The detectability of climate engineering

Abstract We assess the detection and attribution (D&A) of climate engineering (CE) as a function of their duration after initiation. We employ “surrogate” climates where observations are mimicked by simulations. Unlike classical, stationary D&A, the null hypothesis for this analysis is the nonstationary gradual warming caused by continued greenhouse gas (GHG) forcing, which creates a number of theoretical and technical complications. Adapting D&A to this nonstationary setting requires several ad hoc assumptions whose validity is analyzed and discussed. We study the stratospheric sulfur injection scenarios G3 and G4 of the Geoengineering Model Intercomparison Project. For G3, which smoothly balances global warming with a corresponding cooling, the effect is smaller initially and harder to detect. Temperature and precipitation signals are detectable about a decade after commencing CE and attributable a few years later (details depending on model and scenario). The G4 scenario consists of a continuous injection of 5 Tg SO 2 (roughly one fourth of the Pinatubo eruption per year), which represents a shock‐like forcing that is easier and earlier detectable, just after a few years. Later into the century, uncertainty in GHG sensitivity increasingly dominates the background noise, hampering G4 detection. Spatiotemporal CE fingerprints produce more stable D&A results, with smoother dependence on time. Spatial resolution (within the range of a few spherical harmonics) is less relevant. We argue that especially for early detectability, climate predictions (with proper initialization from observations) are more promising. Many details depend on the choice of climate model for observation and fingerprint. We discuss the potential and limitation of using multimodel ensembles.