Segmenting Superbubbles in a Simulated Multiphase Interstellar Medium using Computer Vision
Authors: Jing-Wen Chen, Alex S. Hill, Anna Ordog, Rebecca A. Booth, Mohamed S. Shehata
Organizations: Department of Computer Science, Math, Physics & Statistics, University of British Columbia, Okanagan Campus 1177 Research Rd, Kelowna, BC V1V 1V7, Canada · Dominion Radio Astrophysical Observatory, Herzberg Research Centre for Astronomy and Astrophysics, National Research Council Canada, PO Box 248, Penticton, BC V2A 6J9, Canada · Department of Physics & Astronomy, University of Western Ontario, 1151 Richmond Street, London, ON, N6A 3K7, Canada · Department of Physics and Astronomy, University of Calgary, 2500 University Drive NW, Calgary, Alberta, T2N 1N4, Canada
We developed a computer vision-based methodology to achieve precise 3D segmentation and tracking of superbubbles within magnetohydrodynamic simulations of the supernova-driven interstellar medium. Leveraging advanced 3D transformer models, our approach effectively captures the complex morphology and dynamic evolution of these astrophysical structures. To demonstrate the technique, we specifically focused on a superbubble exhibiting interesting interactions with its surrounding medium, driven by a series of successive supernova explosions. Our model successfully generated detailed 3D segmentation masks, enabling us to visualize and analyze the bubble's structural evolution over time. The results reveal insights into the superbubble's growth patterns, energy retention, and interactions with surrounding interstellar matter. This interdisciplinary approach not only enhances our understanding of superbubble dynamics but also offers a robust framework for investigating other complex phenomena in the cosmos.
Figures & tables
Figure 1: Segmentation results demonstrating superbubbles isolated from the input density ( n ) cube using our proposed tracking framework. The upper row features a supernova remnant case at t=191.8 Myr, while the lower row shows a superbubble identified at t=213.0 Myr in the same simulation.
Figure 2: Density ( n ), temperature ( T ), and vertical velocity ( vz ) from the original MHD simulation dataset [ 32 ] , captured at t=191.9 Myr. The leftmost column represents z=−256 pc, the second column z=0 (the Galactic plane), and the rightmost column z=+256 pc.
Figure 3: Visualization presenting 3D density cubes of hot gas extracted from the magnetohydrodynamic simulation dataset [ 31 ] . The cubes are thresholded at T>105.5 K to highlight hot gas structures at t=209 Myr. The rotation around the z -axis provides different perspectives, offering a clearer view of the spatial distribution and morphology of these high-temperature regions.
Figure 4: Model architecture of the proposed Astro-UNETR model. The 3D data cubes are input to the Astro-UNETR, which employs swin-transformer blocks [ 45 ] to learn the 3D semantic representation of superbubble morphology. High-level features are refined by a bottleneck layer and then upsampled to their original dimensions via deconvolution layers and ResNet blocks, producing a 3D semantic segmentation of all bubbles.
Figure 5: SAM2 model architecture illustration. To isolate a specific bubble, the semantic segmentation output from Astro-UNETR is fed into SAM2 [ 57 ] , a video object segmentation model that identifies the bubble corresponding to the superbubble designated by the supernova (SN) log, which yields an isolated 3D instance segmentation of the target bubble. Subsequent data cubes of the same bubble are segmented over time using the stored model features.
Figure 6: A comparison between SB230 and the larger central structure observed in: (a) the 3D simulation cube, and (b) a 2D x - y plane density slice at z=202 pc. (c) SB230 is seen in 3D as a distinct, well-defined superbubble, making it an ideal candidate for tracking its evolution. (d) The structure below SB230, though appearing bubble-like in 2D, is revealed in 3D to be part of an extensive interconnected system forming a tunnel-like structure linking the upper and lower Galactic halo. Such a feature does not meet the criteria for an isolated superbubble driven by localized supernova explosions. The central structure largely represents the diffuse, interconnected hot gas within the ISM, whereas the isolated bubble on the right more accurately corresponds to a discrete superbubble formed by localized supernova activity.
Figure 7: 3D density plots visualizing the evolution of the tracked superbubble SB230 (opaque) from its initial explosion through 30 Myr of its lifetime. The first frame in this illustration corresponds to the same time slice ( t=209 Myr) depicted from various viewing angles in Figure 3 . The bubble transitions from an initial spherical shape to a more complex structure, developing multiple tunnels that facilitate matter outflow into neighboring regions and the upper halo. This evolution highlights the capability of our computer vision-based model in accurately tracking the full superbubble evolution trajectory.
Figure 8: Illustration of the evolution of volume and energy based on the segmented superbubble morphology. The volume undergoes a rapid surge within the first 5 Myr, followed by periodic increases after supernova-like energy injections (red crosses), and decreases as matter flows into adjacent structures or escapes into the halo through formed tunnels. The energy remains consistently around 1044 J throughout the bubble’s lifetime, indicating its ability to retain energy while maintaining a discrete structure.
Figure 9: Density slice at z=167 pc for SB230 (masked in orange) at each timestep. This plot illustrates the evolution of the target bubble, starting as a distinct structure, temporarily connecting with a nearby formation at t=213 Myr, and later detaching again at t=219 Myr. This progression highlights the challenges of accurately tracing bubble morphology when relying solely on 2D observations.
Figure 10: Visualization of the vertical feature of density ( n ) and vz for superbubble SB230. The left column shows a 2D density slice at z=202 pc with a red line marking the chosen vertical cut to intersect the identified bubble. The middle column presents the cross-sectional density slices along that cut, which pass through SB230, and the right column shows the corresponding vertical vz slices that capture the bubble’s dynamical evolution. We compare these slices with direct views of the isolated bubble to validate the algorithm’s selections. Applying this automated procedure across many bubbles enables robust statistical extraction and demonstrates that detection-based analysis reproduces the physical behavior obtained by manual inspection, supporting its use for the full sample.
Figure 11: Density slice at z=167pc for SB230 (masked in orange), shown at 2 pc resolution for each timestep. The 211−213 Myr timesteps are also shown in the first row of Figure 9 at 4 pc resolution. This figure demonstrates the model’s robustness when applied directly to inputs of different resolutions.
Figure 12: Histogram of cell volumes binned by temperature for the bubble shell (black step lines) and bubble interior (gray step lines), used to compare their temperature distributions. When the bubble is distinctly separated from its surroundings ( 209 Myr; top panel), the gas in the surrounding area remains in a multi-phase state, with a notable peak in warm gas. As the chimney forms and connects with nearby entities (bottom panel), the exterior ring begins to encompass portions of the hot-hot interface of the chimney. This results in a significant increase in the distribution of hot gas within the exterior of the bubble. Concurrently, the interior of the bubble predominantly consists of hot gas, reflecting the intense thermal activities occurring within it.
Figure 13: Illustration of the temperature composition within the bubble shell ring across various timesteps. Initially, the bubble is surrounded by cool and warm gas. As it expands and heats up, cool gas diminishes while hot gas forms chimney-like structures that eventually wane as matter is expelled into the upper halo. In the final stage (Figure 7 ), the bubble develops a full cap with increased surrounding hot gas, highlighting enhanced connectivity with the interstellar medium.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 14: Visualization of vertical slices of SB230 superbubble at its mid-life stage, spanning from 216 to 224 Myr, following Figure 10 . The plots show the formation of tunnels triggered by a supernova injection and their subsequent disintegration over the next few Myr.
Figure 15: Final stage in the life of the SB230 superbubble from 226 to 234 Myr, showing how the bubble’s tunnel progressively expands towards the upper halo and eventually opens up the northern cap. In the vz channel, significant upward matter flow is visible, supporting this expansion. This visualization provides deeper insight into the movement of matter within the bubble that influences its morphological changes.
Bubbly flows involve strongly coupled bubble motion, deformation, and hydrodynamic interactions, making high-fidelity simulation computationally expensive. We propose a neural surrogate that jointly predicts bubble trajectories and interface deformation in a compact, shared representation, replacing costly mesh-resolved simulation with fast generative inference. Because two-phase flow dynamics is nonlinear and chaotic, we treat forecasting as a conditional generative problem rather than deterministic regression, predicting entire future windows in a single pass. We train and evaluate the surrogate on BubbleSH, a Front-Tracking dataset of rising deformable bubble swarms, assessing performance using trajectory, shape, and distributional metrics over kinematic, morphological, and interaction-based properties. The surrogate reproduces key flow statistics while running several orders of magnitude faster than the underlying DNS solver.
Kiet Bennema ten Brinke, Rachna Ramesh, Douwe Orij +2
Department of Mathematics and Computer Science, Eindhoven University of Technology · Department of Chemical Engineering and Chemistry, Eindhoven University of Technology
Superpixels offer a compact image representation by grouping pixels into coherent regions. Recent methods have reached a plateau in terms of segmentation accuracy by generating noisy superpixel shapes. Moreover, most existing approaches produce a single fixed-scale partition that limits their use in vision pipelines that would benefit multi-scale representations. In this work, we introduce H-SPAM (Hierarchical Superpixel Anything Model), a unified framework for generating accurate, regular, and perfectly nested hierarchical superpixels. Starting from a fine partition, guided by deep features and external object priors, H-SPAM constructs the hierarchy through a two-phase region merging process that first preserves object consistency and then allows controlled inter-object grouping. The hierarchy can also be modulated using visual attention maps or user input to preserve important regions longer in the hierarchy. Experiments on standard benchmarks show that H-SPAM strongly outperforms existing hierarchical methods in both accuracy and regularity, while performing on par with most recent state-of-the-art non-hierarchical methods. Code and pretrained models are available: https://github.com/waldo-j/hspam.
Julien Walther, Rémi Giraud, Michaël Clément
Univ. Bordeaux, CNRS, Bordeaux INP, IMS, UMR 5218, France · Univ. Bordeaux, CNRS, Bordeaux INP, LaBRI, UMR 5800, France
Simultaneous perception of 2D objects in perspective view and 3D objects in Bird's Eye View (BEV) is challenging for multi-camera autonomous driving. Existing two-stage pipelines use 2D results only as a one-time cue for 3D detection. We propose SimPB++, which simultaneously detects 2D objects in perspective and 3D objects in BEV from multiple cameras. It unifies both tasks into an end-to-end model with a hybrid decoder architecture, coupling multi-view 2D and 3D decoders interactively. Two novel modules enable deep interaction: Dynamic Query Allocation adaptively assigns 2D queries to 3D candidates, and Adaptive Query Aggregation refines 3D representations using multi-view 2D features, forming a cyclic 3D-2D-3D refinement. For multi-view 2D detection, we use Query-group Attention for intra-group communication. We also design a Crop-and-Scale strategy for long-range perception and a Propagating Denoising strategy with an auxiliary RoI detector. SimPB++ supports mixed supervision with 2D-only and fully annotated data, reducing reliance on expensive 3D labels. Experiments show state-of-the-art performance on nuScenes for both tasks and strong long-range detection (up to 150m) on Argoverse2.