Nash Sabti

Projects

Codes

About Me

I am currently a senior research engineer at Soroco, Boston. I earned my PhD in physics from King's College London, UK. My passion lies in developing and applying state-of-the-art theoretical and numerical frameworks to address pressing questions at the intersection of machine learning, software engineering, and product development.

Besides my day-to-day research, I am also very interested in scientific outreach, teaching and mentoring students. In addition to my work activities, I am very much into music, movies, and everything involving art (if time allows!).

CV.pdf

Contact Me

Nash Sabti
Soroco
123 South St 4th Floor, Boston, MA 02111
Email: nash.sabti@gmail.com

The eras of cosmic dawn and reionization are the new frontiers in our exploration of the cosmos. One way to observe what went on during that time is by looking at emission lines of neutral hydrogen, which is very abundant in the Universe. A prominent example here is the spectral line with a wavelength of 21cm, which is induced by the hyperfine structure of hydrogen, making it a prime target to detect with radio interferometer observatories on Earth, such as HERA. One problem with these kind of observations is that the interaction between the interferometer and the smooth astrophysical foreground emission induces a wedge-shaped region in Fourier space that dominates over the 21cm signal by 5-6 orders of magnitude, leaving only a small window in which the 21cm signal may be cleanly observed.

Instead of modeling the foreground noise down to 1 part in 10^5 (an extremely difficult task), we can take advantage of the inherent non-Gaussian correlations that exist between 21cm modes inside and outside the wedge. These correlations are a result of stars at small scales sourcing the fluctuations we observe at large scales. Neural networks provide a valuable means of learning these correlations, which can then be employed to reconstruct the 21cm signal within the wedge region by analyzing the modes outside of it. In this project, we utilize generative models and condition the inpainting of the 21cm wedge region based on the modes that are cleanly observable.

Stay tuned...

Image credit - Jingchuan Yu, Beijing Planetarium

N-body simulations have been long used to model the gravitational evolution of particles in a box. In scenarios where there are two distinct particle species, and one has a negligible influence on the other (as is the case for example with neutrinos and dark matter), we can independently simulate the evolution of the dominant species without considering the lighter one. This approach is known as the 'N-1 body' method, where our focus shifts to evolving the lighter particles in a pre-computed gravitational potential sourced by the dominating species. The efficiency of this method scales with the number of particles N, making it massively parallelizable. For this purpose, I have developed Cheetah, a GPU-accelerated N-1 body simulation in JAX that is designed to construct the profile of a species clustering onto dark-matter halos over time.

Paper coming soon...

Image - Created with DALL·E 3

The application of generative AI in astrophysics offers a pragmatic solution for simulating galaxies. This technique could be especially promising for testing pipelines for future telescopes to set expectations for new astronomical observations. By utilizing existing images from current telescopes, generative models such as diffusion probabilistic models could learn intricate details about galaxies - their shapes, compositions, and spectra, essentially capturing their 'distribution'. This is very exciting, as it reduces our reliance on costly hydrodynamical simulations to simulate those galaxy properties. In this project, we employed a diffusion probabilistic model (specifically DDPM) in conjunction with the Galaxy10 DECals Dataset (link) to effectively learn the galaxy distribution and generate new, representative images of galaxies. The corresponding GitHub page can be found here.

Image - Output of simulation

Big Bang Nucleosynthesis (BBN) provides us with a window into the very early stages of the Universe, when it was only about 1 second old. How is that possible? Around that time, many of the revelant processes happen to proceed almost out-of-equilibrium, such that even small deviations in the primordial history can lead to big changes in the observed values of the primordial abundances. Given the marvelous agreement between the measured and predicted abundances of helium-4 and deuterium within the Standard Model, BBN has proven to be an important checkpoint in cosmology. And this, of course, has interesting consequences for new physics models of all kinds.

For example, it can be used to constrain thermal dark sectors (1910.01649; see also 2107.11232 for a re-analysis with the updated LUNA rates), deviations in the gravitational constant (1910.10730), and Heavy Neutral Leptons (2006.07387). In the last paper, something interesting is observed: HNLs can decrease the effective numner of relativistic species in the early Universe even if they decay mostly into neutrinos. An explanation of this effect - which happens with any FIMP that decays into neutrinos with high enough energies - is provided in 2103.09831.

Image credit - Astronomy: Roen Kelly

There is still a lot unknown about the conditions of the Universe at high redshifts and small scales. UV galaxy luminosity functions (UV LFs) can provide us with some insightful information in this regards. Indeed, UV LFs tell us something about the abundance of young galaxies in the Universe, which means that they track the evolution of the mass function of dark matter halos, both at earlier times and lower masses than local Universe probes. We are talking here about redshifts 4 to 10-ish, an epoch we don't really know that much about and which makes it an interesting regime to constrain, for example, the nature of dark matter and primordial fluctuations at small scales. There is already some good quality data available from the Hubble Space Telescope, and upcoming telescopes such as the James Webb Space Telescope and Roman Space Telescope, could further improve our current understanding. This of course motivates us to do some exciting physics with this probe!

We tested the grounds with this probe by looking how it can constrain small-scale non-Gaussianities (2009.01245). The word "small-scale" here indicates scales that are smaller than can be probed by Cosmic Microwave Background observations. Interestingly, with current data we can already constrain fNL to be consistent with zero within a few hundred.

In two follow-up papers (2110.13161 and 2110.13168), we presented a galaxy luminosity function pipeline, GALLUMI, that allows for inferring cosmological and astrophysical parameters using Hubble luminosity function data. The code is built as a likelihood code in the MCMC sampler MontePython, which makes it easy to use with other data sets. We found that we are able to measure the clustering of matter up to scales of about k ~ 10 Mpc^-1, making it an interesting addition of current small-scale probes.

Image credit - NASA, ESA, G. Illingworth, D. Magee, P. Oesch, R. Bouwens, HUDF09 Team

We still know quite little about the nature of neutrinos compared to other particle species in the Standard Model. When going back to the early Universe, we have good reasons to believe that neutrinos should have been in thermal equilibrium through weak interactions. Naturally, this then would mean that they follow a Fermi-Dirac distribution. But now the question arises: what can current cosmological probes, in particular the CMB, actually tell us about the properties of neutrinos? In 2111.12726, we demonstrated that in principle the CMB is only able to measure the relativistic energy density and bound the non-relativistic energy density. This means that if the neutrino distribution, for some new-physicsy reason, would deviate from the Fermi-Dirac one, then the cosmological mass bound would change accordingly!

This would have interesting consequences in neutrino experiments like the proposed PTOLEMY experiment that aims to detect the relic neutrino background. In particular, given the currently-existing experimental challenges, a larger cosmological neutrino mass bound would open up a window to more interesting detection prospects at PTOLEMY. We explored this idea and more in 2111.14870.

Image credit - Sanford Underground Research Facility

Before launch, I think most of us already had a feeling that the James Webb Space Telescope (JWST) would revolutionize our understanding of the early Universe. Little did we know that the first-year science results would be this exciting. In case you missed it, observations by JWST seem to indicate that the first few generations of galaxies are not only much heavier than anticipated, but also much brighter. Is our understanding of cosmology flawed? Or could the reason be our ignorance of how astrophysics and data analysis should be dealt with?

In a recent work, 2305.07049, we showed that modifications to cosmology that could explain the JWST observations are in tension with data from the Hubble Space Telescope. The Hubble covers similar redshifts and distance scales as JWST, which means that it is unlikely cosmology can provide a solution to this problem. Our second option is that we simply do not understand the astrophysics at play well enough. This is a harder point to check using the more traditional data such as the UV galaxy luminosity function. This is mainly because different astrophysical solutions exhibit degenerate behavior at this level, which makes it hard to separate them. In a second paper, 2306.09403, we showed that using clustering (bias) measurements allows to disentangle two of the currently popular solutions: a high star-formation efficiency or a large stochasticity in the halo-galaxy connection.

Image credit - Northrup Grumman

So far, everything we know about the Cosmic Neutrino Background is deduced either at the level of background physics (think of Neff and T_nu) or linear-theory perturbations (think of neutrino masses). Given that at least 2 out of the 3 neutrino mass eigenstates have non-zero mass, the non-linear evolution of primordial perturbations would lead to the emergence of neutrino superhalos that are gravitationally bound to dark-matter halos. We took a first step (2306.15715) in studying the potential of detecting such neutrino halos in CMB lensing maps obtained with next-generation CMB experiments. While a more in-depth analysis is required to understand the expected signal-to-noise ratio, detecting these neutrino halos through their gravitational effects would constitute a new, albeit indirect, observational signature of the Cosmic Neutrino Background!

Stay tuned...

Image credit - ESA and the Planck Collaboration