Explain the low-level benefits of omitting pressure scaling

Determine why removing linear pressure scaling from the loss function of the Nested-EAGLE machine-learning weather-prediction model improves or preserves skill in low-level atmospheric temperature and near-surface variables.

Background

Nested-EAGLE uses a loss function that treats all vertical pressure levels identically, unlike models such as GraphCast and AIFS that apply linear pressure scaling to prioritize lower atmospheric levels. Sensitivity experiments reported in the supporting information indicate that removing pressure scaling improves performance in the middle and upper troposphere while also having no negative effect, or even improving skill, for low-level temperature and near-surface variables.

The authors explicitly state that they do not understand the mechanism responsible for the low-level and near-surface benefit. Resolving this issue would clarify how vertical weighting in the training objective affects different atmospheric levels and could inform the design of loss functions for other machine-learning weather-prediction systems.