In neutral in-situ leaching (CO 2 + O 2 ) of sandstone-hosted uranium deposits, oxygen injection, lixiviant injection, and hydrochemical conditions are mutually coupled, leading to nonlinear responses of daily uranium output to operating parameters. Meanwhile, field monitoring data are also noisy and incomplete, limiting empirical optimization. This study focuses on seven-spot leaching units in a mining area, constructs a daily-scale dataset including injection/production parameters, hydrochemical indicators of the produced solution, residual uranium inventory, and time-series derived features, and develops a multilayer perceptron prediction model (VAE-SEMLP) that integrates a variational autoencoder (VAE) with a Squeeze-and-Excitation (SE) attention mechanism. Across three representative unit tests, the proposed model achieves an average R 2 of 0.95 and a normalized RMSE of 0.06, outperforming support vector regression (SVR), random forest (RF), and XGBoost overall. Ablation experiments further confirm the synergistic gain from combining the VAE and SE modules. Parameter scanning under typical static operating-condition slices reveals a pronounced unimodal response of daily uranium output to lixiviant injection volume, indicating an optimal injection interval that shifts with oxygen-injection level. This work provides data-driven support for daily uranium-output prediction and quantitative optimization of coupled oxygen-lixiviant injection schemes in neutral in-situ uranium leaching.
Liu et al. (Wed,) studied this question.