Advances in generative climate emulation to support impact-assessment

Policy targets evolve faster than the Couple Model Intercomparison Project (CMIP) cycles, complicating adaptation and mitigation planning that must often contend with outdated projections. Climate model emulators address this gap by offering inexpensive surrogates that can rapidly explore alternative futures while staying close to Earth System Model (ESM) behavior. Here we present recent advances in probabilistic climate emulation aimed to provide inputs for impact models. We show that a generative emulator can reproduce key climate variables at a small fraction of the computational cost of ESMs, while retaining skill in reproducing probability distributions, cross-variable dependencies, time of emergence, and tail behavior. The emulator is informative even for scenarios with aggressive emissions reductions to meet Paris targets. We further show how generative emulators can extend beyond traditional ESMs by directly integrating bias-correction strategies, thereby avoiding separate post-processing steps commonly used in impact assessment pipelines. Finally, we present a framework to design emission scenarios optimized for emulator training, that yields emulators with comparable or improved skill while reducing the volume of ESM simulations needed to train the emulator. We suggest that modeling centers allocate dedicated resources to such "emulator-training" experiments, enabling the rapid generation of large, impact-relevant ensembles across Shared Socioeconomic Pathways while freeing computational capacity for other scientific applications of full-scale Earth system models.