Skip to main content
Back

KNMI's ASUMO Trains Climate Models to Fix Biases No Single Model Could Reach

A new technique called adaptive supermodeling tunes climate model parameters online using synchronization, closing the gap to a perfect model.

Paper, Full Access
  • KNMI researchers introduce adaptive supermodeling, jointly training model parameters and ensemble weights online.
  • Method uses synchronization of chaos plus Adam-based gradient descent, equivalent to the classic delta rule.
  • Tested on the Marshall-Molteni 3-level quasi-geostrophic global atmospheric model with 1449 degrees of freedom.
  • ASUMO reaches within 8% of a perfect model's RMSE, where classical tuning and supermodeling both fail.
  • Only 2.6% compute overhead per parameter, comparable to operational tuning systems like ECMWF's EPPES.
  • No code released; results depend on idealized truth and modest model complexity, real-data tests pending.

Climate models are famously fiddly. They contain dozens of internal knobs whose values cannot be measured directly, and testing whether a given setting actually produces a realistic climate requires integrating the model over decades of simulated time. That combination of a huge state space and long integration windows is why a new paper from KNMI notes that tuning is still largely done by hand, with experts nudging parameters and re-running simulations until the output looks right.

The paper introduces adaptive supermodeling (ASUMO), a method that fuses parameter tuning and supermodeling into a single online training loop that learns from short-timescale weather-like data but improves long-term climate behavior.

Two flavors of climate model correction

There are broadly two ways to reduce climate model bias. The first is to tune internal parameters to improve consistency with observed climatology. The second is to combine several models so their biases partially cancel. Standard multi-model ensembles like CMIP average outputs in postprocessing, which removes some bias but never lets one model's better physics influence another during integration. Supermodeling fixes that by coupling multiple models into a single dynamical system whose members influence each other at every timestep.

A supermodel can be built either by dynamic coupling between members or by taking a weighted average of their tendencies. Until now, researchers have only tuned those coupling weights, leaving the internal parameters of each member frozen. ASUMO trains both at the same time.