Intelligent soft sensing and regulation of daily uranium production in in situ leaching using domain-guided recursive feature construction and ant colony optimization

Geological heterogeneity coupled with hydrodynamic conditions strongly constrains uranium recovery efficiency and operational control in in situ leaching (ISL). Mechanism-based simulators can be physically consistent but often require costly geological parameters and intensive computation, whereas purely data-driven models typically provide limited interpretability for regulation. This study proposes an engineering-oriented framework (DFS-ACO-LightGBM) for soft sensing and regulation of daily uranium production. A DFS-based recursive feature construction module (DFS-FC) explicitly generates high-order geology–process interaction terms (e.g., permeability × liquid extraction rate), transforming implicit physical constraints into learnable features. Ant Colony Optimization (ACO) is introduced to robustly tune LightGBM hyperparameters in a high-dimensional, non-convex search space. Using over 15,000 daily records from multiple sandstone-type ISL areas (Sep 2023–Oct 2024), DFS-ACO-LightGBM achieves the best performance among 12 schemes ( R 2 = 0.9688, MAE = 0.1385 kg, training time ≈ 0.58 s). Response probing further reveals a phased production response to extraction intensity—gradual increase, threshold transition, and post-threshold differentiation—and indicates a startup threshold around 150–165 m 3 /d for low-to-moderate endowment units. Accordingly, a differentiated fluid allocation strategy (“threshold crossing + graded efficiency”) is proposed to support practical production optimization under heterogeneous ore conditions.