Multi-modal and multi-resolution integration in a single model

Ascertain whether a single physics-aware molecular foundation model can jointly handle small molecules, proteins, nucleic acids, and materials across multiple resolutions—including quantum mechanics, all-atom, and coarse-grained representations—and determine architectures and training regimes to enable such integration.

Background

Molecular data are heterogeneous across chemical classes and resolutions, spanning QM-level electronic detail, all-atom simulations, and coarse-grained models. The paper explicitly questions whether a single foundation model can unify these modalities and resolutions without losing physical consistency.

Achieving this would require architectures and training strategies that respect symmetries, conservation laws, and scale-dependent behaviors while transferring knowledge across domains.

References

Key open questions: Multi-modal integration: Can a single model handle small molecules, proteins, nucleic acids, and materials at multiple resolutions (QM, AA, CG)?

— Learning Biomolecular Motion: The Physics-Informed Machine Learning Paradigm  (2511.06585 - Deshpande, 10 Nov 2025) in Section 7, Future Directions—Physics-Grounded Foundation Models

The present model also uses a fixed real-space discretization: the larger-cell benchmark tests transfer in volume at unchanged grid spacing, while transfer across spatial resolution and multiscale representations remain open directions.

— Equivariant learning of a transferable three-dimensional classical density functional  (2608.13506 - Cheng, 13 Aug 2026) in Discussion