Breakthrough-curve prediction is central to fixed-bed biosorption column design, but data-driven modeling remains challenging with small, unevenly distributed datasets. This study compares a feedforward multilayer perceptron (MLP) with recurrent long short-term memory (LSTM) and bidirectional LSTM (Bi-LSTM) networks for simultaneous prediction of hexavalent chromium [Cr(VI)] and total chromium breakthrough curves in columns packed with Quercus crassipes acorn shell. A dataset of 14 experimental curves was used, covering variations in pH, influent concentration, flow rate, bed height, and biosorbent mass. Model performance was assessed using Mahalanobis distance (D_M)-based holdout selection, curve-shape evaluation, inter-model coefficient of variation (CoV), and SHAP attribution. All architectures achieved high internal cross-validation accuracy (R² = 0.9879–0.9979; RMSE = 0.0149–0.0309). However, complete-curve holdout performance depended on the excluded condition and chromium response. Under the central holdout condition (D_M = 0.317), R² remained high for both outputs (0.954–0.990). In non-central holdouts, errors also affected curve shape, including breakthrough timing and the approach to saturation. The largest deterioration occurred for Cr(VI) under the peripheral low-flow condition, where the MLP decreased to R² = 0.325, while LSTM and Bi-LSTM retained higher agreement. SHAP attribution confirmed process-consistent effects, with influent concentration and flow rate as dominant predictors and pH contributing more strongly to Cr(VI). These findings show that recurrent architectures can provide more robust representations of complex breakthrough profiles, while CoV–D_M analysis helps identify regions where additional experiments are needed.

Heavy metal pollution from anthropogenic sources poses significant risks to ecosystems and human health. Biosorption offers a sustainable removal method, but kinetics are poorly captured by traditional neural networks. This study introduces optimized Bidirectional Long Short-Term Memory (Bi-LSTM) networks for multivariate modeling of Ni(II) biosorption on Quercus crassipes acorn shells, trained using experimental (EKD), synthetic (SKD), and combined (CKD) datasets. A two-stage hyperparameter optimization with Optuna yielded models with R² above 0.995 and low RMSE in 5-fold cross-validation. Second-stage models showed high stability, with coefficient of variation (CoV) values below 10% for RMSE. Based on unseen kinetics, production models showed slightly lower performance (R² = 0.89–0.996): EKD1, EKD2, and CKD1 showed the most consistent performance across challenging conditions with R² values of 0.9617, 0.9769, and 0.9415, respectively; SKD models achieved strong results under standard conditions (kinetic 1, SKD1 R² = 0.9963). SHapley Additive exPlanations (SHAP) analysis identified contact time and initial Ni(II) concentration as key variables, with temperature, cation charge, and a salt interference code also contributing to model interpretability. These findings demonstrate that optimized Bi-LSTM networks offer a robust and interpretable data-driven solution for modeling Ni(II) removal under multivariate conditions, including the presence of salts.