To evaluate the performance of both ERA5 versions, relative errors in output parameters are assessed as compared to the Integrated Global Radiosonde Archive version 2 (IGRA2). For this analysis, the dataset consists of historical atmospheric profile observations from over 200 distributed stations throughout the continental U.S. To ensure a representative comparison, exhaustive quality control is undertaken to address data issues within the IGRA2 dataset. The relative errors are evaluated from a multi-approach perspective, exploring directly observed quantities (temperatures, dewpoint temperature, winds), commonly used atmospheric stability indices (convective available potential energy; CAPE, and convective inhibition energy; CIN), and kinematic parameters. Furthermore, performance for stable and unstable environments is compared to evaluate the relative performance under convective scenarios. Finally, to identify regime-specific biases, we use Self Organizing Maps (SOM), as a clustering technique, to identify the different groups of locations with similar error distributions.
This research enhances the understanding of ERA5's utility and the implications of selecting specific vertical-level versions for future studies. Additionally, it underscores the importance of considering error distributions, topography, and regional variations in utilizing reanalysis data for scientific applications, particularly those related to severe weather conditions. Preliminary results illustrate that the ML version provides a superior option in certain scenarios, particularly close to topography changes or where large vertical or horizontal gradients are found. Such nuances can significantly impact the precision in calculating atmospheric indexes like CAPE or CIN and the position of storms simulated through downscaling.

