Graph-Based Deep\nLearning Models for Thermodynamic\nProperty Prediction: The Interplay between Target Definition, Data\nDistribution, Featurization, and Model Architecture

In\nthis contribution, we examine the interplay between target definition,\ndata distribution, featurization approaches, and model architectures\non graph-based deep learning models for thermodynamic property prediction.\nThrough consideration of five curated data sets, exhibiting diversity\nin elemental composition, multiplicity, charge state, and size, we\nexamine the impact of each of these factors on model accuracy. We\nobserve that target definition, i.e., using formation instead of atomization\nenergy/enthalpy, is a decisive factor, and so is a careful selection\nof the featurization approach. Our attempts at directly modifying\nmodel architectures result in more modest, though not negligible,\naccuracy gains. Remarkably, we observe that molecule-level predictions\ntend to outperform atom-level increment predictions, in contrast to\nprevious findings. Overall, this work paves the way toward the development\nof robust graph-based thermodynamic model architectures with more\nuniversal capabilities, i.e., architectures that can reach excellent\naccuracy across data sets and compound domains.