Lorenz-96 ensemble post-processing — can saved forecasts be repaired without retraining?
Undergraduate research in Prof. Elizabeth Barnes's group at BU: testing whether post-processing closes a variogram-score gap in AI forecast ensembles.
PythonNumPyPyTorch
Private repository
Question
The group’s paper Do AI Forecast Ensembles Sample the Correct Conditional Distribution? (arXiv:2608.08954) compares AI-generated ensembles with a dynamical ensemble on the Lorenz-96 chaotic system. My question: can post-processing saved ensembles, with no retraining, close the variogram-score gap?
Method
Diffusion-model (DDPM) and deterministic-emulator ensembles are compared against the dynamical ensemble, judged by variogram score and CRPS with bootstrap confidence intervals.
Status
In progress through the fall 2026 semester, ending with a presentation to the group. No results are published yet, so none are claimed here.