“…In this work, the dataset consists of 112 experimental data, which were summarized by Akrami et al.,
2 as can be seen in Table S1. According to the obtaining pathways, the input features can be classified into three categories, including the atomic/precursory‐based descriptors, the experience‐independent features, and the characteristics acquired by high‐throughput simulation
41 . Here, in order to avoid time‐consuming and expensive high‐throughput calculations, the first two types of features were selected to form the feature space where the first type descriptors consist of anion‐to‐cation radius ratio (
), difference in Pauling electronegativity (
),
42 difference in Mulliken electronegativity (
),
43 entropy of mixing (
)
44 as well as atomic size mismatch (
),
42 and the corresponding calculation formulas are as follows
31,45 :
…”