Navigating to Islands of Photostability—Multi-Objective Optimization of Perovskite Absorber Compositions for Targeted Photovoltaic Applications Using High-Throughput Robotic Experimentation

The compositional flexibility of halide perovskites presents an opportunity to tailor-make materials for a variety of optoelectronic applications. In photovoltaics alone, perovskite absorbers carry promise in single-junction and wide and low bandgap tandem configurations. However, this flexibility brings with it a high cost of exploration and optimization. Design of an effective absorber is further complicated when all figures of merit, including optoelectronic quality, film morphology, and intrinsic material stability, are considered. To approach this task, we combine PASCAL (Perovskite Automated Spin-Coating Assembly Line), a robotic platform for automated deposition and processing of spin-coated perovskite films, with multi-objective active learning to make tractable compositional search, effectively reducing cost from two sides - increasing experimental throughput, and efficiently traversing the search space with machine learning. We demonstrate compositional optimization of perovskite films in multiple campaigns, each targeting applications as either a single-junction, top-tandem, or bottom-tandem absorber.<br/>PASCAL combines a liquid handler, high-throughput spin-coating (&gt;10^2 samples/day), and in-line characterization. This characterization includes darkfield and brightfield imaging to provide metrics for film morphology and coverage, transmission spectroscopy for material bandgap, and photoluminescence spectroscopy for radiative recombination rate and photostability of the absorber. After an initial grid-sampling study, multi-objective Bayesian Optimization is applied in batches to optimize absorber composition and process variables. Pareto-optimal compositions are identified to co-optimize target bandgap, film quality, photoluminescent yield, and photostability.<br/>We first revisit the well-known “kitchen sink” (MA/FA/Cs)-Pb-(I/Br/Cl) space and confirm that PASCAL converges to previously discovered favorable compositions. At the same time, the fine-grained search provides new clarity on tradeoffs in photostability and quantum yield. We then apply the same experimental approach to discover photostable absorbers with bandgaps for applications in wide and low bandgap tandem cell configurations. The volume and homoscedasticity of datasets generated by PASCAL enables faithful and holistic evaluation of compositional levers on material performance. Overall, our work indicates the clear value of active learning and high-throughput search in realizing the opportunities presented by perovskites across optoelectronic applications.