Mixing cations has been a successful strategy in perovskite synthesis by solution-processing, delivering improvements in the thermodynamic stability as well as in the lattice parameter control. Unfortunately, the relation between a given cation mixture and the associated structural deformation is not well-established, a fact that hinders an adequate identification of the optimum chemical compositions. Such difficulty arises since local distortion and microscopic disorder influence structural stability and also determine phase segregation. Hence, the search for an optimum composition is currently based on experimental trial and error, a tedious and high-cost process. Here, we report on a machine-learning-reinforced cubic-phase-perovskite stability predictor that has been constructed over an extensive dataset of first-principles calculations. Such a predictor allows us to determine the cubic phase stability at a given cation mixture regardless of the various cations’ pair and concentration, even assessing very dilute concentrations, a notoriously challenging task for first-principles calculations. In particular, we construct machine learning models, predicting multiple target quantities such as the enthalpy of mixing and various octahedral distortions. It is then the combination of these targets that guide the laboratory synthesis. Our theoretical analysis is also validated by the experimental synthesis and characterization of methylammonium–dimethylammonium-mixed perovskite thin films, demonstrating the ability of the stability predictor to drive the chemical design of this class of materials.

混合阳离子一直是通过溶液处理合成钙钛矿的成功策略，可改善热力学稳定性以及晶格参数控制。不幸的是，给定的阳离子混合物与相关结构变形之间的关系尚未确定，这一事实阻碍了对最佳化学成分的充分鉴定。由于局部变形和微观无序会影响结构稳定性并决定相分离，因此会出现这种困难。因此，对最佳组合物的搜索目前基于实验试错，这是一个繁琐且成本高的过程。在这里，我们报告了机器学习增强的立方相钙钛矿稳定性预测器，该预测器是在第一性原理计算的广泛数据集上构建的。这样的预测器使我们能够确定给定阳离子混合物的立方相稳定性，而不管各种阳离子的对和浓度如何，甚至评估非常稀的浓度，这是第一性原理计算的一项众所周知的具有挑战性的任务。特别是，我们构建了机器学习模型，预测多个目标量，例如混合焓和各种八面体畸变。然后是这些目标的组合指导实验室合成。我们的理论分析也通过甲基铵-二甲基铵混合钙钛矿薄膜的实验合成和表征得到验证，证明了稳定性预测器驱动此类材料化学设计的能力。