Climate projections from Earth system models are computationally expensive, limiting the number of emission scenarios, ensemble members, and mechanism-oriented experiments that can be explored. Here we present an AI climate emulator that reproduces key capabilities of state-of-the-art CMIP6 models at approximately one hundredth of the computational cost. Trained on simulations from ten CMIP6 models, the emulator performs stable multi-century integrations while preserving model-specific climatologies, inter-model structural differences, and physically reasonable responses to greenhouse gas forcing. It projects twenty-first-century climate change across multiple emission scenarios, reproducing major features including global and regional warming, precipitation redistribution, sea-ice loss, and large-scale atmospheric and oceanic circulation changes. Because the emulator can rapidly generate large ensembles, it enables efficient quantification of projection uncertainty and systematic testing of climate mechanisms. We demonstrate this capability by separating the thermodynamic and sea-surface-temperature-gradient contributions to the projected weakening of the Walker Circulation. By combining multiple climate-model representations within a single computationally efficient framework, the emulator provides a practical complement to conventional Earth system models for scenario exploration, climate attribution, uncertainty assessment, and mechanistic research.

暂无