Active Learning for the Hybrid Perovskite Discovery—Co-Navigating the Literature and Experimental Synthesis

Hybrid perovskites has emerged as one of the most fascinating materials classes, attracting broad attention due to their applications in photovoltaic, optoelectronic, quantum optics, and many other applications. The fundamental bottleneck for exploring HP now is the enormity of the corresponding search space and strong dependence of functionalities on the processing parameters. Even for the 3D endmembers the substitutions on all three sublattices create the enormous combinatorial space of solid solutions. For the 2D perovskites, materials discovery becomes a problem of joint navigation of vast compositional spaces of inorganic matrix and small molecule spacers, further complicated by the strong dependence of phase evolution and stability on synthesis parameters (solvents, antisolvents, temperature, etc). Given the rapidly growing volume of literature in this field, complexity of the coupled chemical, microstructural, and physical phenomena in perovskites, and multiple communities with often dissimilar languages exploring these materials, building comprehensive picture of *<b>known</b>* phenomena in these materials are a daunting task. While the use of LLMs can in principle allow for comprehensive analysis, the associated costs are well outside of a single group. Here, we present an active learning framework for co-navigating the literature and planning experiments on hybrid perovskite synthesis. The ML agent is used dynamically with human in the loop to rapidly assess the literature and build the cause-and-effect relationships. Based on these, the new suggestions for papers analyze based on abstract are formulated as dynamic exploratory loop. The automated synthesis and rapid characterization of optical and photovoltaic properties form the experimental loop run in tandem. The first results of this dynamic co-scientist approach will be presented.