Predicting bond-currents in polybenzenoid hydrocarbons with an additivity scheme

Abstract

We report on the construction and application of a new bond-current additivity scheme for polybenzenoid hydrocarbons. The method is based on identification of the smaller substructures contained in the system, up to tricyclic subunits. Thus, it enables the prediction of any cata-condensed unbranched polybenzenoid hydrocarbon, using a library consisting of only four building blocks. The predicted bond-currents can then be used to generate NICS values, the results of which validate previous observations of additivity with NICS-XY-Scans. The limitations of the method are probed, leading to clearly delineated and apparently constant error boundaries, which are independent of the molecular size. It is shown that there is a relationship between the accuracy of the predictions and the molecular structure and specific motifs that are especially challenging are identified. The results of the additivity method, combined with the transparent description of its strengths and weaknesses, ensure that this method can be used with well-defined reliability for characterization of polybenzenoid hydrocarbons. The resource-efficient and rapid nature of the method make it a promising tool for screening and molecular design.

Revealing Structure-Property Relationships in Polybenzenoid Hydrocarbons with Interpretable Machine-Learning

“…Interestingly, the relative energy also shows a dependence on a specific substructure, "LAL". This particular subsequence was previously noted as behaving in an anomalous manner 24 in the prediction of magnetic behavior in PBHs. A similar analysis using the CUSTODI model trained on SMILES strings yielded no meaningful results, as the substrings used in CUSTODI models are short (this results from the hyperparameter optimization; see Section S4.2 in the Supporting Information for more details).…”

“…Our group has demonstrated in a series of reports over the past few years that catacondensed PASs can be broken down into their smaller components (monocyclic, bicyclic, and tricyclic), and the magnetic properties of the larger molecules can be predicted by summing the contributions of these smaller subunits using an additivity scheme. [24][25][26] For the particular case of the PBHs, molecular properties can be predicted by the type and order of the tricyclic components themselves, where the two tricyclic subunits differ only in their annulation: linear or angular, i.e., anthracene or phenanthrene, respectively. This conclusion allows for a reduction of the molecular structure to the sequence of tricyclic subunits (i.e., the annulation sequence).…”

“…This set of chemically intuitive features (detailed in Table 1) was inspired by our collective experience studying PBHs and by structure-property relationships previously found in smaller datasets. 24,27 The purpose of using the LFV as input was threefold: (1) to validate the intuition we developed previously, (2) to check the predictive power of these descriptors, and (3) to compare the conclusions derived from this set of PBH-specific features to those derived from more general chemical representation. The CUSTODI Framework CUSTODI is a recently developed tokenization technique for text-based molecular representations.…”

Simple and efficient visualization of aromaticity: bond currents calculated from NICS values

“…Moreover, we recently observed a very good agreement between NICS ZZ values calculated with density functional theory (DFT) and NICS ZZ values calculated from bond-currents using the Biot–Savart equation. 82 Based on the surprisingly accurate predictions obtained with this approach, we hypothesized that a scheme converting NICS to unique bond-current graphs could be established.…”

Visualizing electron delocalization in contorted polycyclic aromatic hydrocarbons

“…16,29 NICS has routinely been computed as single points 1 Å above the middle of each ring. 36,37 Richer pictures of NICS in molecules can be obtained by mapping NICS using 1D, [38][39][40][41] 2D, 37,[42][43][44][45][46][47][48] and 3D grids [49][50][51][52][53][54][55][56][57] of Bq. The latter approach can produce 3D maps of the magnetic properties of molecules as isovalue surfaces.…”