Researchers have introduced a geometry‑based fragmentation technique that splits molecules and polymers into clusters based on distance limits rather than chemical rules. The method treats fragmentation as a radius‑constrained k‑center problem, choosing representative atoms from structures and allowing fragment size to be set by a chosen radius. This produces flexible, multiscale representations. When applied to polymer datasets, models that used the new geometry fragments together with 3‑D conformational information achieved lower root‑mean‑square errors (RMSE) than models that relied only on p‑SMILES strings. The improvement arises because the 3‑D layout, generated by classical force‑field optimisation, gives the model access to spatial interactions that p‑SMILES omit. However, the authors acknowledge that using a single force‑field‑derived conformer can misrepresent the true energetic conformations, and the approach does not yet capture the full ensemble of possible shapes.

The study provides a practical framework for building scale‑adaptive simulations and highlights the importance of 3‑D structure in machine‑learning predictions, while noting the need for more accurate conformer generation.