From Observations to Simulations: A Neural-Network Approach to Intracluster Medium Kinematics

TL;DR

The paper tackles the challenge of constraining intracluster medium (ICM) kinematics by linking XMM-Newton LOS velocity maps of four nearby clusters to synthetic maps from IllustrisTNG-300. It introduces a Siamese CNN that learns a data-driven similarity metric in a learned embedding space, enabling orientation-invariant matching between observations and simulations. By generating $5016$ synthetic velocity maps across $101$ viewing angles and using a triplet-loss objective, the approach identifies best-matching TNG300 halos whose velocity structures reproduce both large-scale gradients and localized substructures, consistent with gas sloshing, AGN feedback, and minor mergers. The results demonstrate a robust, scalable framework for connecting high-resolution X-ray kinematic data to cosmological simulations, with implications for understanding turbulence, feedback, and cluster evolution, and set the stage for future tests with XRISM and Athena.

Abstract

We present a systematic comparison between {\it XMM-Newton} velocity maps of the Virgo, Centaurus, Ophiuchus and A3266 clusters and synthetic velocity maps generated from the Illustris TNG-300 simulations. Our goal is to constrain the physical conditions and dynamical states of the intracluster medium (ICM) through a data-driven approach. We employ a Siamese Convolutional Neural Network (CNN) designed to identify the most analogous simulated cluster to each observed system based on the morphology of their line-of-sight velocity maps. The model learns a high-dimensional similarity metric between observations and simulations, allowing us to capture subtle kinematic and structural patterns beyond traditional statistical tests. We find that the best-matching simulated halos reproduce the observed large-scale velocity gradients and local kinematic substructures, suggesting that the ICM motions in these clusters arise from a combination of gas sloshing, AGN feedback, and minor merger activity. Our results demonstrate that deep learning provides a powerful and objective framework for connecting X-ray observations to cosmological simulations, offering new insights into the dynamical evolution of galaxy clusters and the mechanisms driving turbulence and bulk flows in the hot ICM.

Figures (5)

• Figure 1: Diagram of the CNN model used in our analysis. The CNN acts as a feature extractor, converting each input velocity map into a compact embedding vector. Convolution and max-pooling layers progressively identify hierarchical spatial and kinematic features; the output is then flattened and passed through fully connected layers to produce the final embedding.
• Figure 2: Diagram of the Siamese CNN training phase. During training, pairs of velocity maps are passed through two identical CNN encoders with shared weights. Each triplet consists of a reference (Anchor), a matched projection from the same simulation (Positive), and a mismatched projection from a different simulation (Negative). The network learns to produce similar embeddings for Anchor--Positive pairs and dissimilar embeddings for Anchor--Negative pairs by minimizing the triplet loss. This process teaches the model to recognize the underlying kinematic structure of clusters, independent of projection and resolution differences.
• Figure 3: Diagram of the Siamese CNN matching phase. During matching, the trained Siamese CNN compares each observed XMM-Newton velocity map to the full library of simulated velocity maps. Both the observed map and each simulation are passed through the shared-weight CNN encoder to generate corresponding embedding vectors. Similarity between the observation and a given simulation is quantified by the distance between their embeddings in the learned latent space. Simulations with the smallest embedding distance are identified as the closest kinematic matches, enabling a quantitative, model-driven comparison between observed cluster dynamics and the cosmological simulations.
• Figure 4: t-SNE plots comparing the observed velocity maps of four galaxy clusters to a library of TNG simulations. Points represent the similarity of velocity maps, with closer points being more similar. The distribution of TNG simulations is non-uniform, exhibiting curves and filaments corresponding to continuous evolutionary sequences and underlying physical gradients (e.g., merger activity) learned by the network. The XMM-Newton velocity maps are shown in red, and the best-match simulations based on Euclidean distance are highlighted in blue. The axes in these t-SNE plots are arbitrary and do not correspond to physical quantities. The quantitative similarity ranking is based on the Euclidean distances in the original embedding space.
• Figure 5: Best-matching velocity maps from the TNG300 simulation obtained with the Siamese CNN analysis for the XMM-Newton observations. The zoom-in panel shows the XMM-Newton data for a region centered on the physical origin (0,0 kpc) of the simulated cluster. These figures illustrate kinematic similarity based on the CNN-learned metric, not pixel-wise visual resemblance.