Machine Learning-Based Process Optimization for Highly Uniform Flexography-Printed Perovskite Thin Films

Developing a reliable and scalable fabrication technique is one of the most critical challenges for perovskite solar cells to achieve commercialization at terawatt-scale photovoltaic capacity. However, current methods for optimizing the processing parameters heavily rely on traditional manual screening, which is time-consuming and labor-intensive due to the large number of closely correlated variables. In this work, we leverage sequential machine learning (ML) methods to simultaneously optimize perovskite ink properties and printing parameters, aiming to maximize the optoelectronic uniformity of printed perovskite thin films. The perovskite films are fabricated using high-speed flexography (60 m/min), and spatial uniformity is quantified through scanning photoluminescence (PL) characterization. After five rounds of Bayesian optimization (BO) iterations, we achieve continuous improvement in film uniformity while reducing low-quality prints that are unsuitable for device fabrication. As a result, we demonstrated highly uniform perovskite films with a low coefficient of variance (COV) of < 4.1% in PL intensity and <0.26% variance in bandgap, resulting in solar cells with high photovoltaic power conversion efficiency (PCE) > 19.8%. This work provides valuable insights into how fluid dynamic properties influence the printing process and affect the final film uniformity, helping to overcome bottlenecks in the development of perovskite photovoltaic technology and advance its commercialization.