New preprint on evaluating and guiding crystal structure generation using foundation model representations

New preprint on evaluating and guiding crystal structure generation using foundation model representations.

---

We introduce the Coarse-Fine Transport Distance, a new metric based on compressed feature representations learned by the pre-trained MACE interatomic potential. We show that these representations enable a more meaningful assessment of the stability of generated crystal structures while also providing guidance for generative materials discovery models.

Find the preprint here: https://doi.org/10.48550/arXiv.2607.28776