Improve Zatom-2 energy and force prediction

Improve the energy and force prediction performance of Zatom-2 beyond the accuracy achieved by the reported auxiliary-task training configurations.

Background

Energy and force prediction are auxiliary objectives in Zatom-2 rather than the model's primary purpose, and the reported accuracy remains below that of specialized machine-learning interatomic potentials. The authors explicitly state that the results leave room for improvement and designate further improvement as future work.

References

Lastly, while we treat force and energy prediction as auxiliary tasks for generative training and do not intend to match specialized MLIPs' performance, results in Appendix~\ref{sec:results-mlip-finetuning} leave room for improvement, which we leave for 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'