Environmental water analysis remains challenging due to the complexity of the matrix and low concentrations of emerging contaminants. Liquid chromatography coupled to high-resolution mass spectrometry via an electrospray ionization source (LC/ESI/HRMS) is a powerful technique that offers high sensitivity and selectivity for detecting organic compounds. Therefore, non-targeted screening (NTS) with LC/ESI/HRMS has been widely used to detect and identify emerging contaminants in environmental water samples. Nevertheless, many aspects of detection and identification of unknown LC/ ESI/HRMS features remain challenging due to the large amount of data and lack of many analytical standards.

The main aim of this thesis is to utilize machine learning to improve the detection and identification of emerging contaminants throughout the NTS workflow, starting from sample preparation to reporting the results. In Paper I, different solid-phase extraction (SPE) cartridges were compared to aid the selection of a suitable SPE cartridge for NTS of environmental water samples. Furthermore, machine learning was used to model the SPE recoveries and the developed model was assessed and validated using an external dataset. Paper II was a study of the impact of mobile phase (pH, organic modifier, additive) and stationary phase on the LC/ESI/HRMS sensitivity towards 78 selected emerging contaminants. The results guided the selection of chromatographic conditions for LC/ESI/HRMS analysis of wastewater samples, where three selected mobile phases were further tested. Paper III describes the impact of mobile phase pH, organic modifier, additive and stationary phase on liquid chromatography retention of 78 selected emerging contaminants. Here I developed a MultiConditionRT model to predict liquid chromatography retention times in four retention mechanisms in combination with two organic modifiers, different pH-s (2.1 to 10.0), and seven additives. MultiConditionRT was validated using internal and external datasets containing 408 new compounds. In Paper IV, a new approach was developed to estimate the limit of detection (LoD) based on predicted ionization efficiency. This approach can be utilized for prioritization of unknown or tentatively identified features and to assess the detectability of chemicals with NTS methods.

Overall, this thesis illustrates how machine learning can be used to improve the detection and identification of emerging contaminants from NTS of environmental waters with LC/ESI/HRMS. In particular, the findings and data presented in this thesis offer valuable insights into the importance of accounting for the analysis conditions while improving the NTS toolbox.