Noise Conditioned Bayesian Optimization Strategy Selection for Self-Driving Labs

Abstract

Self-driving labs have become a widespread paradigm for automating research and development processes across disciplines such as biology, chemistry, and pharmaceutical research. Such systems combine robotic hardware and software to iteratively optimize experimental outcomes in a closed loop. Noise inherent in sensors and robotic actuation not only affects the speed at which the optimal result is found, but also which optimization strategies perform best. We present a simulation-based methodology to analyze selfdriving lab loops and to select the optimal acquisition function and hyperparameters for Bayesian optimization by incorporating known uncertainty in realistic simulated system behavior. Our results indicate that optimizer selection informed by noise characteristics can significantly reduce the number of required iterations.