Closed-Loop Transfer Enables AI to Yield Chemical Knowledge

AI-guided closed-loop experimentation has recently emerged as a powerful method to optimize objective functions, yet its potential to generate new scientific insights remains underexplored. In this talk, I will discuss the integration of closed-loop experiments with physics-based feature selection and supervised learning, denoted as Closed-Loop Transfer (CLT), to simultaneously generate chemical insights and optimize objective functions. Our work shows that CLT generated a robust predictive model for the photostability of light-harvesting oligomers, revealing that high-energy regions of the triplet state manifold influence molecular photostability across a diverse library of light-harvesting donor-bridge-acceptor oligomers, which challenges the conventional focus on the lowest energy triplet state. This pivotal insight emerged after automated modular synthesis and experimental characterization of only ~1.5% of the theoretical chemical space. The model’s reliability was enhanced by multiple experimental test sets and validated by adjusting triplet state energies to surpass photostability plateaus. After completing the CLT campaign, we designed a new set of molecular building blocks to control photostability by covalently linking the triplet state quencher (TSQ) cyclooctatetraene (COT) to a thiophene block. Our results show that the photostability of light-harvesting oligomers is enhanced when COT is non-specifically added to solution, but enhanced rates of photodegradation were observed for light-harvesting oligomers containing covalently linked COT, which was fully consistent with our predictive models based on density functional theory (DFT) calculations of triplet state energies for these molecules. Importantly, this outcome reinforced our main findings from the CLT campaign that COT’s high-energy regions within the triplet state manifold significantly contribute to photodegradation in light-harvesting molecules. Motivated by these insights from the model validation phase, we are now designing new molecular building blocks based on redox-mediated TSQs such as anthraquinone. Current efforts are focused on using DFT to evaluate the thermodynamic viability of quenching reactions between TSQ candidates and light-harvesting molecules, and promising candidates are being synthesized and characterized in terms of solution-state and solid-phase photostability. Overall, these results demonstrate that combining physics-based modeling with closed-loop discovery campaigns and automated synthesis methods holds the potential to accelerate the discovery of new materials with advanced function.