Earth‑system models usually compress atmospheric chemistry into reduced mechanisms because near‑explicit schemes such as the Master Chemical Mechanism (MCM) are prohibitively expensive to solve in every grid cell. Neural emulators promise to lift this limitation, yet most existing approaches advance chemistry in very short time steps and only cover a handful of species or a reduced mechanism.
We introduce FNO‑MoE, a Fourier Neural Operator equipped with a Mixture of Experts. In a single forward call it predicts the full near‑explicit chemical state over a 30‑minute window, encompassing thousands of species.
FNO‑MoE also guides its own training: the disagreement among its experts serves as an informativeness metric to actively select the most valuable box‑model simulations for training. At equal computational cost this active sampling yields a more accurate emulator than random or uniformly spaced sampling.
Fine‑tuning on longer trajectories keeps the model stable over multi‑hour runs, whereas models trained only on short simulations tend to diverge. The resulting emulator reproduces the evolution of thousands of species with high fidelity, remains accurate for unseen chemical compositions, and runs orders of magnitude faster than the traditional numerical solver.
These results open a new direction for emulating complex atmospheric systems: neural operators that predict the full chemical state over long windows and autonomously choose their training data. The next step is to train under varying meteorology, bringing near‑explicit chemistry into full Earth‑system models.
Review