Constraining Cosmology from Modeling Strong Gravitational Lenses on Multiple GPU Nodes
Saul Perlmutter, Professor
Physics
Applications for Spring 2026 are closed for this project.
Strong gravitational lenses are very rare occurrences and are a powerful tool in studying dark matter and dark energy, two mysterious entities that together account for 95% of the energy in the universe. The strong lensing team works on a range of projects with cutting-edge AI/ML techniques. We have found ~12,000 strong lenses in the DESI imaging (covering the entire extragalactic sky, http://legacysurvey.org/) and spectroscopic (https://www.desi.lbl.gov/) surveys, using four of the best neural network architectures and other techniques. We have increased the number of known lenses by two orders of magnitude! Our Hubble Space Telescope programs (6 of them!) provide exquisite images for these systems. This allows us to construct detailed models of these lenses in order to better understand dark matter and dark energy, and measure the expansion rate of the universe (H0). To this end, we developed a fast GPU-accelerated lens modeling code (the fastest in the world! https://ui.adsabs.harvard.edu/abs/2022ApJ...935...49G/abstract) and have applied it to our systems with Hubble images (first ever in the world; one paper accepted by ApJ: https://arxiv.org/abs/2502.03455, and another submitted to ApJ: https://arxiv.org/abs/2512.07823). DESI is currently measuring redshifts for all of our lensing candidates (with spectra for 6000 systems already in hand, we submitted one paper, https://arxiv.org/abs/2509.18089, with another in internal review and more to come). Also, we have found an incredible lensing system with 13 (!) background lensed sources (https://iopscience.iop.org/article/10.3847/1538-4357/ad65d3; and see this APS article: https://physics.aps.org/articles/v17/148, and a Scientific American article: https://www.scientificamerican.com/article/epic-gravity-lens-lines-up-seven-galaxy-view/; note that at the time of this paper, we only detected 7 of the 13 sources) -- we have now constructed a more detailed lens model using GIGA-Lens and used it to measure cosmological parameters (paper to be submitted soon).
Finally, we are about to submit a paper on running GIGA-Lens using 128 GPU nodes, or 512 GPUs(!), for the first time in the world. This is an unprecedented time to work in strong lensing to address some of the most fundamental questions in cosmology and physics. Our team has all the requisite data and tools to constrain cosmological parameters in a competitive way for the first time ever!
The strong lensing team has also been highly successful in training undergraduate researchers (here's one example: https://www.hpcwire.com/off-the-wire/nersc-honors-early-career-researchers-with-2022-achievement-awards/; and also see: https://education.lbl.gov/outstanding-mentors-award/). Approximately 40 URAP students have worked with us since 2019. Many are co-authors or lead authors of papers published in ApJ. Well over half of them are in top Physics, Astrophysics, or CS grad/PhD programs around the world, some of whom continue to work with our team!
Role: The student will focus on one or more of the following:
1. Using the Carousel Lens to measure the properties dark energy and dark matter (https://iopscience.iop.org/article/10.3847/1538-4357/ad65d3), using multiple GPU nodes. As mentioned above, the first Carousel Lens cosmology paper will be submitted soon. We already have new data in hand that will allow us to significantly improve the accuracy and precision of cosmological parameter measurements. We anticipate that the resulting constraints will be competitive with those from established cosmological probes such as Type Ia supernovae, BAO, and CMB. This will constitute a second cosmology paper.
2. Strong lens modeling using Hubble images on multiple GPU nodes. This can lead to one or more publications on measuring the mass distribution of the lensing galaxies and/or measurements of cosmological parameters (characterizing properties of dark matter and dark energy) based on lens modeling results.
3. Analyzing DESI spectra for our lens candidates: we want to measure the star formation rates to estimate the rates of lensed supernovae. This will lead to an independent and competitive way of addressing the tension in the H0 measurements.
(For the above three areas, we expect multiple papers.)
4. Applying Multimodal and Agentic AI to Strong Lensing Cosmology.
Qualifications: Proficiency in Python coding a necessity. Experience with numpy/scipy/matplotlib/pandas and TensorFlow/JAX/PyTorch highly desirable, and background in AI/ML a plus.
Preferred but not required (and the order is not so important):
1) experience with parallel (and distributed) computing (we use supercomputers at NERSC, https://www.nersc.gov/, both CPUs and GPUs);
2) experience with statistical sampling methods (e.g., Hamiltonian Monte Carlo, normalizing flows)
3) training neural nets on GPUs (at NERSC);
4) knowledge in gravitational lensing, or even some experience in modeling gravitational lensing systems;
5) experience with any of the following: modeling galaxy light profiles, image coaddition, and transient detection by image subtraction;
6) spectroscopic observations, data reduction, and fitting for spectral absorption and emission lines;
7) experience with web development.
8) simulations of gravitational lenses, supernovae, and gravitationally lensed supernovae for the Roman Space Telescope, James Webb Telescope, and LSST/Vera-Rubin Observatory.
Day-to-day supervisor for this project: Dr. Xiaosheng Huang, Staff Researcher
Hours: 9-11 hrs
Off-Campus Research Site: Lawrence Berkeley National Lab, Building 50. This project can also be carried out remotely with regular Zoom meetings.
Related website: https://www.youtube.com/watch?v=BdpgOhxEyMs
Related website: https://www.youtube.com/watch?v=BdpgOhxEyMs