Vadim Semenov
Computational astrophysicist
Center for Astrophysics | Harvard & Smithsonian
I am a computational astrophysicist at the Center for Astrophysics | Harvard & Smithsonian, where I have been a NASA Hubble and ITC Postdoctoral Fellow since 2019. I am spending the summer of 2026 at Anthropic as a STEM Fellow. I received my PhD from the University of Chicago in 2019.
I work at the interface between the physics of the interstellar medium and galaxy formation. Together with my collaborators, I design, run, and analyze supercomputer simulations that span scales from individual star-forming regions to cosmological volumes. In my current work, I am also exploring how next-generation numerical simulations can take advantage of AI and machine learning methods.
Modeling turbulent flows on unresolved scales is a general problem in fluid dynamics simulations across domains, from aerospace engineering and climate models to astrophysics. In each case the grid is far too coarse to follow the motions that still shape the large-scale solution, so their effect has to be modeled rather than computed. Such subgrid models are a natural place for next-generation simulations to take advantage of AI and machine learning methods, and to leverage the rise of GPU computing. In my current work I explore how such AI/ML-powered methods can be built into large-scale simulations to incorporate the missing physics, learned from expensive but sparse high-fidelity simulation data.
Why did galaxies in the early Universe form stars so quickly, and how did they settle into disks so soon? The James Webb Space Telescope (JWST) has found far more bright young galaxies in the first billion years than expected, and many already show disk morphologies and kinematics, which fits earlier evidence from the Atacama Large Millimeter/submillimeter Array (ALMA) for cold, rotating gas disks in the young Universe. We find that detailed modeling of turbulence in early galaxies is crucial for explaining both of these at once: the early vigorous evolution, and the transition to settled disks.
Did our Galaxy build its disk unusually early? Modern surveys such as Gaia and H3 let us reconstruct our Galaxy’s history by digging into the kinematics and chemistry of its stars, an approach known as Galactic archaeology. We studied the same archaeological signatures in simulated Milky Way analogs and found that such galaxies form their disks later on average, indicating that the Milky Way is unusual but not rare. Comparisons of this kind let us learn about both the physics of galaxy formation in general and the possible history of our own Galaxy.
Galaxies turn their gas into stars remarkably slowly, much slower than any relevant dynamical timescale would suggest. This has been one of the biggest puzzles in the physics of galaxies for many decades. We developed a framework that ties a galaxy’s overall star formation rate to how long gas spends in each stage of its life cycle on the scales of individual star-forming regions, from assembly into dense star-forming clouds to dispersal by the stars they form. The framework explains many puzzles about star formation in galaxies, including why it is globally inefficient, the emergent scaling relations, and how it is self-regulated in galaxy simulations. The same feedback loop also shapes the structure of the gas between stars, so that structure records how the loop operates and enables us to constrain the physics of this cycle from observations.
Cosmic rays are high-energy particles produced mainly by supernovae. Closer to home, the particle showers they set off in the atmosphere are a known nuisance for electronics, occasionally flipping a bit in computer memory, which is why aviation and spacecraft systems are hardened against them. On the scales of galaxies their effect is far more consequential: they can shape the distribution of gas inside and outside galaxies, and so help control how galaxies evolve. How far and how fast they travel is the critical unknown that limits our understanding of these effects. Observations indicate that cosmic rays move much more slowly near star-forming regions than through typical interstellar gas, likely because they scatter off the turbulent magnetic fields there. We showed that this slower transport changes galactic structure dramatically: in gas-rich, gravitationally unstable galaxies, the cosmic ray pressure that builds up locally prevents gas from collapsing, stabilizing the galaxy as a whole.
Why does one clump of dense gas turn much of its mass into stars while another turns almost none? A large part of the answer lies in how turbulent gas motions shape the internal structure of such regions. These motions and structures exist on scales far smaller than state-of-the-art galaxy simulations can resolve, so they have to be modeled. I develop models for such unresolved turbulence, following the methods outlined above, and test how they shape galaxy evolution on large scales. This approach, also known as Large Eddy Simulation, helps make simulations predictive in extreme regimes where the usual approaches based on calibration against observations break down — for example, the early Universe described above.
Galaxy formation is an inherently multiscale and multiphysics problem. The range of scales involved is enormous, comparable to the range between the size of the Earth and an orange, and those scales are tightly coupled through the feedback loops described above. Together these make the problem hard to tackle numerically. Much of my work develops and improves the numerical methods needed to treat the relevant processes, including many of those described above: turbulence, cosmic rays, star formation, stellar feedback, and the broader algorithms behind adaptive mesh refinement hydrodynamics.
