A world model of the Sun. 1.16M parameters of MLPs, pooling, and gates.
The task: hide one of the Sun's 9 wavelength images. Reconstruct it from the
other 8 plus the magnetic field. It trains in 2,000 steps, about five minutes
on one GPU.
Layout
model.py the network: encoders, latent graph, dynamics, readout
train.py masked channel training, one flag per lever
lib.py constants and shared utilities
run_eval.py the scorecard: per channel accuracy vs baselines (immutable)
evals/ metrics, baselines, physics probes (immutable)
data/ SDOML v2 fetching, alignment, caching (immutable)
tools/ rendering, cache warming, Ensue publishing
SDOML v2 (Galvez et al.), NASA's machine learning dataset for the Solar
Dynamics Observatory, read from the public S3 bucket gov-nasa-hdrl-data1.
August 2010: 9 AIA wavelength channels and the HMI vector magnetogram,
aligned to a shared cadence and pooled to 256 x 256. Train and validation
days never mix, and the validation days are never touched by training.
Results
Score is a harmonic mean over the 9 channels of accuracy relative to the strongest baseline, on held out days.
Replicated champion: 0.409, std 0.017 across three seeds
Best observed run: 0.456
Built by an autonomous research loop: 35 experiments in one day, 10 record breaks
The largest single gain: the loop found its own learning rate 15x too small and walked it up to the divergence cliff
Ten checkpoints, one day: the score climbing from 0.292 to 0.456 as the dream sharpens. ckpt.pt is the trained champion.
