A Self-Driving Lab Autonomously Learns to Accurately Dispense Samples via Dynamic Experimental Control

Self-driving labs – automated experiments guided by machine learning (ML) algorithms – have the potential to accelerate discovery of new molecules, materials and processes across a range of disciplines. They can enable researchers to think about their experiments in the context of trends, datasets, and other higher level questions rather than individual experiments by minimizing the necessity for manually performing routine or repetitive tasks.<br/> <br/>A central challenge of self-driving labs is building robotic systems capable of carrying out the same wide range of actions and experiments as humans. For example, accurately dispensing both large and very small quantities of material. Currently, this requires careful calibration and testing to determine the correct process parameters for each material, and substantially eats into valuable researcher time. It would therefore be desirable for a robotic systems to be able to teach itself the optimal process parameters autonomously. This would reduce the necessity for human involvement in routine tasks, and boost the efficiency of self-driving labs.<br/> <br/>In this talk, I will detail our work toward developing a fully autonomous synthesis platform, enabled by a closed loop of Bayesian optimization and automated experiments, that can optimize its own instrument parameter to accurately dose samples, and learn from its previous experience with optimizing the parameters for other samples in order to further speed up its learning rate. Such systems have the potential to significantly improve the capabilities of self-driving labs, the development of which is often limited by the need to adapt them to each new material.