Transferability of atom-based neural networks

Machine-learning models in chemistry—when based on descriptors of atoms embedded within molecules—face essential challenges in transferring the quality of predictions of local electronic structures and their associated properties across chemical compound space. In the present work, we make use of adversarial validation to elucidate certain intrinsic complications related to machine inferences of unseen chemistry. On this basis, we employ invariant and equivariant neural networks—both trained either exclusively on total molecular energies or a combination of these and data from atomic partitioning schemes—to evaluate how such models scale performance-wise between datasets of fundamentally different functionality and composition. We find the inference of local electronic properties to improve significantly when training models on augmented data that appropriately expose local functional features. However, molecular datasets for training purposes must themselves be sufficiently comprehensive and rich in composition to warrant any generalizations to larger systems, and even then, transferability can still only genuinely manifest if the body of atomic energies available for training purposes exposes the uniqueness of different functional moieties within molecules. We demonstrate this point by comparing machine models trained on atomic partitioning schemes based on the spatial locality of either native atomic or molecular orbitals.