Structure-Adaptive Generative Atom Substitution

Atom substitution is central to discovering new crystalline materials, and generative models such as Multimodal Crystal Flow (MCFlow) offer a way to automate it by masking a target species and regenerating its atom type. Prior MCFlow substitution experiments kept the host structure frozen during generation, leaving two questions open: does letting the structure relax during generation change which substitutions the model favors, and how should explicit chemical knowledge be combined with the model's learned preferences? We study structure-adaptive generative substitution on NaCl, Fe$_2$O$_3$, and CsPbI$_3$, letting atomic coordinates and lattice parameters co-evolve with the masked atom types, and adding a tunable elemental-similarity guidance term. Structural adaptation does not simply broaden the candidate pool; its effect is host-specific and non-monotonic. For the Na site in NaCl, mild adaptation reverses the leading H/Cl ranking, and stronger adaptation changes the ranking again, whereas the Fe-site distribution in Fe$_2$O$_3$ remains comparatively stable. Chemical guidance behaves more predictably: weak guidance leaves the learned distribution nearly untouched, moderate guidance nudges it toward chemically related elements while preserving much of the model's own ranking, and strong guidance overwhelms the learned prior and collapses diversity onto a handful of hand-favored candidates. The two controls interact — the same guidance strength produces different high-guidance winners depending on how much structural freedom is allowed. This suggests using chemical guidance as a light regularizer on the learned prior rather than as a hard constraint.