayesian Optimization (BO) as a powerful strategy for automating experimental design, highlighting its potential to automate the way we explore and exploit complex search spaces in search of optimal solutions. By sequentially updating a Bayesian model based on empirical data, BO offers a principled approach to navigating the trade-off between exploration of untested configurations and exploitation of known promising areas. Through use cases in product design and ML hyperparameter optimization, we demonstrate how BO can significantly reduce the number of experiments needed to identify optimal settings and designs. Lastly, we briefly delve into the nuances that determine the efficiency of the BO algorithm.