Enhancing data-driven modeling of fluoride concentration using new data mining algorithms

Groundwater is an essential constituent of drinking water in hard rock areas and hence it requires the analysis of contaminant resources. Fluoride contamination with large spatial variation in the part of Sindhudurg district is reported. The present study focuses on the development of data-driven modeling of fluoride concentration using on-site measurement of physicochemical parameters. In this configuration, six machine learning(ML) architectures, namely data mining algorithms were explored including novel algorithms Gaussian process (GP) and long short term memory (LSTM). The results were compared with support vector machine (SVM), random forest (RF), extreme learning machine (ELM), and multi-layer perceptron (MLP) as a benchmark to test the robustness of the modeling process. In total 225 water samples from different dug-wells/bore- wells were obtained from the area (latitude:15.37–16.40 degree, longitude:73.19–74.18 degree) in the period of 2009–2016. Two subsets of data were divided with 80% data in training and 20% in testing. Different 9 physicochemical parameters pH, EC, TDS, Ca²⁺, Mg²⁺, Na⁺, Cl⁻, HCO₃⁻, SO₄²⁻ were used in the modeling of fluoride (F⁻). In this context logarithmic transformation of raw data was employed to improve the correlation between input and target and therefore to enhance the modeling accuracy. Different quantitative and qualitative (visual) measures were taken to establish the prediction power of models. Results revealed that GP outperform all other models in fluoride prediction followed by LSTM, SVM, MLP, RF, and ELM, respectively. Results also revealed that the model’s performance depends on model structure and data accuracy.