Learning from Translation: Seasonal Errors and Feature Importance of the ERA5 Turbulence Predictions
Turbulence is a phenomena that is {\it locally} and statistically characterized by measurements, but it is caused by {\it nonlocal} energy cascades associated with the environment. The presence of turbulence coincides with fluctuations in the refractive index, which impact optical sensing, imaging, and signaling applications. Here, we study the machine learning models that predict near-surface optical turbulence strength $C_n^2$, derived from anemometer-based surface flux measurements through Monin-Obukhov similarity theory, using ERA5 reanalysis data as model inputs. We evaluate the model's ability to perform temporal extrapolation by training on one year of co-located $C_n^2$ observations and ERA5 data, and applying the model to ERA5 data from other years at the same site to reconstruct a multi-year time series. We compare the predictions across Southern California and New York. In spite of varying weather and terrain, the ML models show consistent performance and seasonal behavior across training years. All models show greater correlation, faster convergence, and lower prediction errors in the summer. However, some ERA5 features drive predictions in New York but not California and vice versa, and such feature dependence depends on the season. Seasonal error and feature trends suggest that turbulence is affected by atmospheric composition or other seasonal environmental considerations that are not currently monitored by ERA5. We find, regardless of terrain, the primary feature of importance to turbulence prediction is solar radiation, which underlines the central role of radiative energy transfer in driving atmospheric turbulence. We point toward physics-informed ML translation and feature selection as tools for improving the generalizability of data-driven models.