One2N pushes a 0.7° forecast model to 0.07°: four weeks on one 80 GB GPU, tropospheric skill intact
Synopsis
As Earth-system AI models approach kilometre scales, halving grid spacing quadruples the grid-point count and the end-to-end training cost. One2N splits resolution refinement into successive increments: at each level the complete parent predictor is frozen and only a cross-resolution residual correction module is trained, so the parent's activations and optimizer states are never retained. One refinement level on the NeuralGCM-0.7° baseline yields One2N-0.07°, trained for about four weeks on a single 80 GB A800 GPU; against HRES-0.07° analyses it keeps the parent's tropospheric skill while reducing upper-stratospheric errors and resolving finer structure.
Interpretation
High-resolution global forecasting no longer requires jointly retraining the whole system: an existing coarse-resolution predictor can be frozen as it stands and given one trainable residual correction level on top. Earlier routes either jointly optimized components or resolution levels on the finer grid, or propagated gradients through the dynamical core, so the array and activation growth of the fine grid turned directly into training cost. One2N-0.07°, built from NeuralGCM-0.7° plus one refinement level, was evaluated on 120 matched initializations from the 2023 test set over 10-day forecasts against HRES-0.07° analyses: tropospheric metrics such as Z500, T850 and Q700 stayed broadly comparable to the parent, while lead-time-averaged relative RMSE for geopotential at 1–7 hPa fell by 91.10% to 92.12% and temperature at 1 hPa fell by 59.14%.
The correction module learns residual evolution along the parent trajectory, and the parent supplies conditioning only through a state interface, with its internal architecture unchanged and no differentiability required. Every 2 hours the parent supplies a temporal increment; the module takes the preceding high-resolution residual together with that increment, predicts the next residual, and adds it to the parent forecast mapped to the target grid to reconstruct the high-resolution state. The module is a three-dimensional Swin-Unet that folds the high-resolution residual into 100 non-overlapping tiles along the channel axis, concatenates them with five parent-increment channels, and applies 3D patch embedding and windowed attention; it was trained on 2016–2020 HRES-0.07° analyses with three 2-hour updates inside each 6-hour window, in about four weeks on one 80 GB A800.
Perspective
The demonstration covers one forecasting system, one refinement level (a two-level model) and medium-range forecasting, with training and evaluation following the single path from NeuralGCM-0.7° to 0.07°; of the two starting points, reusing an existing system and training a coarse baseline from scratch, only the former was actually benchmarked.
How total cost scales with resolution and refinement depth remains to be established, since parent forward integration, target-grid state storage and residual-network computation still incur costs; the training target HRES-0.07° is itself an analysis product combining observations with model background, so the learned residual may contain contributions from scale differences, parent-model biases and analysis-system characteristics, and verification against independent observations has not been reported; deeper recursion, the cost advantage of training a coarse baseline from scratch, and refinement on unstructured grids all remain to be evaluated.