Separation of coeluted compounds in capillary gas chromatography: customization via machine learning retention time...

Gas Chromatography is a versatile separation technique widely used in analytical chemistry for constituent determination. However, gas chromatography compound identification is not directly feasible unless the method is coupled with complementary techniques such as mass spectrometry or with referencing methods like retention indices. Statistical retention time prediction of compounds based on the gas chromatography experimental and instrumental parameters could facilitate the gas chromatography characterization process based on compounds' retention time values. Furthermore, these predictions can be used for gas chromatography optimization to achieve optimum separation of coeluted compounds. Six different algorithms were trained on 70% of the manually compiled 608 data points that were gathered from RESTEK company chromatograms and published literature, covering 73 different gas chromatography settings and 61 compounds ranging from C $_1^{}$ to C $_{12}^{}$ hydrocarbons. The chosen model, namely the adaptive boosting support vector regression, outperformed tree-based ensemble methods and showed high accuracy and consistency across the validation and testing sets with $R^2$ scores of 0.992-0.993. We propose a machine learning pipeline that enables accurate retention time predictions, which will facilitate the characterization process and allow for gas chromatography optimization for isomer separation. The predicted settings are in strong agreement with a proposed gas chromatograph designed specifically for isomer separation.