“…SOAP representations, and in general local-density representations, when coupled with kernel and linear methods alike, are proven to be capable of accurate and fast predictions of atomistic systems properties in bulk and of molecular materials with diverse chemistries, also providing fairly accurate predictions when trained on small ($10 3 configurations) datasets. [51,54,55]. Atom-density representation further encode by design the physical symmetries (rotation, permutation, translation invariance) which rule interatomic interaction, and, though not-transparent, correlate with more interpretable "classical" descriptors, for example, coordination [56][57][58].…”