Ablation study on stricter force thresholds to mitigate trade-offs

Determine whether imposing a stricter maximum-force threshold on the AQVolt26 dataset reduces the observed increase in near-equilibrium formation energy errors while preserving the off-equilibrium robustness of AQVolt26-trained eSEN models, through controlled ablation studies.

Background

Co-training with AQVolt26 improves performance in highly perturbed, off-equilibrium regimes relevant to halide dynamics but introduces a modest shift in baseline predictions for stable configurations, reflected in increased formation energy errors and variance on near-equilibrium tests.

The authors propose that a stricter force threshold during dataset curation might reduce this trade-off and explicitly defer the necessary ablation studies to future work.