“…To quantify the predictive skill of a selected Redi variant, we adopt two statistical metrics: the volume‐weighted Nash and Sutcliffe (
1970) coefficient of efficiency R 2 (see also McCuen et al.,
2006; Ritter & Munoz‐Carpena,
2013) and the volume‐weighted relative error Er , defined as
where χ true denotes the MITgcm‐based “true” tracer content (interpolated onto the MAMEBUS grid), χ param stands for the content of the same tracer predicted by MAMEBUS, and the angle bracket
denotes a spatial‐mean operator across a selected region. Notice that R 2 differs from the coefficient of determination in linear least‐squares regressions, and ranges between ‐ ∞ and unity (Wei & Wang,
2021; Wei et al.,
2022). A “perfect” prediction for χ corresponds to R 2 tending to unity and Er approaching zero.…”