A geometry-based method combines persistent homology, ellipsoid packing, and machine learning to estimate molecular shapes in selected organic crystals. The approach offers partial structural information for crystal prediction, while its assumptions and coverage define where it can be applied.