Verifying spatial generalization of fine-tuned FMC models

Assess whether recurrent neural network models fine-tuned on the Oklahoma field study data generalize to other spatial locations by performing validation with independent datasets from additional sites, since current data from a single location prevents verification of spatial generalization.

Background

The thesis fine-tunes models using FMC observations from a single Oklahoma site over a limited period, introducing risk of overfitting to one location and time. Although early stopping is used to control temporal overfitting, the authors note that spatial generalization cannot be verified with the available data.

They suggest future work using broader field datasets (e.g., FEMS) to address this limitation, highlighting the need for multi-site validation to establish robustness of fine-tuned models across different geographic contexts.

References

However, there is no way to verify whether the fine-tuned models generalize to other spatial locations.

— Time-Warping Recurrent Neural Networks for Transfer Learning  (2604.02474 - Hirschi, 2 Apr 2026) in Chapter 3, Section: Comparison of Transfer Learning Methods for Predicting FMC

Second, transfer to other atmospheric environments remains to be established.

— Predictability-Guided Multiscale Probabilistic Forecasting of Wind Direction under Extreme Shear  (2609.16707 - Shu, 15 Sep 2026) in Section 6, subsection “Limitations and Generalizability” (Section 6.4)