Nathaniel Starkman

How often stellar streams curve away from their host galaxy’s center can tell cold from self-interacting dark matter at up to 5σ, with catalogs the size Euclid, Rubin and Roman will provide.

How often stellar streams curve away from their host galaxy’s center can tell cold from self-interacting dark matter at up to 5σ, with catalogs the size Euclid, Rubin and Roman will provide. In a spherical halo a stream cannot curve this way in projection; in a triaxial halo, which cold dark matter predicts, it can. The test needs only the stream’s projected track and the host’s center: no distances, velocities or model of the stream’s progenitor.

Stellar streams around 13 galaxies in Euclid’s first quick data release give a joint measurement of their dark matter halos’ shape: consistent with spherical, with a mild preference for flattening, q = 0.95 (+0.05, −0.10), in line with ΛCDM.

Stellar streams around 13 galaxies in Euclid’s first quick data release give a joint measurement of their dark matter halos’ shape: consistent with spherical, with a mild preference for flattening, q = 0.95 (+0.05, −0.10), in line with ΛCDM. A stream’s shape on the sky constrains the shape and center of its host’s halo, complementing weak lensing. These galaxies are seen up to about 4 billion years in the past, and Euclid is expected to find thousands more streams.

A physics-informed neural network learns only the corrections to an analytic galactic potential, reaching sub-percent acceleration errors while staying interpretable.

A physics-informed neural network learns only the corrections to an analytic galactic potential, reaching sub-percent acceleration errors while staying interpretable. Its orbits are more faithful than the analytic model’s alone, and Bayesian and time-dependent versions add uncertainties and follow a potential as it changes. Led by undergraduate Charlotte Myers; code in galactoPINNs, first shown at NeurIPS 2025.

JAX gives scientific Python automatic differentiation, compilation and GPUs, but no physical units.

JAX gives scientific Python automatic differentiation, compilation and GPUs, but no physical units. unxt adds them: quantities that carry their units through jitted, differentiated and vectorized code, with astropy.units under the hood and an interface astropy users will recognize. It underpins the GalacticDynamics stack, including coordinax and galax.

Mixture density networks model a stellar stream’s track, width, density and kinematics together with each star’s probability of belonging to it.

Mixture density networks model a stellar stream’s track, width, density and kinematics together with each star’s probability of belonging to it. The method uses all the available astrometry and photometry, models the background with normalizing flows, and works even where phase-space data are incomplete. Applied to GD-1 and Palomar 5, it produced public catalogs of stellar membership probabilities.