Skip to content

Dataset inconsistency joint-modeling bottleneck

Home / Open Questions / Dataset inconsistency joint-modeling bottleneck

Background: Jointly training machine-learning models across heterogeneous Earth system datasets often requires reconciling differences in how underlying physical variables are constrained and forced across different reanalysis products.

Question / Future Work: Datasets from different components of the Earth system (e.g., atmospheric reanalysis and ocean reanalysis) are often produced using different forcing data and boundary conditions, creating inconsistencies when used as a unified training set. Developing methods to harmonize these datasets or designing architectures that explicitly account for these cross-component inconsistencies is necessary to prevent performance degradation in near-surface atmospheric variables.

Why It Matters: This bottleneck prevents the seamless scaling of data-driven Earth system models and directly impacts the forecast accuracy of crucial near-surface interactions.

Metadata & Links