<p>This study explores the predictive performance of machine learning (ML) algorithms such as support vector regression optimized by genetic algorithm (SVR-GA), random forest (RF), decision tree regression (DTR), and artificial neural networks (ANN) in modeling the adsorption of Cu²⁺, Ni²⁺, and Zn²⁺ ions from contaminated water using activated carbon. Batch adsorption experiments were conducted under varying operational parameters including pH, contact time, adsorbent dose, temperature, and initial concentration. Among the tested models, RFR outperformed others, yielding <i>R</i>² values of 0.97, 0.91, and 0.94 for Cu²⁺, Ni²⁺, and Zn²⁺, respectively. In comparison, SVR-GA achieved <i>R</i>² values of 0.90 (Cu²⁺), 0.89 (Zn²⁺), and 0.89 (Ni²⁺), while DTR showed slightly lower predictive power. ANN models demonstrated strong fitting capability with regression coefficients ranging from 0.95 to 0.99. Adsorption behavior was further analyzed using isotherms, kinetics, thermodynamics, and dimensionless parameters (<i>N</i>_<i>k</i>, <i>λ</i>, <i>φ</i>). The exploration of adsorption dynamics via dimensionless numbers for ternary ions revealed that adsorption on the surface of activated carbon was transfer controlled, with <i>N</i>_<i>k</i> ranging between 10^–4 and 10^–3. Adsorption on homogeneous surfaces may be correlated with chemisorption, with maximum adsorption capacities of 250&#xa0;mg/g, 500&#xa0;mg/g, and 83.33&#xa0;mg/g for copper, nickel, and zinc ions, respectively. Thermodynamic analysis suggested that the adsorption was spontaneous and endothermic. The desorption investigation showed that activated carbon can be utilized for up to three consecutive cycles of desorption/adsorption before saturation. The study confirms the effectiveness of hybrid and ensemble ML models in capturing complex adsorption behavior and provides a robust predictive tool for designing advanced water treatment systems.</p> Graphical Abstract <p></p>