Applying machine learning (ML) methods to accelerate the search for new materials with improved properties has gained increasing attention in recent years. Using nonadaptive ML approaches that do not have an iterative feedback loop can perform poorly in extrapolations at previously unexplored search space, especially when trained on small data sets. We performed numerical simulations on two data sets that exhibit distinct composition–property relationships and explored the relative efficacies of adaptive ML strategies in identifying the optimal material composition with the highest. Adaptive ML methods show promise for extrapolation and find compositions with properties better than those in the training data, but the rate of discovery is dictated by the nuances of the composition–property landscape. The outcome of this work has key implications in developing strategies that employ ML methods for navigating a vast search space of combinatorial possibilities.

近年来，应用机器学习（ML）方法来加速对具有改进性能的新材料的搜索已引起越来越多的关注。使用不具有迭代反馈回路的非自适应ML方法，在以前未开发的搜索空间进行外推时，效果会很差，尤其是在对小数据集进行训练时。我们对表现出明显的成分－性质关系的两个数据集进行了数值模拟，并探索了自适应ML策略在识别具有最高成分的最佳材料成分方面的相对有效性。自适应ML方法显示了外推的希望，并找到了性能比训练数据更好的构图，但是发现的速度取决于构图－属性图的细微差别。