Lynn Zhou: Rapid Forecast‑Based Attribution via a Climate Emulator
Forecast-based attribution is a widely used approach for quantifying the influence of anthropogenic climate change on the severity of extreme events. The method relies on forecast models to reproduce an extreme event under counterfactual conditions (i.e., climate without human influence), typically through modifying initial conditions and external forcings such as atmospheric CO₂ concentrations. Existing forecast‑based attribution systems, e.g. ACCESS-MICAS, are built on physical models that are complicated to configure and computationally expensive to run. These factors substantially restrict their suitability for operationalised rapid attribution service, where statements must be produced routinely and shortly after events occur.
In this study, we explore the potential use of a data-driven model for forecast-based attribution. As an initial step, we focus on a customised version of LUCIE, a Spectral Fourier Neural Operator (SFNO)-based atmospheric emulator with demonstrated long-term autoregressive stability and unbiased climatology. We assess its forecast skill when trained on 10 years of 6-hourly ERA5 data (including temperature, humidity and wind speed at multiple pressure levels) and explore its response to perturbations in the initial state, including cooler and warmer conditions, without retraining the model. The model is then trialled for an extreme event, and the results are compared with ACCESS-MICAS outputs. This work aims to guide future efforts on training the emulator using CMIP and/or ACCESS-MICAS data.