Replacing the internal standard to estimate micropollutants using deep and machine learning.

Area of Science:

• Environmental Chemistry
• Analytical Chemistry
• Data Science

Background:

• Urbanization increases micropollutant contamination in aquatic ecosystems, posing risks to human health and the environment.
• Accurate monitoring of micropollutants is crucial, but traditional methods using stable isotope labeled (SIL) standards are cost-prohibitive.

Purpose of the Study:

• To develop cost-effective data-driven models for estimating micropollutant concentrations.
• To investigate the potential of natural organic matter (NOM) mass spectrometry data as a surrogate for SIL standards.

Main Methods:

• Developed and trained deep learning (DL) and machine learning (ML) models, including ResNet101, GoogLeNet, VGG16, Inception v3, random forest (RF), support vector machine (SVM), and artificial neural network (ANN).
• Utilized 35 alternative mass spectrum (MS) subsets derived from NOM data as input for model training.
• Trained models on 680 MS data points to estimate concentrations of five micropollutants: Sulpiride, Metformin, and Benzotriazole.

Main Results:

• ResNet101 achieved the highest performance among DL models, with average validation R² of 0.84 and MSE of 0.26 ng/L.
• Random Forest (RF) demonstrated the best performance among ML models, yielding R² of 0.69 and MSE of 0.58 ng/L.
• Trained models showed accurate prediction capabilities for micropollutant concentrations.

Conclusions:

• Data-driven models utilizing NOM-derived MS data can effectively estimate micropollutant concentrations, replacing expensive SIL standards.
• This approach offers a rapid, economical, and accurate alternative for environmental micropollutant monitoring.
• The study highlights the potential of machine learning and deep learning in addressing challenges in environmental analysis.