Machine-Learning-Guided Design of Semirigid Hole-Selective Self-Assembled Monolayers for High-Efficiency and Stable Perovskite Solar Cells.

Inverted perovskite solar cells (PSCs) require efficient, aggregation-resistant hole-transporting self-assembled monolayers (SAMs) at the buried interface. Conventional carbazole-based SAMs, such as 4PACz, suffer from excessive π-π stacking-driven aggregation, leading to inhomogeneous films, elevated trap-state densities, and accelerated device degradation. Here, we report a machine-learning (ML)-guided design strategy employing a pairwise-ranking Siamese directed message-passing neural network (D-MPNN) trained on a curated data set of published SAM control-target pairs. Statistical analysis of the data set reveals that the reduced tail-group planarity improves PCE relative to carbazole benchmarks. Guided by this insight, we designed two semirigid seven-membered-ring SAMs, 4-(5H-dibenzo[b,f]azepin-5-yl)butyl)phosphonic acid (4PABAP, ML score: +2.54) and 4-(10,11-dihydro-5H-dibenzo[b,f]azepin-5-yl)butyl)phosphonic acid (4PAHBAP, ML score: + 5.52) that the model ranked substantially above 4PACz (score: -0.32). Ab initio molecular dynamics (AIMD) simulations confirm that the nonplanar seven-membered rings suppress π-π stacking, enhance ITO-surface anchoring (interaction energies: 4PAHBAP -1.23 eV vs 4PACz -0.74 eV), and promote stronger charge transfer at the SAM/perovskite interface. Experimentally, 4PAHBAP-based inverted PSCs achieve a champion power conversion efficiency (PCE) of 26.36% with a fill factor (FF) of 84.35% significantly outperforming 4PACz-based controls (23.52%, FF 77.51%). Moreover, 4PAHBAP-based devices retain 92.5% of initial PCE after 1,200 h of continuous 1-sun maximum-power-point (MPP) operation under ISOS-L-3 standards, and 95.7% after 1,200 h thermal aging at 85 °C (ISOS-D-2). This work establishes a generalizable ML-first-then-validate paradigm for rational SAM design that simultaneously maximizes efficiency and long-term operational durability.