Mapping of Spacetimes with Neural Networks

arXiv:2611.12857 · gr-qc · Submitted 2026-11-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: General Relativity and Quantum Cosmology. The day in general relativity and quantum cosmology.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Mapping of Spacetimes with Neural Networks".

Mira: Here's the gist of the paper. Mapping of Spacetimes with Neural Networks presents a data-driven neural network method for learning coordinate maps from numerical metric data to a prescribed target gauge The learned map relates numerical spacetimes to analytical backgrounds, with potential applications to cross-code validation and black hole perturbation theory Learning Coordinate Maps from Metric Data The neural network is trained by matching the transformed source metric to the target metric without assuming a functional form for the coordinate transformation The method first validates itself on analytical Schwarzschild and Kerr geometries, recovering known transformations up to the isometries of the target spacetime For a non-spinning black hole, the learned map relates the time-dependent numerical metric to a stationary Schwarzschild metric in ingoing Eddington–Finkelstein coordinates For a spinning puncture, the method provides an approximate identification with a Kerr background despite departures from Kerr in the numerical spacetime Gauge Freedom and Coordinate Mapping The gauge freedom of GR complicates the interpretation of simulation coordinates and the comparison of different descriptions of the same spacetime Cross-code validation typically relies on quantities extracted at future null infinity I such as the Weyl scalar ψ4 or the GW strain obtained via extrapolation or Cauchy-characteristic extraction Establishing correspondence requires an explicit coordinate map in the bulk spacetime The network learns this coordinate map by matching the transformed numerical metric to a prescribed target metric without assuming a functional form for the gauge transformation Neural Network Architecture and Training The problem of finding an unknown smooth map between two coordinate systems is addressed using fully connected feedforward neural networks The network takes as input the four source coordinates xNR and returns the transformed coordinates x′target The network is trained by minimizing a loss that penalises the mismatch between the pulled-back metric and the prescribed target metric The architecture uses SiLU activation in hidden layers, ensuring that the network represents a smooth coordinate map whose Jacobian varies smoothly with the inputs Results and Applications The analytical tests show that the network recovers known coordinate transformations for analytical Schwarzschild and Kerr metrics up to isometries of the target spacetime In numerical cases, the learned map brings the time-dependent moving-puncture metric into ingoing Eddington–Finkelstein (EF) form for non-spinning BHs For spinning punctures, the method jointly fits the coordinate map and the mass and spin of a Kerr background in KS coordinates to construct initial data for Teukolsky evolution The resulting waveforms closely match those extracted from the full nonlinear simulation Limitations and Future Directions The residual mismatch contains physical deviations from Kerr together with numerical and fitting errors A limitation is that the present network architecture does not enforce invertibility explicitly, which could be done using a similar approach used in

77: Furthermore, the loss function could be better adapted to the different scales of the metric components for more balanced accuracy across the domain The perturbative application has additional sources of uncertainty, such as numerical noise from curvature extraction A natural extension is comparing simulations produced by different numerical relativity codes to make comparisons of local spacetime quantities more direct The method provides a measure of the remaining metric mismatch after gauge alignment from the NN training and help quantify deviations from the Kerr background Acknowledgments We would like to thank Abhay Ashtekar, Hengrui Zhu and Vaishak Prasad for helpful comments and discussions ES acknowledges support from The Charles E.

Kai: Kaufman Foundation of The Pittsburgh Foundation SB acknowledges funding from the EU Horizon under ERC Consolidator Grant no.

Mira: InspiReM-101043372 and from the Deutsche Forschungsgemeinschaft (DFG) project “GROOVHY” (BE 6301/5-1 Projektnummer: 523180871) This research used resources of the National Energy Research Scientific Computing Center (NERSC), a Department of Energy User Facility using NERSC award ERCAP0039050 Additional computations were performed on the Roar supercomputer at the Penn State Institute for Computational and Data Sciences (ICDS) Claude AI was used to help with coding and plotting scripts.

Kai: Appendix C: Spacetime Metrics in General Relativity Schwarzschild Metric a.

Title and authors: Kai: So, let's talk about the title "Mapping of Spacetimes with Neural Networks" and who wrote it. This paper introduces this data-driven method for learning coordinate maps from numerical metric data to a prescribed target gauge.

Mira: The authors are Shukla, Radice, Gamba, and Bernuzzi from various institutions at Penn State, Berkeley, and Jena. They’re researchers in different areas of physics but they’ve clearly put together a method that bridges numerical relativity with analytical geometry.

Lev: I’m interested in the fact that they are using a neural network to learn this map without assuming any functional form for the coordinate transformation initially. That takes away a huge amount of guesswork, which is nice because we often have to make those assumptions just to get started.

Kai: It’s about matching the transformed source metric to the target metric directly, which is essentially learning a diffeomorphism from data rather than deriving it through some fixed mathematical procedure like solving field equations for the transformation itself.

Mira: The implication here is that we can bypass some of those tedious, purely analytical steps required for gauge fixing when comparing different numerical solutions.

Lev: I’m wondering how they handle the gauge freedom inherent in general relativity when setting up this matching process; that's usually where things get messy in simulations.

Kai: They address that complication by proposing this learning approach to directly learn the map from the metric data itself, instead of relying on fixed gauge conditions applied beforehand.

Mira: So, they are trying to automate the identification of coordinates between two representations of the same physical reality based purely on how the metrics look numerically.

Lev: That sounds like it could be a powerful diagnostic tool for testing different numerical relativity codes against each other, because if we can map them onto a common gauge space, we can see exactly where they diverge.

Kai: And they validate it first by recovering known transformations for Schwarzschild and Kerr geometries, which is the necessary check to make sure the basic machinery works before scaling up.

Mira: It’s a smart way to build trust in the method because you're confirming it on systems you already have exact solutions for.

Lev: But then they move into more complex things like time-dependent metrics, which is where the real headache usually starts in simulations.

The paper's summary: Kai: Now let’s go over what the paper actually summarizes: it outlines this data-driven neural network method for learning coordinate maps from numerical metric data to a prescribed target gauge.

Mira: The core idea is that instead of assuming a functional form for the coordinate transformation, the neural network learns it by matching the transformed source metric to a prescribed target metric.

Lev: They first show this works on analytical Schwarzschild and Kerr geometries, recovering known transformations up to isometries of the target spacetime.

Kai: Then they apply this to numerical cases, showing that for a non-spinning black hole, the learned map relates a time-dependent numerical metric into ingoing Eddington–Finkelstein coordinates.

Mira: For spinning punctures, it provides an approximate identification with a Kerr background even when the numerical spacetime has departures from Kerr.

Lev: They also mention that this approach has a natural application in black hole perturbation theory, specifically for schemes like the Lazarus project where they need to identify the numerical spacetime with a Kerr background to construct initial data for quantities like Psi four and its time derivative for evolution with the Teukolsky equation.

Kai: So, it’s summarizing that this method is not just an analytical trick; it’s a practical tool for relating complex numerical simulations to standard theoretical backgrounds.

Mira: The implication is that this shifts how we think about comparing different descriptions of spacetime from needing explicit coordinate maps to having the network learn them implicitly from the metric data.

Lev: It means that instead of spending time manually deriving these maps, we let the data guide us toward a physically relevant gauge alignment.

The paper's improvements: Kai: Let’s look at what they suggest are the specific improvements to this method. They are focused on making it more robust and accurate for real-world use.

Mira: One improvement is the goal of enhancing cross-code validation by enabling point-wise comparison of bulk spacetime between codes using this learned coordinate map.

Lev: That’s interesting because they suggest this could provide a complementary diagnostic to asymptotic methods, helping to isolate exactly where in the spacetime two codes begin to disagree.

Kai: They also propose improving black hole perturbation theory by allowing the neural network to jointly fit the coordinate map and Kerr background parameters.

Mira: This allows for an approximate identification with a Kerr background despite departures from Kerr in the numerical spacetime, and it gives us initial data for evolution with the Teukolsky equation.

Lev: And they suggest developing a systematic alternative to existing coordinate identification methods by optimizing an explicit measure of agreement between metrics over the training domain while simultaneously fitting background mass and spin.

Kai: That means they’re aiming for a better measure of remaining metric mismatch after gauge alignment from the neural network training, which is more concrete than just saying "it's close."

Mira: They also mention making it more robust against coordinate singularities by explicitly building invertibility into the neural network layers, potentially using an approach similar to what was used in another paper.

Lev: That’s a big deal for stability; if we can control the map better, we get more robust control over what coordinates we're actually using for our calculations.

Kai: And they also suggest improving the loss function sensitivity by adapting it to different scales of metric components instead of just averaging squared absolute differences.

Mira: That way, components with large coordinate-dependent magnitudes can dominate the optimization, which should give more balanced accuracy across the whole spacetime domain.

Conclusion: Kai: So to wrap up on "Mapping of Spacetimes with Neural Networks," this paper presents a data-driven neural network method for learning coordinate maps from numerical metric data to a prescribed target gauge.

Mira: The main implication is that this lets us relate numerical spacetimes to analytical backgrounds without assuming a functional form for the transformation.

Lev: It’s showing that we can recover known transformations on Schwarzschild and Kerr geometries, and in numerical cases, it successfully maps time-dependent metrics into ingoing Eddington–Finkelstein form.

Kai: And for spinning punctures, it gives us an approximate identification with a Kerr background when things are off-axis.

Mira: This has potential for cross-code validation and helps with black hole perturbation theory by providing initial data for evolution with the Teukolsky equation.

Lev: To weigh in from an engineering side, we need to acknowledge that there’s still that residual mismatch containing physical deviations from Kerr along with numerical and fitting errors.

Kai: A limitation they state is that the present network architecture doesn't explicitly enforce invertibility, which could be done using a similar approach used in other work.

Mira: They also point out that the loss function could be better adapted to different scales of metric components for more balanced accuracy across the domain.

Lev: And another limitation they mention is that if we can't control the map explicitly, we lose some of that robustness against coordinate singularities.

Estuti Shukla, *David Radice, *Rossella Gamba, *Sebastiano Bernuzzi

Institute for Gravitation and the Cosmos, The Pennsylvania State University · Department of Physics, The Pennsylvania State University Department of Astronomy & Astrophysics, The Pennsylvania State University Department of Physics, University of California Berkeley Department of Physics, Friedrich-Schiller-Universität Jena

gr-qc

Submitted: 2026-11-09

Updated: 2026-11-09

Comments: 28 pages, 18 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Mapping of Spacetimes with Neural Networks presents a data-driven neural network method for learning coordinate maps from numerical metric data to a prescribed target gauge The learned map relates…

Key concepts

Learning Coordinate Maps
The neural network learns the function that transforms numerical spacetime coordinates into a desired analytical coordinate system, such as Schwarzschild or Kerr. It does this by training on many examples, matching how the numerical metric looks when viewed through different coordinate systems.
Gauge Freedom
In General Relativity, different coordinate systems can describe the same physical spacetime in different ways due to gauge freedom. This makes comparing simulations difficult. The network learns an explicit map to align these descriptions, helping researchers compare results from different numerical relativity codes.
Neural Network Architecture
Fully connected feedforward neural networks are used to find the unknown smooth map between coordinates. The network takes the four source coordinates and outputs the transformed target coordinates, using SiLU activation to ensure it learns a smooth transformation.
Cross-code Validation
This technique uses the learned coordinate map to compare spacetimes generated by different numerical relativity codes. By mapping both codes' metrics into a common target gauge, researchers can directly see where the two simulations disagree in terms of spacetime geometry.

Terminology

Summary

Mapping of Spacetimes with Neural Networks presents a data-driven neural network method for learning coordinate maps from numerical metric data to a prescribed target gauge The learned map relates numerical spacetimes to analytical backgrounds, with potential applications to cross-code validation and black hole perturbation theory

Learning Coordinate Maps from Metric Data

The neural network is trained by matching the transformed source metric to the target metric without assuming a functional form for the coordinate transformation The method first validates itself on analytical Schwarzschild and Kerr geometries, recovering known transformations up to the isometries of the target spacetime For a non-spinning black hole, the learned map relates the time-dependent numerical metric to a stationary Schwarzschild metric in ingoing Eddington–Finkelstein coordinates For a spinning puncture, the method provides an approximate identification with a Kerr background despite departures from Kerr in the numerical spacetime

Gauge Freedom and Coordinate Mapping

The gauge freedom of GR complicates the interpretation of simulation coordinates and the comparison of different descriptions of the same spacetime Cross-code validation typically relies on quantities extracted at future null infinity I such as the Weyl scalar ψ4 or the GW strain obtained via extrapolation or Cauchy-characteristic extraction Establishing correspondence requires an explicit coordinate map in the bulk spacetime The network learns this coordinate map by matching the transformed numerical metric to a prescribed target metric without assuming a functional form for the gauge transformation

Neural Network Architecture and Training

The problem of finding an unknown smooth map between two coordinate systems is addressed using fully connected feedforward neural networks The network takes as input the four source coordinates xNR and returns the transformed coordinates x′target The network is trained by minimizing a loss that penalises the mismatch between the pulled-back metric and the prescribed target metric The architecture uses SiLU activation in hidden layers, ensuring that the network represents a smooth coordinate map whose Jacobian varies smoothly with the inputs

Results and Applications

The analytical tests show that the network recovers known coordinate transformations for analytical Schwarzschild and Kerr metrics up to isometries of the target spacetime In numerical cases, the learned map brings the time-dependent moving-puncture metric into ingoing Eddington–Finkelstein (EF) form for non-spinning BHs For spinning punctures, the method jointly fits the coordinate map and the mass and spin of a Kerr background in KS coordinates to construct initial data for Teukolsky evolution The resulting waveforms closely match those extracted from the full nonlinear simulation

Limitations and Future Directions

The residual mismatch contains physical deviations from Kerr together with numerical and fitting errors A limitation is that the present network architecture does not enforce invertibility explicitly, which could be done using a similar approach used in [77] Furthermore, the loss function could be better adapted to the different scales of the metric components for more balanced accuracy across the domain The perturbative application has additional sources of uncertainty, such as numerical noise from curvature extraction A natural extension is comparing simulations produced by different numerical relativity codes to make comparisons of local spacetime quantities more direct The method provides a measure of the remaining metric mismatch after gauge alignment from the NN training and help quantify deviations from the Kerr background

Acknowledgments

We would like to thank Abhay Ashtekar, Hengrui Zhu and Vaishak Prasad for helpful comments and discussions ES acknowledges support from The Charles E. Kaufman Foundation of The Pittsburgh Foundation SB acknowledges funding from the EU Horizon under ERC Consolidator Grant no. InspiReM-101043372 and from the Deutsche Forschungsgemeinschaft (DFG) project “GROOVHY” (BE 6301/5-1 Projektnummer: 523180871) This research used resources of the National Energy Research Scientific Computing Center (NERSC), a Department of Energy User Facility using NERSC award ERCAP0039050 Additional computations were performed on the Roar supercomputer at the Penn State Institute for Computational and Data Sciences (ICDS) Claude AI was used to help with coding and plotting scripts.

Appendix C: Spacetime Metrics in General Relativity

Schwarzschild Metric

a. Spherical Polar coordinates

g SC µν = (C1)

b. Ingoing Eddington-Finkelstein Coordinates

g EF µν = (C2)

Kerr Metric

a. Boyer–Lindquist Coordinates

g BL µν = (C3)

b. Ingoing-Kerr Coordinates

g IK µν = (C5)

Cartesian Kerr–Schild Metric

g KS µν = (C7)

Appendix A: 1+1D puncture

We attempted to map the 1+1D puncture data to ingoing EF coordinates with an incorrect target mass, M = 2

Appendix B: Learning mass and spin

To test whether this methodology is capable of recovering the correct mass and spin parameters, we repeat the analytical test (see Sec. IV A 2) of identifying coordinate transformation from BL coordinates to ingoing Kerr coordinates with M and a as learnable parameters

Appendix C: Spacetime Metrics in General Relativity

  1. Schwarzschild Metric

a. Spherical Polar coordinates

g SC µν = (C1)

  1. Kerr Metric

a. Boyer–Lindquist Coordinates

g BL µν = (C3)

c. Cartesian Kerr–Schild Metric

g KS µν = (C7)

c.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve cross-code validation by enabling point-wise comparison of bulk spacetime between codes using a learned coordinate map, which could provide a complementary diagnostic to asymptotic methods and help isolate where in the spacetime two codes begin to disagree.

  2. Enhance black hole perturbation theory by allowing the neural network to jointly fit the coordinate map and Kerr background parameters, enabling an approximate identification with a Kerr background despite departures from Kerr in the numerical spacetime and providing initial data for evolution with the Teukolsky equation.

  3. Develop a systematic alternative to existing coordinate identification methods by optimizing an explicit measure of agreement between metrics over the training domain while simultaneously fitting background mass and spin, providing a measure of the remaining metric mismatch after gauge alignment from the NN training.

  4. Increase robustness against coordinate singularities by explicitly building invertibility into the neural network layers, potentially using a similar approach to [77], to provide more robust control over the coordinate map.

  5. Improve loss function sensitivity by adapting it to different scales of metric components, moving beyond averaging squared absolute differences so that components with large coordinate-dependent magnitudes can dominate the optimization.

Sources

Related papers