Evaluating Snowfall Forecasts over the Midwestern and Eastern United States in the GFDL C-SHiELD Model

Abstract Accurately forecasting snowfall remains challenging, and previous studies indicate that models often overestimate snowfall. However, key drivers of this bias are poorly understood. In this study, we evaluate the performance of the Geophysical Fluid Dynamics Laboratory’s (GFDL) Continental System for High-Resolution Prediction on Earth-to-Local Domains (C-SHiELD) model in predicting snowfall amounts over the midwestern and eastern United States, based on 11 snowfall cases from two winter seasons. We establish a physically based method with temperature constraints to identify snowfall grid points from the Stage IV precipitation data and C-SHiELD precipitation predictions, focusing on the snowfall evaluation under a consistently below-freezing condition. Compared to Stage IV, C-SHiELD generally reproduces the spatial distribution of total time-accumulated snowfall composited from the 11 cases but overestimates the amount, with larger biases in the eastern United States than in the midwestern United States. A consistent overestimation of snowfall accumulation is found in all cases, resulting from a larger snowfall area and/or a higher unit areal mean snowfall. This overestimated snowfall in C-SHiELD is attributed to the consistently saturated conditions at lower-tropospheric levels, influenced by temperature and water vapor. The larger snowfall area is mainly due to a cold temperature bias, while excessive water vapor is the main factor leading to the higher unit areal mean snowfall in the model. These findings suggest that improving the accuracy of low-level and near-surface thermodynamic conditions in the model might be essential for better snowfall forecasts.