Chemical data have been created from many different origins. The chemicals themselves tend to be synthesized out of curiosity or as an industry-led need. Their materials characterization and development for functional applications generate cognate data about their structures and properties. Chemical structures and properties may also be computed ahead of their physical creation. The collation of all this chemical information affords a ‘chemical space’ that encapsulates a rich and diverse set of data. This opinion article considers the shape and size of this chemical space and of its various subdomains, how the relative availability of its structure and property information governs what type of questions one should ask of the data, and what type of machine learning (ML) should be applied to discover a new material. Application examples of ML methods that produce predictive models for data-driven materials discovery are discussed.

化学数据来自许多不同的来源。出于好奇或出于行业主导的需求，这些化学品本身往往会合成。它们的材料表征和针对功能应用的开发会生成有关其结构和特性的同类数据。化学结构和性质也可以在物理创建之前进行计算。所有这些化学信息的整理提供了一个“化学空间”，其中囊括了丰富多样的数据集。这篇观点文章考虑了该化学空间及其各个子域的形状和大小，其结构和属性信息的相对可用性如何控制人们应该对数据提出哪种类型的问题以及应该采取哪种类型的机器学习（ML）用于发现新材料。