Effect of larger MOF pre-training datasets on reconstruction and novel-framework yield

Determine whether pre-training the GLASS autoencoder on larger metal–organic framework datasets improves reconstruction quality and the yield of valid, novel frameworks.

Background

GLASS achieves high structural validity when generating large all-atom metal–organic frameworks, including structures with up to 150 atoms per unit cell. However, its QMOF150 autoencoder exhibits a substantial train–validation reconstruction gap, and most valid generated structures match frameworks in the training set. This indicates that the principal unresolved issue is reliable decoding of structures that are not represented in the training data.

The paper reports that increasing the latent dimension from 32 to 128 does not substantially reduce the QMOF150 reconstruction gap, whereas pre-training on the much larger LeMat-Bulk dataset produces a substantially smaller gap for small crystals. The authors therefore leave unresolved whether larger-scale MOF pre-training can improve both reconstruction generalization and the production of valid, novel frameworks.