An automatic model selection-based machine learning approach to predict seawater intrusion into coastal aquifers

Concerns about seawater intrusion resulting from unplanned mining of groundwater from coastal aquifers have become a global issue. To address this, it is crucial to implement a well-structured plan for groundwater withdrawal that can effectively manage and restrict salinity levels within the aquifer to acceptable levels. An integrated simulation–optimization (S–O) system has been a suitable tool for developing an optimum groundwater withdrawal scheme for managing seawater intrusion. The success of an S–O methodology largely depends on the use of computationally competent surrogates for the intricate simulation model. Although several surrogate models have been recently proposed to create management models for seawater intrusion using integrated S–O approach, the majority of these surrogates have been developed based on the subjective judgement. To fill this research gap, this study proposes an automatic model selection (AutoML)-based machine learning (ML) approach to predict the mechanisms of seawater intrusion in coastal aquifers. The AutoML was performed by optimizing the hyperparameters of a number of ML algorithms and selecting the best performing algorithm utilizing the asynchronous successive halving algorithm (ASHA). The best performing models were developed at 16 monitoring locations (MLs) using the predictor–response training data originated from a simulation model. Results revealed the capability of the ASHA optimization-based AutoML scheme to optimally select the best performing models with adequate prediction accuracies for the particular MLs. The selected best models at various monitoring locations exhibited high performance with accuracy (= 1), R (~ 0.99), NS (~ 0.99), IOA (~ 0.99), and KGE (~ 0.99), which are close to 1, indicating excellent model accuracy. Furthermore, the models demonstrated an RMSE value range of 0.0003–1.4987 mg/L, a relatively small range that is considered favorable for any predictive modeling approach. This study reveals the suitability and efficacy of the automatically selected surrogate models to develop an S–O-based management model to address real-world coastal aquifer management challenges related to seawater intrusion.