Latent Dissipative Neural ODEs for Probabilistic Projection of the Forced Climate Response

Climate projections estimate the response of the climate system to external forcing, such as greenhouse gas and aerosol concentrations, and are used to assess future climate change and climate risk. These projections are traditionally generated with Earth system models (ESMs). However the computational cost of ESM simulation severely limits uncertainty quantification, due to small ensemble sizes, and the number of forcing scenarios that can be studied. This has lead to proliferation of computationally efficient statistical emulators that mimic the behavior of expensive ESM simulations, provided they can project beyond the forcing conditions represented in the training data. We present the latent climate attractor (LCA), a compact neural emulator of the slowly varying climate response to external forcing. LCA encodes temperature fields into a truncated discrete cosine transform (DCT) basis and evolves the latent coefficients under a dissipative neural ODE. A simulated internal-variability component around the projected response can later be added in the latent space without retraining. Against the 11-model CMIP6 mean from 2026--2100 under four SSP forcing trajectories, LCA reduces global-mean RMSE by 53\% and energy distance by 42\% relative to the strongest baseline.