Extending MESMER-M to jointly emulate Earth System Model temperature and precipitation realizations

Abstract

Emulators of Earth System Models (ESM) are runtime efficient models that mimic the behavior of an ESM using simple statistical methods. Because of their low complexity, emulators allow to quickly generate thousands of realizations of high-resolution data. Thus, they have proven to be valuable tools for exploring the emission space, quantifying different sources of uncertainty, and investigating extreme events. In this contribution, we introduce an extension to the Modular Earth System Model Emulator (MESMER) for generating monthly precipitation fields. Precipitation is emulated based off monthly temperature such that also the joint precipitation-temperature characteristics match the distribution of the underlying climate model. The emulation consists of two steps. First, the logarithm of precipitation at each location is assumed to depend linearly on temperatures at selected other locations nearby. The selected locations and the linear coefficients are optimized using a Lasso Regression. This step thus yields a deterministic precipitation response that encodes spatiotemporal relationships between precipitation at a given location and temperature at surrounding locations. Second, the residual variability is assumed to be independent from temperature and is modelled as a multi-dimensional noise process containing spatial correlations. The emulator is trained and tested on CMIP6 data. We show that the emulation set-up performs well in simulating the annual cycle, long-term trends in monthly precipitation as well as spatial patterns and natural variability of the underlying climate model. This offers a promising avenue for, as a next step, extending the MESMER emulation framework to other variables.