Diffusion of Carbamazepine in Hydrophobic Zeolites: A Comparative Study Using Classical and Machine-Learned Potentials.

Hydrophobic zeolites are promising adsorbents for the persistent pollutant carbamazepine (CBZ), yet diffusion within their confining pores remains poorly understood. Computational studies often rely on static interaction energies, conveying only a static picture. We employ umbrella sampling simulations to obtain free energy surfaces (FES) of CBZ diffusion in a range of all-silica zeolites. Results from a classical force field description are compared with a fine-tuned neural network potential (MACE). While both methods show qualitative agreement, the MACE potential mostly predicts higher activation barriers, which we attribute to the more accurate representation of the energy penalty of close atomic contacts at the transition states. MACE simulations show that CBZ can become kinetically trapped in higher-energy, metastable orientations after a transition state. We propose that exergonic adsorption from an aqueous phase would populate an ensemble of lowest-energy and metastable states, providing a plausible kinetic pathway for diffusion, with a lowering of the effective activation barrier with respect to higher-energy states. The diffusion of CBZ is governed by a complex landscape of translational and rotational barriers, a picture only accessible by going beyond an interaction energy-based description. This work supports the rational choice of shape-selective zeolites as effective adsorbents for environmental remediation.