The urgent need to transition to sustainable energy sources demands accelerating traditional research and development cycles. Self-driving labs (SDLs) powered by artificial intelligence (AI) could play a pivotal role in this transformation.

In a perspective paper in the journal Nature Catalysis, researchers from the Fritz Haber Institute's Theory Department discuss the role humans will play in the future of self-driving labs for catalysis research.

A self-driving laboratory integrates AI with lab automation and robotics. The AI plans experiments executed in increasingly automated (robotized) modules. In practice, this process occurs in active-learning loops, where the data from the last loop is used to define a machine-learning model. The AI then uses this model to plan the subsequent experiments in the next loop. This way, only those syntheses, characterizations, and most informative tests are conducted based on all prior collected data. Simultaneously, the automation enhances throughput, reproducibility, and safety—promising a significant acceleration compared to traditional human-led development processes.

Early implementations of this concept for discovering improved catalyst materials often focus on replacing human tasks with synthesis robots. Researchers Christoph Scheurer and Karsten Reuter emphasize that the most time-consuming step of such catalysis research is typically the explicit testing of the materials. Given the increasing importance of sustainability, the degradation behaviour of the materials in the reactor must be monitored over a long time. Therefore, throughput improvements are more likely to be achieved by developing new testing procedures specifically designed for SDLs rather than merely automating existing procedures.

Especially when throughput remains limited, AI's role in experiment planning is crucial. The fewer loops that need to be executed, the better. Also, humans will continue to play a vital role for the foreseeable future. While current AIs can determine optimal experiments within a given overall framework, they cannot yet question this framework or even redefine the scientific questions themselves. For the time being, these creative tasks remain the domain of humans, necessitating a human control function within the loops.

The authors thus advocate the "human-in-the-loop" principle and analyze its implications for AI development in SDLs. Not least, AIs must be capable of responding flexibly, robustly, and accessibly to human modifications of the loop structures—a methodological challenge already addressed by ongoing research in the Theory Department.