In the rapidly evolving field of machine learning (ML) for atomistic simulations of materials and molecules, achieving transferability across different systems remains a key challenge. In this contribution, we demonstrate a case of transfer learning in which an ML model trained exclusively on surface slab data of the cubic ($\alpha$) phase of CsPbI${}_3$ perovskite can reproduce the key features of the dynamic behavior for different phases of the bulk material. Using an equivariant message-passing network (MACE) architecture, we construct an efficient ML force-field trained on the dataset obtained from the molecular dynamics simulations of a CsPbI${}_3$ slab. The resulting model yields stable and reliable trajectories for both slab and bulk, reproducing the experimentally observed phase transitions of CsPbI${}_3$. We demonstrate that pairwise dispersion methods (TS/D3) mispredict the correct phase of CsPbI3 bulk at ambient conditions, while more advanced approaches (MBD-NL/rVV10) recover both the experimentally observed phase sequence with increasing temperature (orthorhombic–tetragonal–cubic) and the corresponding phase-transition temperatures. Through a local analysis of the rotational driving forces acting on PbI${}_6$ octahedra, we show that this discrepancy can be traced to the strong sensitivity of Cs and I atomic forces to the treatment of van der Waals (vdW) dispersion. Overall, our results demonstrate that surface–slab datasets augmented with an accurate treatment of vdW dispersion provide a robust foundation for constructing transferable machine-learning force fields that consistently describe both surface and bulk behavior across multiple length scales.