Astronomy postdoc. Works with machine learning & star clusters.
Queer 🏳️🌈, trans (she/her), vegan.
#Python #Astronomy #Astrodon #MachineLearning #LGBT #Tech
| Location | Germany |
| Website | https://emilydoesastro.com |
| GitHub | https://github.com/emilyhunt |
Emily Hunt
- 1.2K Followers
- 228 Following
- 249 Posts
@markmccaughrean I think just to have a few python environments, pyenv is a smoother experience - conda is very bloated and not necessary in 2025 for any astronomy packages. See guide at https://emily.space/posts/240820-pyenv (which coincidentally includes astropy install instructions)
I now use uv (https://docs.astral.sh/uv/), but it's geared more towards package developers, and if you just want to run python scripts occasionally then pyenv is probably easier
@martinfleis @bkeegan @mszll I think I wrote the blog post aimed at the Conda users I know in astro - most people use it and just use numpy, matplotlib, astropy, and pandas; I think it's complete overkill for them.
Also, conda-forge is hosted by Anaconda still, which costs them 'about $100,000 a month.' Given their recent behaviour, I worry about them deciding to stop doing it for free at some point if everyone switches to it 🙁 https://stackoverflow.com/a/74762864
@martinfleis @bkeegan @mszll that's very fair - how many packages actually need Conda, though? At least for me, I haven't needed Conda in six years of being in science (and I imagine most people in the astronomical data science world don't either). If anything, the incompatibility between pip and Conda has always been a big issue for me, as there are a lot of small and specialist tools not on Conda distros.
@astrojuanlu @bkeegan @mszll I had no idea that uv added a lot more features *today*, oh wow.
I am still waiting for something to become the standard for packaging, though. I don't recommend tools like poetry or pdm to scientists at the moment because they aren't interoperable and it makes it a nightmare to share code with collaborators with a different tech stack, especially as it takes a lot of tech savviness to get around (that most people don't have.)
@astrojuanlu @bkeegan @mszll I explain in the post why I want to recommend one simple tool as a potential alternative. People are welcome to use whatever they want to. Many people still use the main Conda channel, and this post is aimed at them. I also don't think that mentioning every tool ad infinitum is a good idea, uv is still very new too.
Finally, I'm still on the job market - I got really close with a few different postdoctoral fellowships, but it doesn't seem like anything has worked out 😭
Get in touch if you have any postdoc openings. I'd love to keep working on star clusters and/or machine learning!
The paper has been accepted in A&A, but you can read it now on the arXiv!
There's a link to download the data in the comments of the arXiv posting.
Improving the open cluster census. III. Using cluster masses, radii, and dynamics to create a cleaned open cluster catalogue
The census of open clusters has exploded in size thanks to data from the Gaia satellite. However, it is likely that many of these reported clusters are not gravitationally bound, making the open cluster census impractical for many scientific applications. We test different physically motivated methods for distinguishing between bound and unbound clusters, using them to create a cleaned cluster catalogue. We derived completeness-corrected photometric masses for 6956 clusters from our earlier work. Then, we used these masses to compute the size of the Roche surface of these clusters (their Jacobi radius) and distinguish between bound and unbound clusters. We find that only 5647 (79%) of the clusters from our previous catalogue are compatible with bound open clusters, dropping to just 11% of clusters within 250 pc. 3530 open clusters are in a strongly cut high quality sample. The moving groups in our sample show different trends in their size as a function of age and mass, suggesting that they are unbound and undergoing different dynamical processes. Our cluster mass measurements constitute the largest catalogue of Milky Way cluster masses to date, which we also use for further science. Firstly, we inferred the mass-dependent completeness limit of the open cluster census, showing that the census is complete within 1.8 kpc only for objects heavier than 230 M$_\odot$. Next, we derived a completeness-corrected age and mass function for our open cluster catalogue, including estimating that the Milky Way contains a total of $1.3 \times 10^5$ open clusters, only ~4% of which are currently known. Finally, we show that most open clusters have mass functions compatible with the Kroupa initial mass function. We demonstrate Jacobi radii for distinguishing between bound and unbound star clusters, and publish an updated star cluster catalogue with masses and improved cluster classifications. (abridged)
A histogram of all cluster mass function bins (see below) has almost all compatible with expected values from a Kroupa IMF, within a little bit of scatter.
Some notable outliers (probably method issues) are discussed in the paper.
This is really significant because many other cluster mass papers find a varying IMF.
However, we show that almost all variation can be explained by not accounting for selection effects.
Finally, one of the most curious results in this work is that after correcting for selection effects, most clusters have a mass function compatible with the Kroupa IMF.
That's right - **after correcting for selection effects**, the IMF in the Milky Way is pretty much universal!
For the theorists out there, we also set some really cool constraints on cluster formation and destruction with this catalogue!
We derive the first ever Gaia global cluster mass function (+ confirm the Gaia age function too), showing that low-mass clusters are destroyed faster. These results set new constraints on the star cluster formation and destruction rate.
