Modelling multi-scale regional variability with causal networks

Simulation ensembles of the Earth system response to different emission scenarios are a crucial element of climate science. The ability to generate such ensembles using physics-based Earth System Models (ESMs) is limited given their enormous computational and data storage requirements. Lightweight statistical or ML-based emulators calibrated on ESM data successfully capture ESM output for a large range of scenarios. The novel SCALES-MESH emulator framework is designed in a modular way, separately generating the forced response and variability on the regional level (SCALES), then downscaling to gridded emulations using a conditional score-based generative model (MESH).

We here introduce the SCALES variability module, that utilises causal discovery and inference methods to construct natural variability for the emulations. Aggregates of selected variables on the level of IPCC regions are used to derive the interactions between regions and multi-time-lag autocorrelation behavior from ESM data, which are then translated into vector-autoregressive causal network models. With these we can generate ensembles of monthly variability that are able to address key challenges of variability emulation: (i) reproduce the spatiotemporal behavior and relationships of these variables, (ii) in particular also long-range interactions across regions for different time lags, (iii) introduce additional multivariate dependencies, and (iv) superimpose variability on multiple timescales, including multi-annual modes of oceanic variability, which, for instance, (v) enables us to emulate time series of temperature in the Pacific ocean regions that capture ENSO-like patterns.