vis-r is a method and the associated code for rapid radial profile modelling of interferometric visibilities. The article in RASTI outlines the method (summarised in the figure below), and this repo contains some code, which may be sufficient for what you want. If it isn't you can roll your own, and if you need help please ask!

The vis-r executable in the install below will model multiple sets of visibilities exported from an ms file (try python hello_world.py in the examples folder). There are stan and python/emcee versions that are generally the same, but with minor differences. Both assume a single disk plane, but can include multiple radial components with independent vertical scales, though these must all be of the same type (e.g. Gaussian, power). Each set of visibilities can have independent offset parameters in the emcee version. Compact central star, point, and Gaussian sources can also be included.

All of this is fairly easy to change for your use case, e.g. a model with multiple disk components could also have independent orientations for each. If you make such changes let me know since they could be helpful for others.

Install the most recent commit on an M2 Mac with conda, which will create the shell command vis-r (frank wants scipy<1.12):

CONDA_SUBDIR=osx-arm64 conda create -n vis-r -c conda-forge python emcee matplotlib scipy=1.11.4 multiprocess arviz corner pytest numpy cmdstanpy ipython arviz jupyter
conda activate vis-r
pip install https://github.com/drgmk/vis-r/archive/refs/heads/main.zip

If you are likely to fiddle with the code, e.g. to add new radial profile functions, it is better to clone the repo into a folder, cd into it, and install vis-r with pip install -e .. Calls to the installed vis-r executable will use the code in this folder.

First, export your visibilities from the ms file to a text or numpy file. Examples that do this are here and here. These will export whatever is in the ms, so discard any calibration and flagged data, and it is best to do some averaging first (e.g. 30s time intervals, a few channels per spw) to keep the size of the exported file down.

Run the fitting on some visibilities like so

vis-r -v data/HD109573.12m*npy -g 0.012 -0.035 26.6 76.5 -p 0.013 1.07 0.06 0.01

or for the stan version, just add the --stan flag:

vis-r -v data/HD109573.12m*npy --stan -g 0.012 -0.035 26.6 76.5 -p 0.013 1.07 0.06 0.01

The options for each version are mostly the same, but there are some differences. The stan version first estimates the parameter scales with the Pathfinder algorithm, and only the emcee version has independent astrometry fitting for each visibility dataset --astrom implemented. Both default to the same number of warmup/sampling steps, but stan will yield many more independent samples for a given number of steps.

To evaluate the model, viewing residuals is normally very helpful. Subtract the best fit model (*vismod.npy) from the ms from which the visibilities were exported, an example function is here.

The suggested method of fitting is to initially use rather hard u,v averaging with the --sz parameter, setting it to something similar to the disk size, or even smaller. Runs should complete quickly, give reasonable parameter estimates, and an idea of whether the model fits the data well via subtraction and creation of dirty images. Given some initial success, more robust posterior estimates can be obtained with larger --sz, and either longer emcee runs or sampling with the stan implementation.