Improve force-conditioning mechanisms

Improve the force-conditioning mechanisms in Zatom-2, including conditioning on per-atom force vectors, to capture within-regime force variation across all subsets of the OMol25 and OMat24 data.

Background

Zatom-2's force conditioning distinguishes low- and high-force regimes, but the experiments show that it does not reliably capture precise force values within a regime for all data subsets. The authors explicitly leave improvements to the conditioning mechanism unresolved, identifying per-atom force-vector conditioning as one possible direction.

References

We leave improvements on force conditioning mechanisms, including conditioning on the per-atom force vector, to future work.

— Zatom-2: Multitask Pretraining on Atomistic Data for Generative Modeling across Domains  (2610.11454 - Cretu et al., 8 Oct 2026) in Appendix, Section 'Broader Impacts & Limitations', subsection 'Limitations'