Prospector having issues fitting IR filters for a simulated SED

[DhruvZ]

Hi,
I have an issue which I suspect is user error, but I'm not sure how to evaluate it further or if there might be an issue with prospector. I have an SED produced by radiative transfer on a simulation for a z=2 galaxy (I can provide the file to run the script if needed) that I am trying to run prospector on. The script I am running is at https://paste.yt-project.org/show/730/ .

When I attempt to fit the SED with a massive amount of photometry, I am never able to get what I think is an appropriate fit to the far-IR part of the spectrum (plot 1 below, corner plot for reference below that).

However, if I only include the reddest photometric filters to do my fitting, there is no issue with PROSPECTOR reproducing the shape of the dust emission (plot 3 below, associated corner plot below that), though in both cases the metallicity information is not consistent with what I get from the simulation outputs, and it does seem like part of the solution is just a more massive galaxies since there are no lower-wavelength constraints.

I have tried several things, including relaxing the priors I was using previously to what is now on the script, but none of these fixes seem to resolve the problem. I did try this experiment with a simple FSPS galaxy setup, and aside from maybe some minor bias in the inferred metallicity, it was still reproducing the entirely set of photometric filters accurately.

I'm concerned that it might be an issue that PROSPECTOR is having issues because it is focusing too much on the lower wavelength information since there is a higher density of filters in that part of the spectrum and/or that it maybe cannot reproduce what the simulation star-dust geometry might be doing, in which case it would be very hard to resolve.

I'm hoping to check (1) if there is some user error in the code provided and (2) if not, if there is a way to adjust the priors/model setup to enable it to reproduce the provided SED. Additionally (though lower priority), if you have any thoughts on why PROSPECTOR might be biased in metallicity determinations/how to improve that as well, that would be appreciated.

[bd-j]

I suspect that there is a difficulty in reproducing both the shape and normalization of the optical SED and the total absorbed luminosity with a fixed attenuation curve shape. I'm not sure what dust_type you are using but try letting dust_index vary - this will allow you to get a grayer effective attenuation curve, which is to say more absorption for a given optical reddening.

[bd-j]

e.g., note that the mass value is much larger for the IR-only fit. It's pretty concerning though that there is any constraint at all on dust2, logzsol, or the sfh parameters in the IR-only fit. Are these corner plots taking into account the sample weights? I guess dust2 is close to the prior.

[DhruvZ]

Hi, thanks for the response and your good catches on the first results!
A note for your comment about the corner plots - I'm using prospector's own corner plot command, so I assume that factors in the weights, but it depends on what that call does.

I was using dust_type=2 before, which gave the fits I posted before. I tried setting dust_type=0 and letting dust_index vary with a very broad prior but only got this bad fit below.

I instead switched back to dust_type=2 and relaxed the dust2 prior to be a TopHat(mini=0,maxi=5) as opposed to the script I initially posted with a guassian prior and a max value of 2 and instead got this much more appropriate looking result (with the dust2 value basically at the edge of the initial prior).

The last thing I tried was the dust_type=1 and letting dust1,dust1_index,'dust2', mwr, and uvb value all vary with broad priors, and again got this much improved fit.

I'm assuming that these fits appear acceptable. I'd like to avoid any more issues like this, as I'll be running prospector on many simulation SEDs across different redshifts - what do you recommend for dust configurations/priors for good generality/flexibility for this problem?

[bd-j]

Hi @DhruvZ in the first case above it looks like maybe dust2 somehow got removed from the parameters, so there's no dust attenuation and therefore no emission? I had not noticed that the dust2 prior in your script was limited to < 2; you'll know better the context for the simulations you are fitting, but a higher value (or uniform prior) might easily make sense.

Your last version is similar to dust_type=4 and letting dust_index vary, but the added flexibility of of a variable UV bump strength and the dust1 parameters. I don't recommend letting dust1 and dust1_index vary completely independently from dust2 - there are complicated degeneracies between these and the SFH and metallicity (and possibly dust2). But that's really a scientific question that I can't answer for you, as is the question of whether these fits are 'acceptable'. It depends on the purpose. That said, you might look at the "prospector_alpha" set of parameters for some recommended general purpose treatment of dust attenuation, especially for linking dust1 to dust2. See also Leja et al 2017 and the prospector paper.