disk-galaxy-deprojection 0.4.1

Learned mass-conserving deprojection of galaxy images -> 3D stellar density, rotation curve, potential.

Disk Galaxy Deprojection

This repository builds a baseline-plus-residual probabilistic benchmark for deprojecting barred-galaxy stellar mass structure from S4G-like images.

For project background, current Milestone 2b status, and continuation guidance, start with docs/reports/2026-06-10-project-handoff.md.

Deprojecting a galaxy image — the dgdp package

dgdp turns a galaxy image into its deprojected 3-D stellar-mass distribution, rotation curve, and (optionally) gravitational potential, using a bundled pretrained model — no training data or GPU required.

Install

pip install git+https://github.com/lucyundead/disk-galaxy-deprojection

Core install is light (numpy, astropy, matplotlib) and ships the ~1 MB model. Retraining the model additionally needs the TNG table and the [train] extra (pip install '.[train]').

Quickstart

Python:

from dgdp import deproject
r = deproject("galaxy.fits", distance_mpc=15.2, inclination_deg=30, pa_onsky_deg=153,
              ml=1.0, stellar_mass=6e10, n_samples=48)
r.density_3d              # (nR, nphi, nz) cylindrical stellar-mass cube [Msun]
r.v_circ([1, 2, 5, 10])   # rotation curve at those radii [km/s]
r.scale_height([1, 5, 10])       # sech^2 scale height h_z(R) [kpc] (rho ∝ sech²(z/h_z))
r.scale_height_samples([1, 5])   # posterior draws (n_samples, nR) -> uncertainty bands
r.rms_z([1, 5, 10])       # RMS|z|(R) moment [kpc] (tail-weighted; h_z is the obs-comparable one)
r.v_circ_samples([1, 5])  # same for the rotation curve
r.reproj["history"]       # reprojection-consistency residual (see below)
r.edge_on, r.face_on      # 2-D renderings
r.regrid(n_r=256, n_phi=192, n_z=128)   # same model re-evaluated on a finer grid (hydro ICs:
                          # vertical + m<=4 terms are analytic; only Sigma(R,phi) interpolates)
r.save("out/")            # density.npz + rotation_curve.csv + deprojection.png

CLI:

dgdp-deproject galaxy.fits --distance-mpc 15.2 --inclination-deg 30 --pa-onsky-deg 153 \
    --ml 1.0 --stellar-mass 6e10 -o out/

Geometry & M/L

• Geometry: give pix_arcsec + pa_pix_deg (+ center) to bypass WCS (e.g. an S4G cutout), or let a FITS WCS supply the pixel scale and convert an on-sky pa_onsky_deg.
• The deprojected shape is M/L-independent; M/L only sets the v_c amplitude. Fix the absolute scale with either (a) stellar_mass (total M*, M/L already folded in), (b) luminosity × ml, or (c) zeropoint + band_solar_mag + distance to calibrate the image to a luminosity. With none of these the outputs use a relative scale.

Potential (optional)

r.potential(R, z) and dgdp-deproject --potential return the AGAMA CylSpline potential Φ(R,z). AGAMA is not pip-installable — build it from source (https://github.com/GalacticDynamics-Oxford/Agama) and put it on PYTHONPATH; everything else works without it.

Method

The in-plane surface density is anchored geometrically from the image; a learned fixed-dictionary sech² vertical profile q_m(z;R) — trained on 185 TNG50 barred galaxies × 198 projections — supplies the thickness, mass-conserving by construction. See docs/reports/2026-06-30-mixture-qm-fixed-dictionary.md.

The learned vertical profiles cover m ∈ {0, 2, 4}; the remaining azimuthal content of the surface density (spiral arms, odd m) is carried with the local m=0 vertical profile — being φ-mean-free it changes no ring mass and no RMS|z|(R), but keeps the arms in the maps and lets the reprojection loop drive the image residual to the discretization floor.

Two consistency layers on top:

• Reprojection loop (reproject_iters, default 2): the thick reconstruction is projected back to the sky and the Σ anchors are corrected until the model actually reproduces the observed image (the plain geometric stretch assumes zero thickness). Early-stops with revert, so it never returns a worse-reprojecting model; r.reproj["history"] holds the obs-weighted mean |log(obs/model)| per pass and r.reproj["ratio"] the final residual map.
• Posterior sampling (n_samples): the bundled model is a mixture density network; sampling its head propagates the vertical-profile posterior to RMS|z|(R) and v_c(R) bands (rms_z_samples / v_circ_samples). Band coverage is calibrated on the held-out TNG50 val split (src/dgdp/models/dgdp_fixed_dict.calibration.json).

Milestone 1 predicts posterior residuals for physical summaries:

• radial stellar mass profile
• radial surface-density profile
• vertical scale-height summary
• bar thickness summary
• bar amplitude summary
• central concentration summary

Start with:

python -m pytest -q
python scripts/build_synthetic_benchmark.py --config configs/milestone1.synthetic.toml
python scripts/train_summary_residual_mdn.py --data outputs/milestone1_synthetic/residual_table.npz
python scripts/evaluate_summary_residual.py --run-dir outputs/milestone1_synthetic

Synthetic Milestone 1 Benchmark

Run the synthetic benchmark:

python scripts/build_synthetic_benchmark.py --config configs/milestone1.synthetic.toml
python scripts/train_summary_residual_mdn.py --data outputs/milestone1_synthetic/residual_table.npz --output-dir outputs/milestone1_synthetic
python scripts/evaluate_summary_residual.py --run-dir outputs/milestone1_synthetic

The benchmark writes:

• outputs/milestone1_synthetic/manifest.csv
• outputs/milestone1_synthetic/residual_table.npz
• outputs/milestone1_synthetic/summary_residual_mdn.pt
• outputs/milestone1_synthetic/normalization.npz
• outputs/milestone1_synthetic/metrics.json

Milestone 2 Cluster TNG50 Ingestion

Milestone 2 runs TNG-facing work on the remote cluster through the existing HPC wrapper at /home/lucyundead/codex/hpc-agent/hpc.

python scripts/cluster_dgdp.py sync
python scripts/cluster_dgdp.py check-env
python scripts/cluster_dgdp.py reproduce-milestone1
python scripts/cluster_dgdp.py run-tng50
python scripts/cluster_dgdp.py run-tng50-density
python scripts/cluster_dgdp.py fetch-tng50

The cluster workflow keeps full TNG50 snapshots remote and fetches only compact artifacts under outputs/tng50_milestone2/. If local rsync is unavailable, the project CLI falls back to archive transfer through the HPC wrapper and does not delete remote files.