An integrated machine learning framework for predicting anthropogenic and natural iodine isotopes in the South China Sea with uncertainty quantification.

Anthropogenic Iodine-129 (129I) is a critical long-lived radionuclide for tracing ocean circulation and environmental contamination. However, its measurement is costly and yields spatially sparse data, limiting comprehensive environmental assessment. This study proposes a novel, integrated machine learning framework to predict the concentrations of not only anthropogenic 129I but also stable 127I and their isotopic ratio (129I/127I) in the South China Sea using readily available oceanographic parameters. A Bayesian Neural Network (BNN) was developed as the core predictive model to provide robust uncertainty quantification for each estimate. The BNN's complex hyperparameters were systematically optimized using the Improved Snow Goose Algorithm (ISGA), a powerful metaheuristic method designed to efficiently navigate complex search spaces. The optimized ISGA-BNN framework demonstrated high predictive accuracy for all three targets when evaluated on an independent test set. The models achieved R2 values of 0.9623 for 127I, 0.9286 for the 129I/127I ratio, and 0.8148 for 129I. Diagnostic analysis confirmed that the BNN provided well-calibrated uncertainty intervals, accurately capturing the confidence level of each prediction. This framework represents an advancement by providing accurate, multi-target predictions with vital uncertainty estimates, holding potential for enhancing environmental monitoring and optimizing future sampling campaigns.