Adsorption kinetics is a critical feature of the pollutant removal process in water purification, wastewater treatment and environmental remediation that can be modeled as dynamical systems. While the classical Langmuir model can serve as a useful first approximation, its accuracy breaks down under heterogeneous conditions or with experimental noise, motivating the exploration of alternative surrogate models that are more flexible. In this work,a compact Neural Ordinary Differential Equation (Neural-ODE) surrogate with only 49 trainable parameters is developed, designed to reproduce the dynamics of classical Langmuir adsorption kinetics. Synthetic trajectories generated from the analytical Langmuir solution were used for training, and a two-stage optimisation procedure (ADAM followed by BFGS) reduced the mean-squared error from \(4.1\times 10^{2}\) to \(5.2\times 10^{-6}\). On noise-free data, the Neural-ODE achieved an RMSE of \(4.13\times 10^{-4}\), corresponding to a relative error of \(0.0413\%\). Across ten independent \(5\%\) Gaussian-noise replicates, the model remained stable with a mean RMSE of \(0.3746\pm 0.0168\). Additional evaluations under proportional noise and structured drift (linear and sinusoidal) further confirmed the surrogate’s robustness, while global sensitivity analysis demonstrated stable performance under uncertainty in the adsorption parameters. The proposed framework was further evaluated on Weibull and stretched exponential adsorption kinetics, demonstrating that the same Neural-ODE architecture can accurately learn multiple adsorption models. Generalization experiments on previously unseen Langmuir parameter combinations yielded a mean RMSE of 0.126, indicating that the model learned a broader family of adsorption dynamics rather than a single trajectory.Furthermore, the proposed framework was validated using both adsorption kinetics reconstructed from experimentally fitted intraparticle diffusion parameters for tetracycline adsorption and an independent experimental breakthrough dataset for diclofenac sodium adsorption on activated carbon. The experimental validation achieved an RMSE of 7.80 mg L\(^{-1}\) and an \(R^2\) value of 0.873, demonstrating the ability of the proposed Neural-ODE to learn experimentally observed adsorption dynamics. These findings demonstrate that compact Neural-ODEs can faithfully and efficiently reproduce nonlinear adsorption dynamics, serving as an accurate surrogate framework for continuous-time adsorption modeling with potential applications in environmental pollution control.
