astronomy/ai produced the result/RAS Techniques and Instruments 2025 · v2
Neural networks turn simulated cluster mass maps into X-ray and SZ images
Researchers trained U-Net neural networks on simulated galaxy clusters to convert maps of total mass into maps of the X-ray and microwave signals clusters give off, then applied them to dark-matter-only simulations.
spectrum · one line per step, placed by what the step does · bright lines used AI
Deep learning generated observations of galaxy clusters from dark-matter-only simulations
RAS Techniques and Instruments, 2025
doi:10.1093/rasti/rzaf007 · record aix-00082 v2 · checked 2026-10-08
- AI was for
- Simulation surrogate
- Model family
- Convolutional neural network, Clustering
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Galaxy clusters are the largest objects held together by gravity: hundreds or thousands of galaxies, plus hot gas, plus a much larger amount of dark matter. Most of a cluster's mass cannot be seen directly. What telescopes see instead is the hot gas between the galaxies. It glows in X-rays, and it leaves a faint imprint on the microwave background light passing through it, known as the Sunyaev-Zel'dovich effect and measured as a quantity called Compton-y. To compare theory with telescopes, astronomers run computer simulations of clusters. Simulations that follow the gas as well as the dark matter are far more costly than simulations that follow dark matter alone.
The team worked with The Three Hundred project, a set of cluster simulations that exist in both kinds: runs that include the gas and its physics, and runs with dark matter only. They set out to see whether a neural network could take a flat, projected map of a cluster's total mass and produce the matching X-ray and SZ maps, and whether such a network would still behave sensibly when fed mass maps from dark-matter-only runs, which have no gas at all.
Where AI came in
The AI here is a U-Net, a type of convolutional neural network built for turning one image into another image of the same shape. Six of these networks were trained from scratch, one for each observable signal across three choices of training simulation. Each took an 80 by 80 pixel map of projected total mass as input and learned to output the corresponding Compton-y or X-ray map, using the gas-physics simulations as the answer key. The network stands in for the expensive gas physics: rather than simulating how the gas heats, moves and radiates, it is asked to guess what that gas would have looked like given the mass alone.
Trained networks were then run on clusters held back from training, and on mass maps from two dark-matter-only simulations, to generate observable maps where none existed. The generated maps were scored against the simulated ones with image-similarity and difference measures, and through fits of the relations linking signal strength to cluster mass. Above a pivot mass of 2x10^14 h^-1 solar masses the percentage errors of the fitted relation's parameters averaged (0.5+-0.1)%, while lower-mass clusters departed further from the reference maps. Two further algorithms, t-SNE and UMAP, compressed the network's internal activations onto a flat plane to show which input simulations the network treated as alike.
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
U-Net convolutional networks were trained on The Three Hundred hydrodynamical cluster simulations to convert projected total mass maps into Compton-y (SZ) and bolometric X-ray surface brightness maps, and were then applied to total mass maps from dark-matter-only simulations. The generated maps were compared with ground-truth simulated maps using maximum mean discrepancy, mean relative difference and structural similarity, and through fits of the Y-M and LX-M scaling relations. For clusters above the pivot mass of 2x10^14 h^-1 M_sun, the percentage errors of the scaling-law parameters averaged (0.5+-0.1)%, while predictions for lower-mass clusters deviated more from the ground-truth datasets. Dimensionality reduction of the final convolutional layer with t-SNE and UMAP showed the Gadget_dm and Music_dm inputs lying close to Gadget_hy and the Gizmo_hy inputs separated from the rest.
How AI was used
Halos with M200 above 10^13.5 h^-1 M_sun were selected from The Three Hundred hydrodynamical runs (Gadget-X, Gizmo-Simba) and from Gadget-X DM-only and Music DM-only runs, and particle data were projected into 80x80 pixel total mass, SZ and X-ray maps after Gaussian smoothing and downsampling. Maps were normalised by a log10 transform followed by standardisation, chosen after testing five normalisation schemes. A U-Net with an encoder of 32-256 filters, a 512-filter bottleneck with dropout, and a transposed-convolution decoder, with about 7.8x10^6 trainable parameters, was trained with the ADAM optimizer and a mean absolute error loss for 100 epochs, with the learning rate halved after five epochs without improvement, using an 80/20 train-test split. Six such models were trained, one per observable for each of the Gizmo+Gadget, Gizmo and Gadget training sets. The trained networks were then run on held-out hydrodynamical mass maps and on mass maps from the two dark-matter-only simulations to produce observable maps, which were compared to ground truth with MMD, MRD and SSIM and through least-squares broken power-law fits of the observable-mass scaling relations. Activations of the final convolutional layer were embedded in two dimensions with t-SNE and UMAP to inspect how the network separated the input simulations.
The shape of the work
Structural · the record, drawn
no AI
Select galaxy cluster halos from simulations
Obtaining raw data, whether by measurement, download or retrieval.
the criteria include selecting halos with a mass M200>1013.5h−1M⊙ at a redshift close to zerowhere the paper describes this · verbatim
no AI
Project and preprocess mass, SZ and X-ray maps
Cleaning, filtering, normalising or labelling data already obtained.
The SZ maps are computed by using the publicly available library PYMSZ, that integrates the pressure field along the line of sightwhere the paper describes this · verbatim
AI
Train U-Net models to map mass to observables
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
a total of six individual U-Net models (Table 3) were trained to predict observable maps in galaxy clusterswhere the paper describes this · verbatim
AI
Predict observable maps for held-out test clusters
Running a trained model over new data to predict, classify or score. The AI stood in for simulation.
We present the results obtained using the test dataset, which comprises 503 clusters from the Gadget simulationwhere the paper describes this · verbatim
AI
Predict observable maps from dark-matter-only simulations
Running a trained model over new data to predict, classify or score. The AI stood in for simulation.
we further assess its performance using the Gadget-X DM-only from the The300 projectwhere the paper describes this · verbatim
no AI
Score maps with metrics and fit scaling relations
Testing outputs against ground truth.
The metrics considered are the Maximum Mean Discrepancy (MMD), the Mean Relative Difference (MRD), and Structural Similarity Index Measure (SSIM)where the paper describes this · verbatim
AI
Embed final convolutional layer for interpretation
Extracting understanding from model behaviour.
we applied a couple of dimensionality reduction algorithms and visualized the results in a two-dimensional planewhere 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 paper's object of study is the observable maps produced by the U-Net models; every reported result is a property of those generated maps
we specifically utilize U-Net convolutional neural networks, a widely used deep learning architecturewhere the paper describes this · verbatim
For data splitting, the training dataset comprises 80% of the samples, while 20% is allocated to the testingwhere the paper describes this · verbatim
The data is freely available upon request following the guidelines of The Three Hundred collaborationwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of U-NetWhich version of the model was used is not stated.
- Version of t-SNE (Scikit-Learn)Which version of the model was used is not stated.
- Version of UMAP (UMAP-learn)Which version of the model was used is not stated.
- What step 7 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00082, 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