Outliers in Shannon's effective ionic radii table and the table extension by machine learning

In materials informatics and computational design, ionic radius is an essential physical feature needed to predict and model structures and other material properties. Currently, the most used reference for ionic radii is the renowned Shannon's table set in 1976, where each ionic configuration is composed of the element, oxidation state (OS), and coordination number (CN). The original table has 488 radii for 476 unique ionic configurations. The table was extended recently using machine learning (ML) to add 512 missed known configurations, which are not in the original table. Conceptually, the accuracy of ML prediction depends on the reliability of the original data. In this work, multiple techniques are used to detect outliers in Shannon's table to improve the prediction accuracy. It turned out that there are 24 outliers. The ionic radius testing root mean square error using Gaussian process regression is reduced from 3.99 pm using all Shannon's data (488 data points) to 2.71 pm (with 464 data points), with R² scores of 0.989 and 0.995, respectively. This is done by removing only 4.92 % of the data. The cleaned data are then used to estimate the ionic radii of the additional 512 OS/CN ionic configurations and the detected outliers. The consolidated table of 988 unique ionic configurations is provided in the Supplementary Information and is made available online.