ACCEL: Automated Closed-loop Co-Optimization and Experimentation Learning Enables Phase-Pure Identification in Formamidinium-based Dion–Jacobson Halide Perovskites

Self-driving laboratories (SDLs) are poised to transform materials discovery by integrating automation with machine learning (ML) to accelerate data-driven experimentation. However, most SDL frameworks remain limited by single-feedback optimization and lack the multi-modal diagnostics needed to resolve both optical and structural evolution in complex hybrid systems. Here, we introduce ACCEL, an automated, ML-guided closed-loop platform for identifying and synthesizing target pure phases within quasi-2D halide perovskite systems. Using a ternary 3D:2D compositional space comprising 3D FAPbI3 and two Dion–Jacobson (DJ) spacer–based components, ACCEL optimizes the 3D:2D ratios to identify compositions that converge toward a structurally and optically stable α-FAPbI3-like phase within a quasi-2D design space. Automated in-situ photoluminescence (PL) is used to monitor crystallization kinetics in real time, revealing how the relative fraction of co-spacers governs nucleation rate and phase evolution during film formation. These kinetic insights are integrated with high-throughput synthesis, automated PL analysis, and X-ray diffraction (XRD) similarity metrics within a Gaussian Process–Bayesian Optimization framework to suppress phase heterogeneity and stabilize the target phase. Rather than targeting a predefined DJ layer number, ACCEL enables optimization based on user-defined optical and structural signatures, providing a generalizable route for autonomous phase discovery in complex hybrid materials.