No Time to Waste: Quickly Optimizing Perovskite Composition with Off-the-Shelf Active Learning Methods

The compositional flexibility of halide perovskite presents both opportunity for discovery of tailored materials and risk of extraneous effort. Active learning - a class of machine learning methods that alternates between modeling a response surface and suggesting interesting points to be tested next - enables optimization of a system while testing just a small fraction of the total search space, and can alleviate the cost of large compositional searches. We demonstrate that a single generic active learning algorithm is effective across five typical tasks in halide perovskite development under the three research paradigms of hand-done, high-throughput robotic, and computational experiments. These tasks involve tuning the composition at one, two, or all three of the A,B, and X sites. Moving beyond generic active learning, representing ABX3 compositions by their physical properties at each site – which, in practice, entails simply looking up and averaging tabulated values – further improves the rate of composition optimization by 1-1.4x. The experimental budget of a typical successful active learning optimization was about 15 samples – well within the budget of tedious manual experiments, and a drop in the bucket for high-throughput automated or computational studies. Our findings resonate with the growing body of active learning demonstrations across the scientific literature, and advocate for the integration of active learning into composition optimization efforts in halide perovskites.