Categorical modeling of the flow pattern of liquid organic compounds between blade electrodes using semiempirical and ab initio quantum chemical descriptors

For a data set of 30 organic fluids, categorical modeling has been employed to predict the flow pattern under an external electric field. To this end, a previously generated data set was augmented by 10 compounds with new experimental results, and quantum chemical methods have been used to characterize the geometric and electronic structure of the molecules on both the semiempirical and ab initio levels of theory. Both linear discriminant analysis (LDA) and binary logistic regression (BLR) have been employed to model the flow rate (high vs. low) and flow direction (left vs. right). For the flow rate, good LDA and BLR calibration statistics using the dipole moment, hydrophobicity and some charged partial surface area (CPSA) descriptors is accompanied with moderate prediction statistics, as evaluated through simulated external validation, and activity scrambling shows that chance correlation is not relevant. Additional neural network analyses yielded no stable models due to constraints imposed by the data set size. For the flow direction, LDA and BLR calibration and prediction statistics show more variation among the different models generated, with an overall performance inferior to the one for the flow rate. Here, besides CPSA descriptors, two parameters characterizing the softness of the electronic structure are involved. In general, BLR is slightly superior to LDA for both properties. The results are discussed in terms of contingency table statistics and with respect to the mechanistic meaning of molecular descriptors.