Integrating Experimental Expertise with Adaptive Bayesian Optimization for Perovskite Synthesis

The ball milling synthesis of perovskite materials involves exploring a complex, high-dimensional parameter space, where conventional trial-and-error approaches are inefficient. For novel systems, large-scale prior datasets are often unavailable and may introduce bias if over-relied upon. In perovskite synthesis, the representation of process parameters and precursor descriptors plays a decisive role in the performance of predictive models and optimization strategies. To address this challenge, we propose a machine learning (ML)-guidedBayesian optimization (BO) framework for adaptive experimental design to accelerate the optimization of perovskite synthesis parameters. The framework integrates physicochemical descriptors of precursor elements with ball-milling、 and heat-treatment variables to construct a crystallite-size prediction model, which is embedded into a BO loop to dynamically guide experiments toward target crystallite sizes. This enables systematic control and fine-tuning of crystallite size, representing a performance-oriented reverse design paradigm. Preliminaryresults show that the framework converges to optimal conditions within only a few experimental iterations, significantly outperforming traditional trial-and-error methods. Combining ML with BO effectively reduces the experimental search space, lowers costs, and accelerates materials synthesis. The proposed framework provides a promising pathway for intelligent, data-driven synthesis of perovskites and complex inorganic materials, laying the methodological foundation for future self-driving experimental work.