University of Lubumbashi

Hydrometallurgical extraction of copper (Cu) and cobalt (Co) from oxide ores represents a critical industrial process in the Central African Copperbelt, yet operational variability consistently undermines recovery targets. This study presents a comprehensive data-driven analysis of six months (July–December 2024) of daily operational records from an industrial 8,000-tonne-per-day Cu–Co hydrometallurgical plant located in the Democratic Republic of Congo (DRC). A total of 177 quality-filtered daily observations encompassing 23 process variables were systematically preprocessed, analyzed, and modeled. These variables include crushing and grinding throughput, particle size distribution (PSD), leach feed thickener (LFT) and leach discharge thickener (LDT) parameters, countercurrent decantation (CCD) underflow densities, pregnant leach solution (PLS) concentrations, free sulfuric acid (H2SO4, hereafter free acid, g/L), pH, and electrowinning efficiency. Three machine learning algorithms were deployed: Random Forest (RF), Gradient Boosting (GB), and K-Nearest Neighbors (KNN).Under 5-fold walk-forward cross-validation on normal operating folds (excluding the October 2024 maintenance shutdown period), RF achieved mean Cu R2 = 0.698 ± 0.236 and Co R2 = 0.780 ± 0.175, while GB achieved mean Cu R2 = 0.641 ± 0.126 and Co R2 = 0.795 ± 0.190. On the independent December 2024 temporal holdout test set, RF achieved Cu RMSE = 0.807% (R2 = 0.583) and Co R2 = 0.922 (RMSE = 3.74%), while GB achieved Cu RMSE = 0.851% (R2 = 0.537) and Co R2 = 0.932 (RMSE = 3.48%). RF provided the most reliable discrimination of daily Cu target (≥94%) achievement among the three models (AUC = 0.867, accuracy = 69.6%, F1 = 0.788). The Cu residue ratio, average CCD underflow density, mill throughput, and high-grade PLS (HG PLS) Cu concentration were identified as the dominant predictors of leaching efficiency. Cu leaching met the ≥ 94% target on 72.2% of operating days, whereas the more challenging Co target of ≥ 80% was achieved on only 40.3% of days. High free acid episodes (>8 g/L) were recorded on 61.0% of days, with a statistically significant negative association with Co recovery. Response surfaces for process optimization generated from RF and GB predictions identify operating windows of HG PLS Cu > 10.5 g/L and free acid < 8 g/L as necessary conditions for simultaneously achieving both leaching targets. These findings provide actionable, evidence-based guidance for optimizing plant performance and reducing operational variability in complex Cu–Co hydrometallurgical circuits.

This study presents a comprehensive machine learning framework for the optimisation of metallurgical parameters for the grinding, leaching and solid/liquid separation in a Congolese Copperbelt hydrometallurgical plant. Three deep learning architectures namely Deep Neural Network (DNN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) were designed, trained, and comparatively evaluated. Models were trained to predict Cu and Co leaching recoveries from ten critical physicochemical process variables: feed rate, particle size distribution (PSD), percentage solids at key settling or decantation stages (Leach feed thickener (LFT), Counter current decantation (CCD4), Leach discharge thickener (LDT)), metal losses in overflow streams, and flocculant addition rates. The DNN achieved superior predictive performance with MAE = 1.328 % and R2 = 0.721 for Cu recovery, substantially outperforming CNN (MAE = 2.381 %, R2 = 0.017) and LSTM (MAE = 2.421 %, R2 = − 0.024). Permutation-based feature importance identifies PSD (granulometry) as the dominant predictor (ΔMAE = +0.368), followed by metal losses and thickener underflow densities. A 3-month DNN forecast projects Cu recovery improving from the 2024–2025 mean of 96.0 % to 97.2–97.8 % as gradual PSD and LDT improvements are implemented, with Co recovery reaching 79.8–80.0 %. The critical operational recommendation is to improve milling fineness from the current mean PSD of 61 % to the 75–80 % at 75 µm target range through SAG mill circuit optimisation.

Monopisthocotylan parasites have been proposed as tags for studying host population structure due to their direct life cycles and short generation times. However, their effectiveness in reflecting host population connectivity remains under scrutiny. The poorly-understood connectivity of fisheries stocks in Lake Tanganyika, the second deepest lake in the world with a permanently stratified, relatively species-poor and well-delimited pelagic zone, serves as a case-study to test utility of parasites for fish stock identification. This study investigates the population structure of two Kapentagyrus species parasitizing dorosomatid fish in Lake Tanganyika, by analyzing variation in mitochondrial protein-coding genes across a geographic gradient. The study comprised 12 population genomic datasets from the central and southern subbasins of the lake sampled within the same week to account for host migration. We observed differences in geographic population structure of the two parasite species, with restricted gene flow in Kapentagyrus limnotrissae infecting a single dorosomatid species and smaller hosts preferring the littoral zone. Conversely, K. tanganicanus exhibited no geographical structure, reflecting its broader host range and preference for larger hosts preferring the pelagic zone. The results also highlight that other mitochondrial genes such as those from the dehydrogenase family or atp6 provide higher resolution for population genetic studies in these parasites than the frequently used cox1. The study also compared two sequencing strategies—individual versus pooled sequencing (PoolSeq)—for assessing population structure of monopisthocotylan parasites, and found that PoolSeq yielded similar results with lower demands on individual DNA quantity and sequencing costs. This work supports the use of host-specific, directly transmitted parasites as ecosystem tags and provides valuable insights into the role of host ecology and parasite life-history traits in shaping population dynamics.

In the context of accelerating environmental degradation, human-nature relationships have gained renewed attention. In and around protected areas, these relationships are marked by contestation, war and conflict. This study explores the relationships between residents of Kisandji Village, near Upemba National Park (UNP) in the Democratic Republic of the Congo, and the natural resources they depend on. Using mixed participatory methods, we examined resource use and perceptions of conservation versus livelihoods. Residents emphasized provisioning services as essential to daily needs, identity and security. This dependence, alongside institutional and historical constraints and threats of extractive activities, has shaped a feeling of exclusion and precarity. While some view UNP as limiting their livelihoods, others support conservation when it includes participatory management and acknowledges local needs. The findings highlight that the need for inclusive conservation strategies that align ecological goals with community priorities is even more crucial when other drivers of extractive activities threaten.

Reliable fault diagnosis is essential for condition monitoring of rotating machinery, yet the scarcity of experimental data often limits the effectiveness of machine learning algorithms. A promising approach lies in constructing hybrid datasets that combine experimental and numerical data. However, if performed arbitrarily, such hybridization may degrade diagnostic reliability. This study proposes a systematic methodology for building hybrid databases for ball bearing fault diagnosis across multiple fault severities and varying operating conditions. The methodology is evaluated along three complementary axes: (i) single-condition performance, (ii) expandability to a wide range of speeds through the identification of a reliable variation range, and (iii) transferability of calibration parameters to neighboring regime operating conditions. Both homogeneous (quantitative enrichment) and heterogeneous (qualitative enrichment with unseen fault labels) hybridizations are investigated. Results demonstrate that even a small experimental fraction (approximate to 10%) substantially enhances classification accuracy compared with purely numerical models, while stable performance (>90%) is achieved from about 30% of experimental data. In multi-speed scenarios, hybridization maintains generalization provided sufficient experimental anchoring, though the reliable range remains narrow and asymmetric. In heterogeneous configurations, high accuracy is preserved even with a single EC, as numerical enrichment effectively compensates for missing severities, provided the enrichment is balanced with the operating range.