Artificial Intelligence-Guided Cosolvent Design for High-Performance Perovskite/Silicon Tandem Solar Cells

Abstract Realizing high-performance perovskite/silicon tandem solar cells requires precise control of wide-bandgap perovskite crystallization. Solvent engineering is the most direct lever for this task; yet, its intricate, multi-variable mechanisms defy intuition-driven design. Herein, we overcome this bottleneck by pioneering a retrieval-augmented large language model to screen > 8000 solvents, identifying γ -valerolactone (GVL) as a non-toxic, high-performance cosolvent. It is found that the GVL strongly coordinates FA + , thus precisely modulating crystallization kinetics, retarding nucleation, and promoting oriented, micrometer-scale grain growth. The resulting films exhibit not only superior crystallinity, reduced non-radiative recombination, but also improved scalability to large area and the tolerance to increased film thickness. Consequently, both the single-junction and tandem devices achieve efficiencies of 23.3% and 32.5%, respectively, along with excellent stability under moisture and illumination. This study establishes the first artificial intelligence (AI)-guided cosolvent strategy for 1-μm-thick perovskite layers in perovskite/silicon tandem architectures, underscoring the transformative role of generative AI in advancing high-performance photovoltaics.