DELINEATING GROUNDWATER FLOW PATHS WITH ISOTOPIC SIGNATURES AND MACHINE LEARNING TECHNIQUE: A CASE STUDY IN THE LOWER KELANTAN RIVER BASIN, MALAYSIA

Machine learning techniques may provide support in addressing the complexities of biogeochemical processes by characterizing relationships between variables monitored in natural and anthropized environments. The support is demonstrated by the presented study in the lower Kelantan river basin (LKRB) and the interaction between the strategic coastal aquifer and the surface waters. The study is based on a multi-tracer evaluation of 85 samples collected in three different months, including significant ions (Ca²⁺, K⁺, Mg²⁺, Na⁺, Cl⁻, SO₄²⁻, HCO₃⁻, CO₃²⁻), nutrients (NO₃⁻+NO₂⁻), trace elements (F⁻, Br⁻), and isotopes signatures (δ¹⁸O, δ²H, δ¹³C, δ¹⁵N and, ³H). The dataset was treated through fuzzy C-Means (FCM) clustering, resulting in five distinct clusters. The clustering analysis revealed a layered structure within the aquifers. Samples from January in the shallow aquifer and surface water were mainly assigned to FC5. FC3 included a mix from intermediate brackish and deep aquifers, while FC4 represented mixed water. FC1 contained mostly deep aquifer samples, and FC2 had primarily shallow aquifer samples. The FCM clustering indicated recently recharged, actively mixing groundwater exhibiting ³H levels near 6 TU alongside depleted δ¹⁸O and δ²H ratios, with separate clusters showing lighter δ¹⁵N and δ¹³C values compared to those with high ³H (around 10 TU) and enriched δ¹⁸O values. These latter clusters were characterized by relatively heavier δ¹⁵N and δ¹³C composition and elevated TDS. Additionally, the clustering results indicated higher NO₃⁻+NO₂⁻, suggesting potential localized surface contaminant impacts in the LKRB.

This study highlights FCM's capacity for flexibility by considering variations within clusters, aiding in identifying intermediate patterns before assigning the highest salinity brine to separate cluster. Integrating geochemical data and FCM analysis successfully distinguished the main water types and end members. The clustering analysis revealed a layered structure within the aquifers, contrasting geochemical patterns through this unsupervised learning approach. This information is expected to be used to trace seawater intrusion and contaminant movement further, facilitating improved groundwater management of the coastal aquifers in LKRB.