DiscoNet Training Pipeline

Note

The training scripts (located in training/) are not installed as part of the disco-astronomy package and are not required for end-user operation. They are provided for reproducibility and documentation purposes. CASA (Common Astronomy Software Applications) is required for simulate_catalogue.py.

The DiscoNet model is trained on a hybrid dataset combining simulated ALMA observations (generated via CASA) and on-the-fly synthetic disk images. The training pipeline consists of three independent scripts.

---

Step 1: Catalogue Generation (training/generate_catalogue.py)

Generates a CSV catalogue of 100 synthetic disk parameters with randomised physical properties drawn from the following distributions:

Parameter	Sampling distribution
incl_deg	40% in [0°, 30°] (low inclination), 60% in [30°, 80°] (high).
pa_deg	Uniform in [0°, 180°].
rout_arcsec	35% “compact” in [0.10″, 0.50″]; 35% “large” in [0.60″, 1.80″]; 30% “intermediate” in [0.35″, 0.90″].
rmin_arcsec	Set to a non-zero value (drawn uniformly from [0.05″, 0.55 rout]) with probability 30%; otherwise 0.
band	ALMA Band 6 (70% probability) or Band 8 (30%).
flux_jy	Log-uniform: Band 6 in [5, 80] mJy; Band 8 in [5, 40] mJy.
array_cfg	Band 6: cycle9.6/7/8.cfg with probabilities [0.20, 0.35, 0.45]; Band 8: cycle9.5/6/7.cfg with [0.20, 0.35, 0.45].
pwv	Band 6: uniform in [0.8, 2.5] mm; Band 8: uniform in [0.4, 1.2] mm.
time_s	One of: 1200, 1800, 2400, 3600 seconds (uniform random choice).
niter	Uniform integer in [300, 1500].

Output: catalogo_piloto.csv with columns ID, incl_deg, pa_deg, rout_arcsec, rmin_arcsec, flux_jy, time_s, array_cfg, pwv, niter, band, array_lo.

---

Step 2: FITS Simulation (training/simulate_catalogue.py)

For each row in the catalogue, this script:

1. Generates a synthetic sky model FITS file (create_fits_model) on a 1024 × 1024 pixel grid using a physically-motivated brightness temperature model including:

   • A power-law + exponential radial profile

   • Optional inner cavity with wall brightening

   • Randomised gap-ring substructure (0–5 gaps per disk)

   • Optional crescent asymmetry (20% probability)

   • Optional central unresolved component (80% probability)

   • Multiplicative texture noise (Gaussian-filtered)

   • Outer exponential taper

2. Calls CASA simobserve with the sky model to produce a simulated measurement set.

3. Calls CASA tclean with multi-scale CLEAN to produce a FITS image.

Array configurations and Band parameters used:

Configuration	Beam (mas)	Cell (mas)	Comment
alma.cycle9.5.cfg	130	22
alma.cycle9.6.cfg	80	13
alma.cycle9.7.cfg	50	8
alma.cycle9.8.cfg	28	5

---

Step 3: Model Training (training/train_model.py)

The training script combines two datasets:

• FITSDataset — real FITS simulations from training/simulations/, with online augmentation.

• SyntheticDataset — purely synthetic disk images generated on-the-fly by randomly sampling disk parameters without CASA simulation.

The combined dataset is split 90/10 into training and validation subsets. The model trained is DiskNet (equivalent in architecture to disco.core.cnn_inference.DiscoNet).

Training configuration:

Hyperparameter	Value / Description
Input size	3 × IMG_SIZE × IMG_SIZE (3-channel: normalised intensity, beam map, beam scale map)
Outputs	5: [incl/90, sin2PA, cos2PA, dx/0.14, dy/0.14]
Optimiser	AdamW, lr=LEARNING_RATE, weight_decay=1e-3
LR schedule	Linear warmup (5 epochs) + cosine decay to 10% of initial LR
Gradient clipping	max_norm=2.0
Mixup augmentation	Applied with probability 0.5 per batch
Batch dropout	0.45 (first head layer), 0.30 (second head layer)

The best checkpoint (lowest validation loss) is saved to MODEL_SAVE_PATH (configured in the script) with the following checkpoint schema:

{
    "epoch":       int,
    "model_state": OrderedDict,
    "val_loss":    float,
    "img_size":    int,
    "n_out":       int,
    "outputs":     ["incl/90", "sin2PA", "cos2PA", "dx/0.14", "dy/0.14"],
}

This checkpoint is the file distributed as disco/models/disco_model_stable.pth.