Low-toxicity solvent processing in ambient air for perovskite solar cells via two-step Bayesian machine learning

The low-cost industrial application of perovskite solar cells requires an environmentally friendly and scalable fabrication process. However, achieving high-quality perovskite layers under these requirements is challenging because the multi-step optimization with multiple intercorrelated experimental variables typically requires the development of a new deposition process. To address this, we propose a two-step machine learning approach for creating a new method for perovskite deposition in ambient air and anti-solvent-free processing with a low-toxicity solvent triethyl phosphate (TEP). The two-step machine learning approach integrates a precursor solubility prediction model and a device-efficiency prediction model within a Bayesian optimization framework. This framework enables the information of solubility to be passed as a constraint function when optimizing the efficiency of perovskite solar cells, facilitating a quick optimization of a TEP-based, vacuum-quenching-assisted deposition in ambient air. Furthermore, the optimal precursor solution is subsequently applied to FAPbI₃ perovskite devices, achieving a device power conversion efficiency of 24.26% under ambient conditions (23 °C and ∼50% relative humidity). This work demonstrates the promising potential of machine learning to expedite new fabrication processes to fulfill industrial needs.