astronomy/ai produced the result/arXiv 2023 · v2
Neural networks recover a simulated barred galaxy's gravity from one snapshot
Researchers applied a method called Deep Potential to a computer-simulated barred galaxy, using neural networks to learn the stars' motions and then fit the galaxy's gravitational field and the rotation speed of the bar.
spectrum · one line per step, placed by what the step does · bright lines used AI
Recovering the gravitational potential in a rotating frame: Deep Potential applied to a simulated barred galaxy
arXiv, 2023
doi:10.48550/arxiv.2310.00040 · record aix-00087 v2 · checked 2026-10-08
- AI was for
- Structure determination
- Model family
- Normalising flow, Multilayer perceptron
- Checked by
- Held-out3 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Many galaxies, including our own, have a central bar: a long, dense bundle of stars that sweeps round like the hand of a clock. Working out the gravity inside such a galaxy is awkward. Gravity comes from all the matter present, including dark matter, which gives off no light and can only be noticed by its pull. Astronomers would like to read the gravitational field off the positions and speeds of stars. The usual trick is to assume the galaxy is unchanging. A bar breaks that assumption: seen from a fixed viewpoint, the stars keep rearranging themselves as the bar turns.
The way around this is to imagine riding along with the bar. In a frame of reference that spins at the same rate as the bar, the pattern can look steady again. But the right spin rate is not known in advance. The researchers set out to test whether a method called Deep Potential could work out the gravitational field and that spin rate at the same time, from a single frozen snapshot of where stars are and how fast they are moving.
Where AI came in
Two neural networks did the work. The first, a normalising flow, was trained on the positions and velocities of a sample of simulated stars. Its job was to turn a scattered cloud of particles into a smooth, continuous description of how stars are spread through space and speed. This stands in for the statistical smoothing an astronomer would otherwise do by hand, and it has the useful property that its slopes can be read off exactly.
The second network took those slopes and learned a gravitational field, together with the rotation speed of the frame in which the star pattern looks most unchanging. It was tuned not against known answers but against a physical rule about steady motion. Accelerations and the total amount of matter, including dark matter, were then read off from that learned field, and compared with the simulation's true values.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
Deep Potential represents the stellar distribution function of a simulated barred disc galaxy with normalizing flows, then fits a neural-network gravitational potential together with the rotation speed of the frame in which that distribution is most stationary, using only a single snapshot of stellar positions and velocities. Applied to three snapshots of an N-body barred galaxy, the method recovered the bar rotation speed to within 20 %, accelerations to within 20 % and densities to within 50 % in a 16 kpc dataset, and the dark-matter radial profile to within a factor of two for r > 2 kpc. In spherical sub-volumes centred 8 kpc from the galactic centre, the bar rotation speed was recovered to within 20 % for a 2 kpc radius and 15 % for a 4 kpc radius. Constraining the same procedure to the non-rotating laboratory frame left residual non-stationarities along the bar.
How AI was used
Tracer particles (n = 524 288 per snapshot) were drawn from a cylindrical volume of the N-body simulation and split into four concentric cylindrical shells, each padded with virtual particles to smooth the boundaries. For each shell a chain of three FFJORD continuous normalizing flows was trained by maximum likelihood with stochastic gradient descent and Jacobian and kinetic regularization, giving a continuous, auto-differentiable estimate of the six-dimensional phase-space distribution function. m = 2 097 152 phase-space points were then sampled from the stitched flows and the position and velocity gradients of the distribution function computed at each point by automatic differentiation. These gradients were used to train a feed-forward network with four hidden layers of 512 neurons and tanh activations that maps position to gravitational potential, with the frame rotation speed fitted concurrently, by minimising a loss penalising non-stationarity in the rotating frame and negative mass density, with l2 weight regularization. Both fits used the rectified Adam optimiser with warm-up and patience-based learning-rate decay in TensorFlow 2, reserving 25 % of the input data as a validation set. Accelerations and total matter densities were obtained from the gradient and Laplacian of the trained potential by automatic differentiation; the same procedure was repeated for 2 kpc and 4 kpc spherical sub-volumes and their mirrored counterparts.
The shape of the work
Structural · the record, drawn
no AI
Run N-body simulation of barred disc galaxy
Numerical or physics simulation, including where a learned surrogate replaces it.
we make use of an N -body simulation of a collisionless stellar disc that subsequently develops a strong barwhere the paper describes this · verbatim
no AI
Select kinematic tracers and partition into padded shells
Cleaning, filtering, normalising or labelling data already obtained.
we randomly select n=219 (524 288) stellar particles at each time stepwhere the paper describes this · verbatim
AI
Train normalizing flow on stellar phase space
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
we train a normalizing flow f(x→,v→) using stochastic gradient descentwhere the paper describes this · verbatim
AI
Sample flow and auto-differentiate distribution-function gradients
Running a trained model over new data to predict, classify or score.
we draw m=221 (∼2 million) phase-space coordinates, and calculate the gradients ∂∕f∕∂x→ and ∂∕f∕∂v→ at each pointwhere the paper describes this · verbatim
AI
Train potential network and frame rotation speed
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
We concurrently train the parameters of the potential and Ω to minimizewhere the paper describes this · verbatim
AI
Evaluate accelerations and total density from the potential
Running a trained model over new data to predict, classify or score.
We obtain the accelerations predicted by the gravitational model by calculating the gradients of the model via auto-differentiationwhere the paper describes this · verbatim
no AI
Infer dark-matter density profile by subtraction
Extracting understanding from model behaviour.
By considering the modeled density and subtracting the ground-truth baryonic density, we can build an estimate for the dark matter densitywhere the paper describes this · verbatim
no AI
Compare recovered quantities with simulation ground truth
Testing outputs against ground truth.
it is possible to compare the predicted accelerations and densities (from the gradients and Laplacian of the potential, respectively) with the ground truthwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The recovered gravitational potential, total density, bar pattern speed and dark-matter profile are all outputs of the trained normalizing flow and potential network; the paper's results are these outputs.
we train a normalizing flow f(x→,v→) using stochastic gradient descentwhere the paper describes this · verbatim
In all three time steps, we recover the pattern speed of the bar to within ∼ 3 kms−1kpc−1where the paper describes this · verbatim
All of our code, trained models, training data, and simulation snapshots, as well as Python notebooks to generate paper plots, are publicly availablewhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Version of FFJORD normalizing flow (chain of three)Which version of the model was used is not stated.
- Version of Deep Potential gravitational-potential networkWhich version of the model was used is not stated.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00087, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error