Abstract
The gas chromatography retention indices of 100 different components of essential oils, on three columns with stationary phases of different polarities, were used to develop robust quantitative structure–retention relationship (QSRR) models. Two linear models with only one variable, i.e. solvation entropy, were developed, which explain 95 and 94% of variances of the test set for dimethyl silicone and dimethyl silicone with 5% phenyl group columns, respectively. These models are extremely simple and easy to interpret, but they show higher errors compared with more robust models such as partial least square (PLS) and ridge regressions. For the third column (polyethylene glycol (PEG)), 24 hydrogen bonding descriptors were calculated and were used for modeling. Kernel orthogonal projection to latent structure (KOPLS), as a non‐linear technique, was applied for the modeling of the retention indices of the compounds on the PEG column. R^2^ values for the test set established by Monte Carlo cross‐validation and SPXY (sample set partitioning based on joint x–y distances) test set of the KOPLS were 0.92 and 0.94, respectively. y‐Randomization indicated that chance plays no role in constructing the KOPLS model.