Decomposing Chemical Space: Applications to the Machine Learning of Atomic Energies

Abstract

We apply a number of atomic decomposition schemes across the standard QM7 data seta small model set of organic molecules at equilibrium geometryto inspect the possible emergence of trends among contributions to atomization energies from distinct elements embedded within molecules. Specifically, a recent decomposition scheme of ours based on spatially localized molecular orbitals is compared to alternatives that instead partition molecular energies on account of which nuclei individual atomic orbitals are centered on. We find these partitioning schemes to expose the composition of chemical compound space in very dissimilar ways in terms of the grouping, binning, and heterogeneity of discrete atomic contributions, e.g., those associated with hydrogens bonded to different heavy atoms. Furthermore, unphysical dependencies on the one-electron basis set are found for some, but not all of these schemes. The relevance and importance of these compositional factors for training tailored neural network models based on atomic energies are next assessed. We identify both limitations and possible advantages with respect to contemporary machine learning models and discuss the design of potential counterparts based on atoms and the intrinsic energies of these as the principal decomposition units.

Properties of Local Electronic Structures

“…This uniqueness holds true in the results of the MO-based scheme but not in those based on AOs, as was previously alluded to in our recent work on decompositions across small model compound spaces in organic chemistry. 33 Having discussed how AO-and MO-based partitioning schemes yield somewhat contrasting fingerprints of local electronic structures within typical functional groups of organic molecules, we next turn our attention to the different manners in which these schemes decompose changes to atomic properties during the simulation of chemical reactions, e.g., by tracking these along a well-defined reaction coordinate. 44 As an initial application, we study the simplest of S N 2 substitution reactions, in which a halide (nucleophile) replaces a halogen of the same kind on a methyl moiety in a concerted attack from the rear side, i.e., X − •••CH 3 X (X = F, Cl, or Br).…”

“…Combined with the fact that corresponding calculations of atomic dipole moments, partial charges, and energies in a basis set lacking diffuse functions yield fundamentally different results (see Figures S2−S4), the usefulness of native AO-based schemes for interpretative purposes is once more brought into question. 33 For partitioning schemes based on spatially localized MOs, on the other hand, results in both of the (aug-)pc-2 basis sets are indistinguishable, allowing for unbiased interpretations of local electronic structures. Our initial results hence motivate the application of these schemes to even more complex chemistry, whereby finer mechanistic details may be exposed through changes to local, rather than total, molecular properties.…”

“…In a recent work of ours, we applied a number of AO-and MO-based partitioning schemes across a small model of chemical compound space, consisting of less than 10k constitutional and structural molecular isomers, each limited to a mere 7 heavy atoms (C, N, O, and S). 33 Its small extent notwithstanding, this model space still comprised a number of fundamentally distinct local atomic environments and our work was concerned, in part, with the study of how atomic contributions to cohesive energies were distributed across the space, but also how the aggregate and diversity in the representation of each element from these decompositions influenced the rate by which atomic contributions could be learned using various kinds of machine learning architectures based on neural networks. In the course of the present work, we will instead seek to quantify what factors fundamentally determine the magnitude and sign of atomic contributions to molecular atomization energies.…”

Interpretable Attribution Assignment for Octanol–Water Partition Coefficient

“…Attribution assignment is also gaining significance in chemistry. − Most attribution assignments focus on which atoms and bonds are important in predicting chemical properties. Although the attribution assignments of atoms are vital when designing molecules and debugging models, less attention has been paid to whether attribution assignments are made based on the right reasons.…”

Interpretable attribution assignment for octanol-water partition coefficient

“…Attribution assignment is also gaining significance in chemistry. [5][6][7][8][9] Most attribution assignments focus on which atoms and bonds are important to predict chemical properties.…”