An Integrated Machine Learning Approach for Real‐Time Prediction, Diagnostics and Optimization of Uranium‐Leaching Groundwater System

Abstract Reliable and efficient simulation and optimization (SO) approach are crucial for groundwater management. Traditionally, SO of groundwater system relies on process‐based numerical models, which often feature low computational efficiency, and unsatisfactory accuracy under limited amount of supportive data and time‐varying aquifer conditions. This study establishes a bidirectional SO approach for adaptive prediction, diagnostics and optimization of groundwater system. Prediction module uses recurrent and convolutional neural network to tackle the spatiotemporal relationship between well operations and responses; particularly the lightweight model is developed under limited historical observations, and transfer learning is leveraged for model updating based on new observations to accommodate evolving aquifer conditions. Diagnostics module uses expected gradient algorithm to detect abnormal situation that the predicted well responses are biased from the object, and to identify controlling factors (e.g., well positions and pumping rates) sensitive to the abnormal responses. Optimization module uses iterative ensemble smoother to optimize the controlling factors. The effects of this real‐time analysis approach are exemplified in a uranium leaching system in north China. The forward lightweight model facilitated with transfer learning achieves fast and accurate prediction of uranium concentrations under varying pumping rates. The diagnostics module allows for the dynamic detection of well positions and pumping rates controlling the uranium production, and are explainable in comparison to residual uraninite distribution in the aquifer simulated by reactive transport model. Finally, the optimization of pumping rates at the controlling wells in real time enhances the uranium production by approximately 20% higher than that without SO.