Bias correction of SMILEs: A bulk approach to preserve internal variability

Abstract

Single Model Initial-condition Large Ensembles (SMILEs) represent a pivotal progress in climate modeling, offering multiple simulations from a single model to address the inherent uncertainties in climate projections (Maher et al., 2021). However, biases intrinsic to climate models can distort SMILEs' outputs, potentially misrepresenting climate risks and uncertainties. In climate impact studies, bias correction of Earth System Models (ESMs) typically aligns model outputs with observed historical data, using statistical methods to adjust climatic variables. While essential, this correction may suppress the range of climatic conditions, particularly when applied individually to each ensemble member, thus diminishing the ensemble's diversity and its ability to represent varied climate futures. Instead, we explore whether a bulk approach to bias correction is more appropriate for SMILEs. This method involves applying a consistent correction across the entire ensemble, thereby maintaining the relative differences and natural variability among the ensemble members and preserving the unique capacity of SMILEs to represent a broad spectrum of climatic conditions, in particular under current and near-future climate. Our analysis used the 100-member dataset from the Community Earth System Model Large Ensemble Project Phase 2 (CESM-LENS2, Rodgers et al., 2021), covering historical and future climate simulations. We adjusted key climate variables—precipitation, temperature, relative humidity, and surface pressure within the CONUS domain—using the ISIMIP3basd algorithm (Lange, 2019), with MSWX reanalysis data as the historical reference (Beck et al., 2022). Our experiment involved a twofold comparison: We first evaluated the results after adjusting the entire ensemble at once using (the bulk approach) and, secondly, after adjusting each individual ensemble member separately (member-by-member approach). This comparative analysis allowed us to discern the effects of these two different bias correction methodologies on the ensemble's ability to represent climate variability and extremes. Our results show the effect of both bias correction approaches on the variability of crucial climate extreme statistics and the correlation between ENSO and climate variables. Additionally, we discuss how the choice of bias adjustment method can influence the magnitude of projected changes under future climate scenarios, a key consideration in climate impact studies.