Deep Potential: Recovering the gravitational potential and local pattern speed in the solar neighborhood with GDR3 using normalizing flows astropy ( Astropy Collaboration et al. 2013 ; Astropy Collaboration et al. 2018 ; Astropy Collaboration et al. 2022 ) , Jupyter ( Perez & Granger 2007 ; Kluyver
Taavet Kalda, Gregory M. Green · 2025
Read Deep Potential: Recovering the gravitational potential and local pattern speed in the solar neighborhood with GDR3 using normalizing flows astropy ( Astropy Collaboration et al. 2013 ; Astropy Collaboration et al. 2018 ; Astropy Collaboration et al. 2022 ) , Jupyter ( Perez & Granger 2007 ; Kluyver by Taavet Kalda, Gregory M. Green (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-03742.txt.
Introduction
Taavet Kalda and Gregory Green map the Milky Way's gravity well around the Sun with a learning machine in a 2025 paper. Gravity encodes where all matter sits, dark matter included, and Gaia's third release shows a messier galaxy than smooth models allow. Their sample holds over five million well-behaved stars.
The method, Deep Potential, assumes only that stars form a collisionless system steady in a rotating frame. A normalizing flow learns the stars' six-dimensional spread, while a neural network carries the gravity field itself. Out fall the local pattern speed, near twenty-eight kilometers per second per kiloparsec, plus local densities for all matter and for dark matter alone. The recovered field ripples across the kiloparsec bubble, a hint that the neighborhood is not as still as the method presumes.
Catalog record
| CREATOR | Taavet Kalda, Gregory M. Green |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 81 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-03742
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Deep Potential: Recovering the gravitational potential and local pattern speed in the solar neighborhood with GDR3 using normalizing flows astropy ( Astropy Collaboration et al. 2013 ; Astropy Collaboration et al. 2018 ; Astropy Collaboration et al. 2022 ) , Jupyter ( Perez & Granger 2007 ; Kluyver et al. 2016 ) , matplotlib ( Hunter 2007 ) , numpy ( Harris et al. 2020 ) , python ( Van Rossum & Drake 2009 ) , scipy ( Virtanen et al. 2020 ; Gommers et al. 2025 ) , scikit-learn ( Pedregosa et al. 2011 ; Buitinck et al. 2013 ; scikit-learn developers 2025 ) , Cython ( Behnel et al. 2011 ) , dustmaps ( Green 2018 ; Green et al. 2024 ) , gala ( Price-Whelan 2017 ; Price-Whelan et al. 2024 ) , galpy ( Bovy 2015 ) , h5py ( Collette 2013 ; Collette et al. 2023 ) , tensorflow ( Developers 2025 ) , tqdm ( da Costa-Luis et al. 2024 ) , CMasher ( van der Velden 2020 ; van der Velden et al. 2024 ) , healpy , HEALPix package ( Zonca et al. 2019 ; Górski et al. 2005 ; Zonca et al. 2025 ) http://healpix.sourceforge.net Taavet Kalda, Gregory M. Green Abstract The gravitational potential of the Milky Way encodes information about the distribution of all matter – including dark matter – throughout the Galaxy. Gaia data release 3 has revealed a complex structure that necessitates flexible models of the Galactic gravitational potential. We make use of a sample of 5.6 million upper-main-sequence stars to map the full 3D gravitational potential in a one-kiloparsec radius from the Sun using a data-driven approach called “Deep Potential”. This method makes minimal assumptions about the dynamics of the Galaxy – that the stars are a collisionless system that is statistically stationary in a rotating frame (with pattern speed to be determined). We model the distribution of stars in 6D phase space using a normalizing flow and the gravitational network using a neural network. We recover a local pattern speed of Ω p = 28.2 ± 0.1 km / s / kpc \Omega_{p}=28.2\pm 0.1\mathrm{\,km/s/kpc} , a local total matter density of ρ = 0.086 ± 0.010 M ⊙ / pc 3 \rho=0.086\pm 0.010\mathrm{\,M_{\odot}/pc^{3}} and local dark matter density of ρ DM = 0.007 ± 0.011 M ⊙ / pc 3 \rho_{\mathrm{DM}}=0.007\pm 0.011\mathrm{\,M_{\odot}/pc^{3}} . The full 3D model exhibits spatial fluctuations, which may stem from the model architecture and non-stationarity in the Milky Way. show][email protected] I Introduction A central goal in galactic dynamics is to determine the spatial distribution of dark matter, which can only be traced through its gravitational effects. One gravitational probe of dark matter is the kinematics of stars in the Milky Way, which are guided by the gravitational potential, sourced in part by dark matter. Directly measuring stellar accelerations induced by the Galactic potential (typically in the order of 1\text{\,}\mathrm{c}\mathrm{m}\,\mathrm{s}^{-1}\,\mathrm{y}\mathrm{r}^{-1} ) is extremely challenging with current technology ( Silverwood & Easther 2019 ) , although some measurements exist for binary pulsars ( Moran et al. 2024 ) and future probes are anticipated ( Chakrabarti et al. 2020 ; Chakrabarti et al. 2022 ) . Consequently, dynamical inferences primarily rely on an instantaneous snapshot of stellar positions and velocities provided by surveys like Gaia . Without further assumptions about the dynamical state of the Galaxy, it is impossible to constrain the gravitational potential from a snapshot of stellar positions and velocities. A common assumption – but not the only one – is that the tracer population is statistically stationary in a chosen reference frame, typically the Galactocentric rest frame. Inserting this assumption into the collisionless Boltzmann equation (CBE) then generally yields an overconstrained system of equations, which one can use to solve for the gravitational potential. After recovering the total gravitational potential, which is sourced by the total matter density, one can obtain the dark matter density by subtracting a model for the baryonic component. Historically, reconstructions of the gravitational potential have relied on simplifying assumptions about its form, such as axisymmetry. However, recent surveys like the Gaia mission ( Gaia Collaboration et al. 2016 ; Gaia Collaboration et al. 2023a ) have highlighted the Milky Way’s disk’s true complexity, uncovering a wealth of non-axisymmetric and disequilibrium features, including the spiral arms, velocity waves, the Galactic warp, and the phase spiral. There is a complex interplay between the internal dynamics and external perturbations that drives the shape and growth of these features, meaning that it is often insufficient to look at one of these features in isolation. To address this, we leverage the six-dimensional phase-space data of stars within one kiloparsec from the Sun and employ the “Deep Potential” method ( Green & Ting 2020 ; Green et al. 2023 ) , which borrows methods from deep learning to model the distribution of matter and the local pattern speed within one kiloparsec from the Sun. For a thorough overview of the literature on Milky Way dynamics, we refer to a recent review by Hunt & Vasiliev 2025 . Traditional methods typically make assumptions about the form of the stellar distribution function (DF) or the gravitational potential, often motivated by a lack of data or its incompleteness. The most prominent approach is to use Jeans modeling, which takes velocity moments of the CBE at a given position to obtain a set of equations directly relating acceleration, and hence the gradients of the gravitational potential, to a set of observables. This is typically coupled with assumptions about spherical or rotational symmetries and isotropy in velocity dispersion ( Binney & Tremaine 2008 ) , though these assumptions can be partially relaxed ( Cappellari 2008 ) . Other distribution function based approaches include action-angle approaches ( Bovy & Rix 2013 ; Ting et al. 2013 ) , orbit-superposition approach ( Schwarzschild 1979 ) , or the Made-to-Measure method ( Syer & Tremaine 1996 ) . One can additionally make use of chemical features as proxies for integrals of motion ( Price-Whelan et al. 2025 ) or for a completely different approach, utilize non-equilibrium features such as the Gaia phase spiral ( Widmark et al. 2022 ; Guo et al. 2024 ) . The motivation of Deep Potential comes from the desire to make use of the full six-dimensional kinematic Gaia data with minimal structural priors. This is achieved by fitting both a non-parametric DF, represented by a normalizing flow (NF), and a non-parametric gravitational potential, represented by a neural network (NN). This is coupled with a robust model for the Gaia selection function in order to account for the incompleteness of the dataset. In contrast with typical approaches for modeling the selection function, we employ a spatially smooth model instead, as it would otherwise produce a non-continuous distribution function, which is difficult to model with normalizing flows. Here we summarize the assumptions of Deep Potential: 1. The motions of stars are guided by a gravitational potential \Phi(\mathbf{x}) . 2. We observe the phase-space coordinates of stars (the “kinematic tracers”) that are statistically stationary in a rotating frame around the Galactic Center. 3. The overall matter density is non-negative everywhere: \rho(\mathbf{x})\geq 0 . We can express