Joint field and latent MLP¶

BottleneckUNetMLPRegressor shares an encoder between a wind-field decoder and an MLP that predicts continuous IBTrACS maximum wind. Spatial mean and max pooling connect the bottleneck feature map to the scalar head. This is the joint representation studied in the paper.

The input has 23 condition channels with ERA5 or 14 without it, plus a condition-validity mask. The decoder predicts a normalized field that is converted to physical wind using the target statistics. The explicit ERA5 residual addition belongs to the field-only model. The scalar head predicts intensity directly rather than taking the maximum of the decoded field.

Targets and loss¶

The data module interpolates IBTrACS USA_WIND to each SAR timestamp only between valid fixes at most three hours apart; an exact fix uses its recorded value. Knots are converted to m/s. Categories are derived diagnostics, not training labels.

The default objective is

\[ L = L_{\mathrm{Huber,field}} + L_{\mathrm{Huber,intensity}}, \]

with transitions of 2 and 5 m/s. Both terms update the encoder; the field term also updates the decoder. Checkpoints use combined val/loss.

Optional radii supervision¶

The structure head predicts five nonnegative values in kilometres: eye size, RMW (called \(R_{\max}\) in the paper), and equivalent-area R34/R50/R64. Each target has an independent validity mask. A missing eye or radius label contributes no loss. Radius-enabled presets add masked Huber loss with a 20 km transition and weight 0.25.

Direct head predictions and radii diagnosed from the decoded image are reported separately; image-diagnosed radii do not contribute to this scalar loss. The published supervision ablation tests the contribution of SAR and radius targets.

Train¶

uv run geo2wf-train experiment=latent_mlp_sar_no_era5_max_wind_radii \
  data.root=/path/to/paired \
  data.ibtracs_file=/path/to/ibtracs.ALL.list.v04r01.csv

Use bottleneck_unet_mlp for the ERA5-conditioned field/intensity model without radius supervision. The preset table lists the other retained combinations. data.stats_file follows data.root in these joint presets. Hardware and logging overrides are described in training.

Encoder-only control¶

BottleneckEncoderMLPRegressor retains the encoder and scalar heads but removes the decoder and field loss. Select a latent_mlp_no_sar_* preset to compare scalar learning with and without SAR supervision. The matched data cohort can still require a SAR-valid center; “no SAR” describes the objective, not a claim that the training samples have no SAR record.