The dependence on trial-and-error approaches and human expertise is a persistent barrier in the commercialization of perovskite photovoltaics, making materials discovery and device fabrication inefficient and hard to reproduce. Now, writing in Nature, Samuel Stranks, Xiao Cheng Zeng, Zonglong Zhu and collaborators demonstrate a fully integrated, closed-loop system that links machine-learning-guided molecular design with robotic device fabrication, enabling both the discovery of new passivation molecules and the reproducible manufacture of high-efficiency perovskite solar cells.

Perovskite devices were then manufactured by a robotic fabrication platform using Bayesian optimization. The system iteratively adjusts processing parameters to maximize power-conversion efficiency. Devices incorporating the best-performing molecule identified in the previous step, 5-(aminomethyl)nicotinonitrile hydroiodide (5ANI), reached a power conversion efficiency of 27.22% in small-area cells and 23.49% in mini-modules. “The discovery of the passivation molecule 5ANI proves that artificial intelligence is far more than just an ‘accelerator’ for experiments,” comments Danpeng Gao, first author of the study. “It possesses a sophisticated analytical capability that allows it to navigate vast chemical spaces and identify high-performance materials that traditional trial-and-error methods might easily overlook.” Notably, automated fabrication resulted in a nearly fivefold improvement in reproducibility. Stability tests showed only 1.3% degradation after 1,200 h of continuous operation.