Machine Learning-Based Quantitative Structure–Property Relationships for the Electronic Properties of Cyano Polycyclic Aromatic Hydrocarbons

Abstract

In this study, quantitative structure−property relationships (QSPR) based on a machine learning (ML) methodology and the truncated degree of π-orbital overlap (DPO) to predict the electronic properties, namely, the bandgaps, electron affinities, and ionization potentials of the cyano polycyclic aromatic hydrocarbon (CN-PAH) chemical class were developed. The level of theory B3LYP/6-31+G(d) of density functional theory (DFT) was used to calculate a total of 926 data points for the development of the QSPR model. To include the substituents effects, a new descriptor was added to the DPO model. Consequently, the new ML-DPO model yields excellent linear correlations to predict the desired electronic properties with high accuracy to within 0.2 eV for all multi-CN-substituted PAHs and 0.1 eV for the mono-CN-substituted PAH subclass.

Atom-Based Machine Learning Model for Quantitative Property–Structure Relationship of Electronic Properties of Fusenes and Substituted Fusenes

“…In this study, we use a data set of a total of 2131 molecules from four classes of polycyclic aromatic compounds: PAHs, thienoacenes, singly to quadruply cyano-substituted PAHs, and singly to quadruply nitro-substituted PAHs. Electronic properties data on PAHs, thienoacenes, and cyano-substituted PAH molecules are from our previous studies. − Electronic properties of the nitro-substituted PAH chemical class were calculated in this study at the same level of theory as other chemical classes, namely the B3LYP-D3/6-31+G­(d) level of theory, by using the GAUSSIAN package . Electron affinity (EA) and ionization potential (IP) can be respectively approximated to be the highest occupied molecular orbital (HOMO) and the lowest occupied molecular orbital (LUMO) energy levels according to Koopman’s theorem.…”

“…Electron affinity (EA) and ionization potential (IP) can be respectively approximated to be the highest occupied molecular orbital (HOMO) and the lowest occupied molecular orbital (LUMO) energy levels according to Koopman’s theorem. The band gap values are approximated to be the HOMO–LUMO gaps. − …”

“…For each experiment, we randomly split the data set into two separate subsets, namely, the training and test sets, with approximately 70% and 30% of the data, respectively. To ensure that the training and test sets have similar distributions of band gap values, we employed a band gap–bins-wise stratified splitting method, as previously described in our work. , …”

“…To ensure that the training and test sets have similar distributions of band gap values, we employed a band gap−bins-wise stratified splitting method, as previously described in our work. 18,19 Figure 3 shows the root-mean-square deviations (RMSD) of band gaps for the ECFP-based models (Figure 3A−C) and WLbased model (Figure 3D−I) as functions of the number of WL iterations. Because of overfitting, some GPR/WL kernel methods with high numbers of iterations produce RMSD values larger than 2.00 eV, and therefore, they are not presented in Figures 3D−F for the sake of visibility of other models' performances.…”

“…Electronic properties, such as band gap, frontier orbitals’ energies, electron affinity (EA), and ionization potential (IP), of organic semiconductors correlate with the materials’ stability, charge transport properties, and other devices’ qualities ,, and, thus, make pragmatic and vital objective functions for data-driven optimization of materials. Fortunately, these properties correlate well with the molecule’s 2-dimensional (2D) structure, especially for polycyclic aromatic hydrocarbons (PAHs) by the use of the degree of π-orbital overlap (DPO) descriptor, which is based on the quantum mechanic 2D particle-in-the-box model. − Nevertheless, the attachment of electron acceptors/donors to PAH molecules complicates such a particle-in-the-box physical model and, hence, undermines the simple “box” model. This calls for a better approach to this problem.…”