specsr sharpens low-resolution galaxy spectra by roughly a factor of ten in resolving power (R ~ 100 → R ~ 1000), recovering narrow emission-line structure — including blended doublets such as [O III] λλ4959,5007 and Hβ — that is unresolvable at prism resolution. It is trained on paired JWST/NIRSpec observations from JADES, where each galaxy contributes a prism spectrum and a stitched medium-resolution grating reference.

from specsr.inference import SpecSRPipeline

pipeline = SpecSRPipeline.from_pretrained()      # weights fetched from the Hub
result = pipeline(flux_low, wavelength_low)      # your own wavelength grid

result.sr1           # super-resolved spectrum
result.sr2           # after physics-informed line refinement
result.sr2_sigma     # wavelength-dependent predictive uncertainty
result.z             # inferred redshift, with result.z_sigma
result.wavelength    # the model's grid, in microns

Pass wavelength_low unless your spectrum is already on pipeline.wavelength: the input is then resampled onto the model's grid by integration rather than interpolation, so line flux is preserved. SpecSRPipeline.from_checkpoints(dir) loads a local set of weights instead of the Hub.

Installation

pip install "specsr[hub]"

PyTorch is required. On a GPU machine, install the build matching your driver from the official index before installing specsr, so pip does not resolve a mismatched CUDA stack. Extras: [hub] for automatic weight download, [train] for the training stack, [all] for everything. To track main instead of a release: pip install "git+https://github.com/aryana-haghjoo/specsr.git#egg=specsr[hub]".

Tutorials

Three executable notebooks in tutorials_for_user/, with real held-out JADES spectra bundled so they run with no survey data and no configuration:

How it works

Three stages, trained in sequence, each freezing the one before it:

Every number on the diagram is read off the released checkpoints at draw time by scripts/make_architecture_figure.py, and its inset panels are real predictions for a held-out galaxy — a retrain that changes a width or a depth changes the figure rather than silently invalidating it. ARCHITECTURE.md explains the design and the invariants the code enforces.

What it does and does not do

Deblends features unresolved at prism resolution, improves emission-line S/N for [O II], Hβ, [O III] and Hα, reaches noise-limited residuals above ~2 µm, and recovers redshift information approaching that of the high-resolution reference. It applies a learned prior from the training distribution: it cannot reconstruct features that have no statistical relationship with the prism data.

Limitations, all measured rather than suspected:

• Absolute flux scale is not predicted — only spectral shape.
• Line-flux recovery is a strong function of line brightness. Bright lines keep most of their flux; faint ones are systematically under-recovered. Any flux-accuracy figure quoted without a signal-to-noise cut attached is meaningless.
• Super-resolution does not improve line fluxes over the input. Measured with profile fits, the prism already recovers the total flux of strong lines (median ratio to the reference 0.91–1.09, 2–11% of lines off by more than a factor of two); the reconstruction is lower and more scattered (0.72–0.87, 24–30%). Take total line fluxes from the input spectrum.
• The redshift head is a conditioning stage, not a redshift pipeline. Its catastrophic-outlier rate is high enough that a redshift should be checked against the line positions it implies.
• Super-resolution does not beat direct fitting of the prism on the [O III] doublet-ratio test.
• The paired training sample is small by machine-learning standards, and generalisation to populations underrepresented in JADES — very dusty systems, low-mass galaxies, the highest-redshift sources — is not established.
• Performance degrades below ~1.5 µm, where both the input and the reference have low S/N.

The model card carries the full performance tables, with scatter and sample sizes.

Command line

specsr build-dataset --release DR4      # raw JADES x1d products -> paired dataset

specsr train sr1   --config configs/sr1.yaml --out-dir runs/sr1
specsr train zhead --config configs/zhead.yaml --source sr1 --out-dir runs/zhead \
    --sr1-ckpt runs/sr1/best_superres_model.pth --sr1-config configs/sr1.yaml
specsr train sr2   --config configs/sr2.yaml --out-dir runs/sr2 \
    --sr1-ckpt runs/sr1/best_superres_model.pth --sr1-config configs/sr1.yaml \
    --zhead-ckpt runs/zhead/best_zhead.pth

specsr infer --dataset data/paired_DR4_logR.npz --idx 0 10 42 --save predictions.npz
specsr evaluate line-flux

Upstream checkpoints are named explicitly rather than discovered, so every model records what it was trained on top of. Everything the package writes — figures, predictions, evaluation tables — goes under SPECSR_OUTPUT_DIR (default ./outputs). See the training guide.

Data

Raw JWST/JADES data is not distributed with the package. You need it only to rebuild the dataset or retrain — running the released models on your own spectra requires nothing but the weights, which download on first use.

export SPECSR_JADES_ROOT=/path/to/JADES_data   # contains DR3/, DR4/, ...
export SPECSR_DATA_DIR=/path/to/derived        # where built products are written
specsr build-dataset --release DR4

The data guide covers where to download each release and the directory layout the builder expects.

Splits are drawn over parent galaxies, never over rows. Each galaxy contributes 21 highly correlated rows — one real spectrum plus 20 augmented realisations — and only training galaxies are augmented, so a held-out galaxy contributes exactly its real spectrum and nothing synthetic derived from it. The splitter refuses to return a split that leaks a galaxy across the boundary.

from specsr.data.splits import get_training_split
train_idx, val_idx, path = get_training_split("data/paired_DR4_logR.npz")

Repository layout

src/specsr/          all library code (see ARCHITECTURE.md)
configs/             stage configs and W&B sweep definitions
scripts/             thin entry points: argument parsing and file IO only
tests/               unit tests plus checkpoint-reproduction guards
tutorials_for_user/  executable notebooks with bundled sample spectra
docs/                Sphinx documentation

Documentation is at https://aryana-haghjoo.github.io/specsr/, or build it locally with pip install "specsr[docs]" then sphinx-build -b html docs docs/_build/html.

Citation

Please cite the paper for the method, and the Zenodo record if you need to reference a specific version of the code:

@software{haghjoo2026specsr_code,
  author    = {Haghjoo, Aryana},
  title     = {specsr: physics-informed super-resolution of JWST/NIRSpec prism spectra},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21943196},
  url       = {https://doi.org/10.5281/zenodo.21943196},
  note      = {Concept DOI; resolves to the latest version}
}

@article{haghjoo2026specsr,
  title   = {Learning to See Sharper: A Physics-Informed Artificial Intelligence
             Framework for Super-Resolving Galaxy Spectra},
  author  = {Haghjoo, Aryana and Hemmati, Shoubaneh and Mobasher, Bahram and others},
  journal = {The Astrophysical Journal},
  year    = {2026},
  eprint  = {2603.18357},
  archivePrefix = {arXiv},
  primaryClass  = {astro-ph.GA}
}

License

MIT — see LICENSE.