The dynamics of silicate liquids and supercooled liquids plays a key role in glass manufacturing and geology. However, the relationship between the dynamics of silicate glasses and their atomic structure remains unclear, as intuitive structural metrics are only weakly correlated to dynamics. Because of its ability to discover complex patterns within data, machine learning is considered as a powerful tool to map structure to dynamics by revealing complex structure characteristics. Based on molecular dynamic simulations of silica supercooled liquid, a classification machine learning model was developed to investigate the structure origin. By interpreting the results of the model, we extracted a non-intuitive structural metric (called “softness”) that exhibits a strong correlation with the displacement of the oxygen atoms. This non-intuitive structural fingerprint, based merely on the initial atomic positions, serves as an appropriate structural descriptor to predict silica’s dynamics. More generally, this approach offers a promising route to decipher the relationship between atomic structure and dynamics.