Emulating transient climate simulations with generative AI

Global climate models (GCMs), like KNMI’s EC-Earth, are an important tool to study the global climate system, and to understand how the climate responds to changes in external forcing. Large ensembles of climate simulations are necessary to separate the forced response from fluctuations due to the climate system’s internal variability (Maher et al., 2021; Muntjewerf et al, 2023). GCMs are computationally very expensive to run, particularly as they move towards the km-scale, which makes generating large ensembles very expensive. 

The generative modelling framework allows the transformation of a base distribution to the target distribution and easily facilitates the construction of large ensembles. We compare two generative models: 1) “stochastic interpolants”, that learn a pseudo-time dependent stochastic process that directly interpolates between the current state and the conditional target state of interest, and 2) a “flow matching” model, that learns a pseudo-time dependent deterministic process, conditioned on the current state, between a Gaussian distribution and the target state of interest.  Both models use a PDE-transformer backbone (Holzschuh et al, 2025). 

We train an emulator to predict global 2m-temperature at time t+1 using the previous 5 days of temperature, the annual global mean temperature and some static spatial and temporal features as conditioning inputs. We make predictions auto-regressively, feeding each prediction back into the model to generate sequences of arbitrary length at inference time. We use Large Ensembles from the EC-Earth3 model, for which a transient 16-member (1950-2166) ensemble and two 160-member time slices (2000-2009, 2075-2085) are available (Muntjewerf et al., 2023). The training dataset consists of up to 5 transient members and we use a single member for validation during training. We use another member for inference to produce an ensemble of global temperature simulations.  

The flow matching model successfully generates a stable ensemble of temperature fields that simulates the long-term forced trend, interannual variability, and spatial patterns of (global) temperature similarly to the GCM. 

 

 

References: 

Maher, N., Milinski, S. and Ludwig, R., 2021. Large ensemble climate model simulations: introduction, overview, and future prospects for utilising multiple types of large ensemble. Earth System Dynamics, 12(2), pp.401-418. 

Muntjewerf, L., Bintanja, R., Reerink, T. and Van Der Wiel, K., 2023. The KNMI Large Ensemble Time Slice (KNMI–LENTIS), Geosci. Model Dev. 16 4581–4597. doi: 10.5194. 

Holzschuh, B., Liu, Q., Kohl, G., & Thuerey, N. (2025). PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations. arXiv preprint arXiv:2505.24717.