Machine learning accelerates virtual screening in which material candidates are selected from existing databases, facilitating materials discovery in a broad chemical search space. Machine learning models quickly predict a target property from explanatory material features called descriptors. However, a major bottleneck of the machine learning model is an insufficient amount of training data in materials science, especially data with non-equilibrium properties. Here, we develop an alternative virtual-screening process via ensemble-based machine learning with one handcrafted and two generic descriptors to maximize the inference ability even using a small training dataset. A joint representation with the three descriptors translates the physical and chemical properties of a material as well as its underlying short- and long-range atomic structures to describe a multifaceted perspective of the material. As an application, the ensemble-scope descriptor learning model was trained with only 29 entries in the training dataset, and it selected potential oxygen-ion conductors from 13,384 oxides in the inorganic crystal structure database. The experiments confirmed that we successfully discovered five compounds that have not been reported, to the best of our knowledge, as oxygen-ion conductors.

机器学习可加速虚拟筛选，从现有数据库中选择候选材料，从而在广阔的化学搜索空间中促进材料的发现。机器学习模型从称为描述符的解释性材料特征中快速预测目标属性。但是，机器学习模型的主要瓶颈是材料科学中训练数据的数量不足，尤其是具有非平衡特性的数据。在这里，我们通过基于整体的机器学习开发了一种替代的虚拟筛选过程，该过程使用一个手工制作的和两个通用的描述符，即使使用很小的训练数据集，也能最大程度地提高推理能力。具有三个描述符的联合表示可翻译材料的物理和化学特性及其底层的短程和远程原子结构，以描述该材料的多面性。作为一种应用，对集合谱描述符学习模型进行了训练，仅在训练数据集中进行了29个条目的训练，并从无机晶体结构数据库中的13384种氧化物中选择了潜在的氧离子导体。实验证实，就我们所知，我们成功地发现了五个尚未报道的化合物，它们是氧离子导体。无机晶体结构数据库中有384种氧化物。实验证实，就我们所知，我们成功地发现了五个尚未报道的化合物，它们是氧离子导体。无机晶体结构数据库中有384种氧化物。实验证实，就我们所知，我们成功地发现了五个尚未报道的化合物，它们是氧离子导体。