Proceedings of MATSUS Fall 2026 Conference (MATSUSFall26)
Publication date: 22nd July 2026
Self-driving laboratories have accelerated materials experimentation by coupling robotic synthesis with Bayesian optimization in closed-loop campaigns.1 These campaigns typically operate within a design space defined a priori by human-specified variables, parameter bounds and objective functions. Such workflows can rapidly identify high-performing candidates within a given domain but do not independently formulate new hypotheses that challenge the underlying assumptions or redefine the scope of investigation. The resulting findings therefore remain constrained by the scientific framework established before experimentation begins.
Here we present an agentic architecture that integrates knowledge mapping, hypothesis generation, experimental testing and analysis in a closed loop with automated experimentation. Distinct agent roles exchange an explicit record of falsifiable claims and evidence, allowing each cycle to revise the hypothesis tested next. New hypotheses can introduce previously excluded variables and redirect experiments toward unexplored compositions or synthesis pathways. Experiments challenge proposed mechanisms, and both supported and contradicted predictions inform subsequent hypotheses. The design space can thus evolve in response to experimental evidence. Across campaigns, this process could reveal principles that connect material behavior to its underlying mechanisms and support predictions beyond the compositions tested. Such principles could open routes to material families and synthesis strategies absent from the original search, extending laboratory autonomy to the formulation and pursuit of new scientific directions.
This work was generously supported by the taxpayers of South Korea through the Institute for Basic Science, project code IBS-R020-D1.
