<p>Polymeric porous materials have a variety of applications ranging from gas storage and catalysis to sensing. A particularly interesting class comprises covalent organic frameworks (COFs), which are often realized as materials consisting of 2D extended, covalently bonded layers stacked on top of each other. The stacking order in layered COFs (LCOFs) is typically&#xa0;not well defined, but impacts virtually all of their properties. In fact, recent studies suggest that stacking does not occur in a strictly periodic fashion. When modeling such disordered COF stacks, the focus so far has been on the static situation, as simulations of the layer-stacking dynamics have either been too expensive (using ab initio methods) or too inaccurate (using generic force fields). This situation has changed with the advent of system-specific, machine-learned potentials (MLPs). In the present contribution, it is described for the prototypical material COF-1, how to parametrize a machine-learned force field that is capable of accurately reproducing the key features of the DFT-calculated potential energy surface. This is possible, even though the energy differences between the local minima are only on the order of a few ten meV per unit cell. The obtained MLP is then used to demonstrate that in molecular dynamics simulations large enough supercells need to be chosen to avoid a massive overestimation of the tendency of layers to slip dynamically. Moreover, the dynamic stability of local minimum structures is tested and the highly anisotropic thermal expansion coefficients of COF-1 are predicted. The present study outlines a strategy that serves as a basis for future computational investigations of dynamical properties of layered materials.</p> Graphical Abstract <p></p>