Systematic evaluation and integration of high-energy configurations into training

Investigate how to systematically evaluate and integrate high-energy configurations—including those with positive cohesive energies and very large force magnitudes—into the training of eSEN models co-trained on MatPES and MP-ALOE, so as to better map the repulsive regions of the potential energy surface and provide molecular-dynamics stability guardrails without destabilizing the fit.

Background

During model training, configurations with maximum force magnitudes greater than 800 eV/Å and with positive cohesive energies were excluded to avoid unphysical states. The authors hypothesize that including some of these extreme configurations could help capture the repulsive walls of the potential energy surface and improve MD stability, provided labels are reliable and do not destabilize training.

A systematic procedure to assess and incorporate such configurations is not yet established, and the authors explicitly defer this evaluation and integration to future work.

References

However, we posit that their inclusion could be beneficial for explicitly mapping the highly repulsive walls of the PES. Provided these extreme states possess sound energy and force labels that do not destabilize the model fit, they could serve as valuable physical guardrails during molecular dynamics simulations. We leave the systematic evaluation and integration of these high-energy configurations for future investigations.

— AQVolt26: High-Temperature r$^2$SCAN Halide Dataset for Universal ML Potentials and Solid-State Batteries  (2604.02524 - Kim et al., 2 Apr 2026) in Methods: Machine Learning Model Training