A hybrid deep learning architecture that integrates convolutional neural networks, gated recurrent units, and an attention mechanism accurately forecasts extraction uranium concentration while establishing optimized acid injection thresholds in in-situ leaching operations, according to research conducted on four years of continuous field data. The framework, developed to address the complex nonlinear dynamics between injected leaching reagents and recovered uranium solutions, achieved a test-set mean absolute error of 0.0543, a mean squared error of 0.0315, and a coefficient of determination of 0.915. In single-variable simulation analyses based on the trained model, uranium concentration exhibited a characteristic rise-then-stabilize pattern across varying acid concentrations, enabling the derivation of tailored injection set-points across four distinct residual reserve levels: 11.69 grams per liter for 30 tons, 10.85 grams per liter for 25 tons, 11.41 grams per liter for 20 tons, and 10.99 grams per liter for 15 tons of residual uranium.
In-situ leaching has become the predominant approach for nuclear energy resource extraction because it accesses underground ore bodies with minimal surface disturbance, comparatively low operating expenditures, and reduced ecological disruption relative to conventional mining methods. During this hydrometallurgical extraction cycle, an acidic leaching solution, primarily consisting of sulfuric acid, is pumped through injection wells directly into permeable underground ore formations to dissolve and mobilize solid uranium minerals into an aqueous solution. The pregnant, uranium-laden solution is subsequently drawn to the surface via extraction wells for downstream processing, recovery, and concentration. However, managing the chemical balance of this subterranean reaction zone presents substantial operational challenges, as the concentration of uranium in the extracted fluid is nonlinearly governed by the concentration of the injected acid solution.
Operating at an improper acid concentration imposes steep trade-offs on the mining process. When injection acid concentrations are excessively high, side reactions with calcium and magnesium components, such as gypsum precipitation, reduce leaching efficiency and increase acid consumption. Conversely, injecting acid at concentrations below optimal thresholds fails to sufficiently dissolve uranium minerals, slowing reaction kinetics, extending extraction timeframes, and impairing the commercial viability of the recovery site. Most existing mining operations regulate injection acid concentration using empirical threshold-based methods that are inadequately adaptive to dynamic subsurface conditions.
To overcome the limitations of static operational control, researchers evaluated real-time production records gathered between January 2018 and January 2022 from the A1 mining block of an active uranium operation. The four-year dataset recorded daily sampling observations across process variables including injection acid concentration, total injection volume, extraction uranium concentration, residual uranium reserves, instantaneous extraction flow, and extraction volume. Before constructing the predictive architecture, the researchers applied data preprocessing techniques to preserve signal fidelity without introducing lookahead bias. Missing daily entries were imputed using a bidirectional forward-backward filling approach, and extreme outliers were constrained using a quantile-based trimming method that established fixed upper and lower boundary thresholds estimated exclusively on training splits.
To streamline network complexity and eliminate redundant inputs, the researchers performed correlation analyses using the Pearson correlation coefficient to quantify the linear association between operational parameters and extraction uranium concentration. Injection acid concentration, residual uranium, and injection acid volume demonstrated strong correlations with extraction concentration, each exhibiting absolute Pearson correlation coefficients exceeding 0.7. Total injection volume showed a moderate correlation coefficient of 0.553, whereas instantaneous extraction flow and extraction volume demonstrated weak correlations. As a result, weak variables were removed from the input matrix, leaving four core feature variables: injection acid concentration, injection volume, residual uranium content, and injection acid volume. Following feature selection, the data matrix was normalized into the zero-to-one interval and sliced via a sliding window to generate structured three-dimensional time-series tensors.
The neural framework developed in the study couples three distinct architectural modules into a unified multi-model system: a modified multi-scale convolutional module, a stacked recurrent module, and a multi-head attention mechanism. Rather than feeding multi-variable sequences into a single input channel, the model separates each input variable into an independent parallel channel to prevent feature crosstalk and preserve semantic distinctions among operational parameters. The convolutional module uses parallel kernels of varying sizes to simultaneously capture local patterns across different temporal scales, supplemented by a causal dilated convolution layer that broadens the receptive field without violating chronological ordering. Standard fully connected flattening layers were eliminated after the convolutional stages to prevent information loss and preserve temporal sequence continuity.
The temporal sequence modeling stage relies on an optimized gated recurrent unit architecture, which employs a lightweight two-gate structure comprising an update gate and a reset gate to reduce parameter overhead relative to three-gate long short-term memory networks. To model complex nonlinear temporal dynamics and deep lag effects, the design stacks two gated recurrent unit layers with intervening residual connections that mitigate gradient degradation during backpropagation. Additionally, a time-step weighted fusion mechanism introduces learnable weighting parameters across temporal points, allowing the network to prioritize the most influential historical intervals within the input trajectory before generating final latent representations.
To enhance feature discrimination, the architecture incorporates a multi-head attention mechanism inspired by the Transformer design between the convolutional feature extractor and the recurrent layers. Single-head attention networks often lack the representational diversity necessary to isolate coupled interactions across multidimensional industrial features. By projecting queries, keys, and values across multiple parallel representation subspaces, the multi-head mechanism dynamically assigns importance weights to specific variables and temporal steps based on their predictive relevance. This dynamic weighting guides the network to emphasize critical variables and capture key temporal features before passing the selected representations to downstream recurrent layers.
The complete pipeline was implemented in PyTorch using Python on a Windows 11 platform, partitioning the four-year operational dataset into training, testing, and validation splits in a 7:2:1 ratio. Hyperparameter sensitivity investigations revealed that training with the Adam optimizer at a learning rate of 0.001 achieved faster convergence and superior stability compared to higher or lower learning rates. Sensitivity evaluations across discrete input sequence lengths showed that a sequence length of one produced the highest validation performance, with an R-squared of 0.911, indicating that longer historical context windows introduced redundant noise into this specific daily-sampled setting. Model training was capped at 400 epochs, where validation loss converged and remained on a stable plateau through 500 epochs without evidence of overfitting.
Comparative evaluations against standard benchmarks demonstrated the advantages of the unified CNN-GRU-attention architecture. When tested against long short-term memory networks, CNN-LSTM hybrids, and standalone Transformer architectures, the proposed framework demonstrated the lowest errors and the closest tracking of actual uranium concentrations. The baseline long short-term memory model registered the poorest performance, recording a mean absolute error of 0.0767 and a mean squared error of 0.0497. The proposed model achieved an R-squared of 0.915, explaining 91.5 percent of the variance in the test dataset. In an ablation study designed to isolate the contributions of individual components, a standalone GRU baseline produced an R-squared of only 0.783, which increased to 0.856 with the addition of the CNN module and to 0.872 when paired with attention, confirming the synergistic contribution of all three components.
Beyond performance evaluation on static test data, the trained model was applied in a single-variable simulation to map the quantitative relationship between acid injection concentration and extraction uranium concentration under fixed average injection volumes. The simulation revealed a distinct two-phase response: at injection concentrations below 8 grams per liter, uranium extraction concentrations climbed steeply as acid availability stimulated mineral dissolution. Between 8 and 11 grams per liter, the rate of increase began to decelerate, and above 11 grams per liter, extraction concentrations entered a plateau where additional acid inputs yielded diminishing returns. By balancing the point of maximum uranium recovery against reagent consumption and operating costs, the analysis established optimal acid concentration thresholds of 11.69, 10.85, 11.41, and 10.99 grams per liter across residual uranium reserve tiers of 30, 25, 20, and 15 tons, respectively.
The researchers note several operational boundaries and limitations associated with the current study. Because model training and validation relied exclusively on operational records collected from a single mining block, external validity across deposits with differing permeability, lithological stratigraphy, and groundwater geochemistry remains to be verified. In addition, the current feature matrix relies on operational injection and extraction metrics without explicitly incorporating hydrogeological descriptors or lithological classifications, and data points at extreme high or low acid injection levels were comparatively sparse. Future research plans outlined by the investigators include expanding the dataset through collaborative cross-mine monitoring, integrating direct hydrogeological parameters, and implementing periodic retraining pipelines to deploy closed-loop adaptive acid regulation across diverse operational sites.
Subject of Research: Deep learning prediction and acid optimization in in-situ uranium leaching
Article Title: Prediction of extraction uranium concentration and optimization of injection acid concentration using a CNN–GRU–attention model in in-situ uranium leaching
Article References: Wei, Z., Wang, L., Liu, Z., & Zhou, Y. (2026). Prediction of extraction uranium concentration and optimization of injection acid concentration using a CNN–GRU–attention model in in-situ uranium leaching. Journal of Analytical Science and Technology, 17(1), Article 14. https://doi.org/10.1186/s40543-025-00526-9
Image Credits: AI Generated
DOI: 10.1186/s40543-025-00526-9
Keywords: in-situ leaching, uranium extraction, deep learning, convolutional neural networks, gated recurrent units, attention mechanism, multi-head attention, acid injection optimization, time-series prediction, sulfuric acid leaching, mining technology, A1 mining block, predictive modeling, hydrometallurgy, process parameter control, recurrent neural networks
Cite Scienmag News
Blake Davidson. (November 21, 2026). Deep Learning Framework Optimizes Acid Injection Concentration in In-Situ Uranium Leaching. Scienmag. https://scienmag.com/deep-learning-framework-optimizes-acid-injection-concentration-in-in-situ-uranium-leaching/
Blake Davidson. "Deep Learning Framework Optimizes Acid Injection Concentration in In-Situ Uranium Leaching." Scienmag, 21 November 2026, https://scienmag.com/deep-learning-framework-optimizes-acid-injection-concentration-in-in-situ-uranium-leaching/. Accessed 21 November 2026.
Blake Davidson. "Deep Learning Framework Optimizes Acid Injection Concentration in In-Situ Uranium Leaching." Scienmag. November 21, 2026. https://scienmag.com/deep-learning-framework-optimizes-acid-injection-concentration-in-in-situ-uranium-leaching/








