Skip to content

STATUS.md

Latest commit

 

History

History
50 lines (41 loc) · 2.84 KB

STATUS.md

File metadata and controls

50 lines (41 loc) · 2.84 KB
Raw

ML-RT / STARDUST — progress summary

Shareable overview for the team. Last updated: 2026-07-28.

STARDUST simulation (code revival)

  • Fixed a build-blocking bug so the C code compiles/links with modern toolchains again; verified end-to-end on the test cases.
  • Real bug fixes + cleanups in the file readers, config handling, and physics I/O (including one that had silently rejected valid density profiles); clearer variable names; a ~20% faster interpolation loop (measured).
  • Made the Strömgren-sphere test a runtime config option instead of a compile flag.
  • Modernised setup: rewritten dependency installer, conda env, Apptainer image, docs, and fixed the Python SED-generator/plotting scripts to run under current Python/SciPy.

ML-RT2 — paper 3 (advanced emulators)

Goal: methods barely used in astrophysical radiative transfer yet, kept comparable to papers 1–2, and studied for how they optimise. All prototypes share one data module, trainer, and metrics, so comparisons reflect the method, not the setup.

Eight emulator prototypes (all built + sanity-verified)

Runnable identically via train.py --model <name>:

family model one-line idea
neural operators FNO, DeepONet learn parameters → profile as an operator; resolution-free, fast inference
physics-informed PINO operator + a soft radiative-transfer equilibrium constraint (data-anchored; avoids the old PINN's failure)
generative flow matching modern generative model giving calibrated ensembles / error bars, cheap sampling
transformer transformer attention couples the ionisation-front physics and the four species
stretch Neural-ODE, JEPA, CNP radial "marching" decoder; joint-embedding; per-radius uncertainty

Getting the data

  • Dataset 053 is fetched from Google Drive via gdown (in the conda env): run data/053_data_set/get_data.sh (a one-time Drive file id must be set at the top of the script).

Supporting infrastructure

  • Automatic hyperparameter search (Optuna, cluster-parallel) + 1-D sensitivity sweeps for the "how models optimise" study.
  • Analysis + plotting (emulator-vs-STARDUST profiles; training-dynamics curves).
  • Cluster-ready: conda env + Apptainer image spanning 3090 → H200 hardware; 4-day-safe checkpoint/resume.

Documents to look at (incl. non-coders)

  • docs/methods.pdf — methods brief: what each method is, the key equation, why it's interesting for us, references, a pipeline diagram, and a schematic per architecture.
  • diagrams/*.pdf — standalone schematic for each architecture (same style as the paper 1–2 figures).

Status

All eight prototypes verified locally; ready for the GPU cluster. Deferred by choice: the table_ion.c HeII physics fix (awaiting paper/author review) and extending the PINO residual (helium / heating terms).