Journal of Chemical Information and Modeling, Год журнала: 2025, Номер unknown
Опубликована: Янв. 9, 2025
In this contribution, we examine the interplay between target definition, data distribution, featurization approaches, and model architectures on graph-based deep learning models for thermodynamic property prediction. Through consideration of five curated sets, exhibiting diversity in elemental composition, multiplicity, charge state, size, impact each these factors accuracy. We observe that i.e., using formation instead atomization energy/enthalpy, is a decisive factor, so careful selection approach. Our attempts at directly modifying result more modest, though not negligible, accuracy gains. Remarkably, molecule-level predictions tend to outperform atom-level increment predictions, contrast previous findings. Overall, work paves way toward development robust with universal capabilities, can reach excellent across sets compound domains.
Язык: Английский