Estimating Forced Variability as a Residual: Application to Basin Mean North Atlantic Sea Surface Temperature

Separating the relative contributions of external forcing and internal variability to North Atlantic Sea Surface Temperatures (NASST) has important implications for attributing and predicting climate changes around the North Atlantic basin. In this work, we approach the separation problem by training a machine learning model to estimate the internal component of basin mean NASST on the basis of NASST patterns orthogonal to the basin mean. With this internal component, the externally forced component is then estimated as a residual. In contrast to many other procedures for separating forced and internal variability, the proposed approach makes no assumption about the temporal structure of forced variability. Since this temporal structure is the subject of active research with no consensus, the proposed approach offers an alternative approach to separating forced and internal variability that avoids such controversial assumptions. In earlier work, we demonstrated that this approach can skillfully estimate basin mean internal variability when trained and applied in preindustrial model simulations. Here, we extend the previous results by training in historical simulations, where the presence of externally forced variability in both the basin mean and orthogonal patterns presents a challenge when estimating basin mean internal variability. This approach is trained and evaluated in a multi-model set of historically forced large ensembles and evaluated alongside other common methods for separating components of variability.