🔭 astrophysics

A Parallel Imaging Simulation Framework for Lunar-based Radio Telescope Arrays

This paper proposes a parallel processing framework utilizing multi-CPU and GPU clusters to enhance the efficiency and accuracy of simulating large-scale lunar-based radio telescope arrays, specifically supporting the Hongmeng Plan mission through optimized baseline simulation, observation modeling, and image reconstruction.

Original authors: Xiang Zhao, Ran Duan, Li Deng, You Song

Published 2026-08-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xiang Zhao, Ran Duan, Li Deng, You Song

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

For decades, astronomers have been blind to a vast, silent corner of the universe. Below a frequency of 30 million cycles per second, the Earth's own atmosphere acts like a heavy curtain, bending and absorbing the faint radio whispers from the cosmos. This hidden band holds the keys to understanding the "Cosmic Dark Ages," the mysterious era before the first stars ignited, and the subsequent "Cosmic Dawn" when those stars began to light up the universe. To see this, scientists must look from space, far above the interference of our planet. One ambitious proposal, known as the Hongmeng Plan, envisions a fleet of satellites orbiting the Moon to create a giant, floating radio telescope. However, the sheer volume of data such a mission would generate is staggering, requiring a new way to process information that current computers simply cannot handle alone.

To solve this, researchers Xiang Zhao, Ran Duan, Li Deng, and You Song have developed a powerful new software framework designed to simulate how this lunar telescope would work. Their work focuses on the immense computational challenge of turning raw radio signals into clear images of the entire sky. In a traditional radio telescope, signals are collected by pairs of antennas, and the distance between them—called a baseline—determines the detail of the picture. The Hongmeng Plan uses eight small satellites that move in a coordinated dance, constantly changing their distances from one another to build a three-dimensional map of the sky. Because these satellites are constantly moving and the Moon blocks out Earth's radio noise only for part of their orbit, the data they collect is incredibly complex and unevenly distributed.

The team created a simulation that mimics the entire process of this mission, from calculating the positions of the satellites to reconstructing the final image. They found that trying to do this on a single computer would be impossibly slow. Instead, they built a system that splits the work across multiple processors, using both standard computer chips and specialized graphics cards to handle different parts of the job. For the heavy lifting of calculating satellite positions and counting how many data points fall into specific areas of the sky, they used a multi-core computer processor. For the intense mathematical work of turning those data points into an image, they harnessed the power of graphics cards, which are designed to perform millions of calculations simultaneously. This hybrid approach allowed them to process the massive amounts of data generated by the mission in a fraction of the time it would otherwise take.

A major hurdle in this process is that the satellites orbit at a specific angle, which means they cannot see every part of the sky equally well. This creates gaps in the data, leading to images that look brighter in some areas and dimmer in others, regardless of what is actually out there. The researchers discovered that this unevenness was not a flaw in the telescope's design but a natural consequence of the satellites' path. To fix this, they developed a correction method that analyzes the distribution of the data points and adjusts the brightness of the final image accordingly. By dividing the sky into different zones and applying specific mathematical adjustments to each, they were able to smooth out these artificial bright spots and reveal a more accurate picture of the universe.

The results of their simulation were striking. When they tested the system at a frequency of 10 million cycles per second, the framework successfully reconstructed a full-sky image in a matter of hours, a task that would have taken days or weeks with older methods. The corrected images showed a high level of detail, preserving the shapes and structures of cosmic radio sources while removing the distortions caused by the satellite formation. The team measured the quality of these images using standard tests for accuracy and found that the corrected versions were significantly closer to the true sky than the uncorrected ones. This success suggests that the proposed framework is not just a theoretical exercise but a practical tool that could be used to guide the actual Hongmeng mission when it launches.

While the simulation is a major step forward, the researchers acknowledge that it is not yet perfect. The current system has not yet been tested at the highest frequencies the mission might reach, and it does not yet account for the Moon itself blocking parts of the view during certain times. However, the work proves that with the right combination of computing power and clever algorithms, the massive data challenges of lunar radio astronomy can be overcome. By providing a clear path to processing these complex signals, this framework lays the technical foundation for a future where we can finally lift the curtain on the universe's earliest moments.

Technical Summary: A Parallel Imaging Simulation Framework for Lunar-based Radio Telescope Arrays

1. Problem Statement

Ground-based radio observations below 30 MHz are severely constrained by ionospheric refraction and radio frequency interference (RFI), creating a bottleneck for studying the Cosmic Dark Ages and the Epoch of Reionization via redshifted neutral hydrogen 21-cm radiation. To overcome this, the "Hongmeng Plan" proposes deploying a linear satellite array in lunar orbit to conduct full-sky imaging.

However, this mission faces significant computational challenges:

Massive Data Volume: The mission design involves 8 sub-satellites forming baselines that undergo "spring-like" telescopic transformations over a 3-year mission. At 30 MHz, the total number of baselines is projected to exceed 3 billion.
Complex Imaging Geometry: The array utilizes omnidirectional antennas with non-coplanar baselines and a wide field of view. Traditional 2D Fourier transform approximations are invalid; the $w$ -term (non-coplanar baseline effects) must be accounted for using 3D Fourier transforms.
Imaging Artifacts: Due to the 30° inclination of the lunar orbit, the baseline distribution in the 3D frequency domain is incomplete, leaving conical gaps (60° half-cone angle) along the polar axis. This results in latitude-dependent non-uniformities in reconstructed brightness temperature images, specifically elevated temperatures within ±30° latitude.
Computational Bottleneck: Processing petabyte-scale data for such missions requires exascale capabilities. Existing frameworks designed for ground-based arrays are not directly applicable to the unique orbital dynamics and data characteristics of the Hongmeng Plan.

2. Methodology

The authors propose a heterogeneous parallel computing framework integrating multi-CPU and multi-GPU architectures to handle the entire imaging pipeline: baseline simulation, baseline counts statistics, observation simulation, and image reconstruction.

2.1 Parallel Workflow Architecture

The framework is divided into four core stages, optimized for specific hardware capabilities:

Baseline Simulation (Multi-GPU):
- Satellite positions are generated per telescopic transformation cycle (14 days).
- Relative positions between 8 sub-satellites are computed to form baseline vectors.
- Parallelization: Multiple GPUs process different cycles in parallel. Within a single GPU, tasks include generating satellite positions, calculating valid baselines, and performing sorting and deduplication using the Thrust library (CUDA toolkit) to manage global baseline arrays efficiently.
Baseline Counts Statistics (Multi-CPU):
- This stage calculates weighting coefficients ( $\Delta c_k$ ) based on the density of baselines in 3D frequency grids (divided by 0.5 wavelength).
- Hardware Choice: Despite the high parallelism of GPUs, the authors selected a multi-CPU approach using Python (Pandas) and multi-threading.
- Reasoning: Due to GPU memory constraints, splitting the massive dataset (billions of baselines) into GPU blocks ( $M$ ) would require excessive data transfer time compared to the CPU approach ( $N$ blocks). The CPU framework processes data in larger segments with fewer iterations, proving more efficient for this specific I/O and aggregation task.
Observation Simulation (Multi-GPU):
- Simulates visibility functions ( $V$ ) based on a full-sky brightness temperature map ( $I$ ) and celestial coordinates ( $l, m, n$ ).
- Optimization: Uses MPI and OpenMP for parallel control across GPUs.
- Kernel Implementation: Celestial coordinates and sky brightness data are stored in GPU shared memory to minimize latency. The baseline simulation kernel executes the Fourier transform (Eq. 2) in parallel, utilizing loop unrolling to accelerate computation.
Image Reconstruction (Multi-GPU):
- Reconstructs the full-sky image via inverse Fourier transform (Eq. 5).
- Optimization: Baseline data, visibilities, and counts are distributed across GPUs. Reusable data (baselines, visibilities, coordinates) are preloaded into shared memory. The inverse transform kernel is executed in parallel with loop unrolling.

2.2 Image Correction Strategy

To address the latitude-dependent brightness temperature artifacts caused by incomplete baseline coverage:

Region Segmentation: The sky is divided into a Central Region ( $C$ , $\approx \pm 30^\circ$ ), Transition Region ( $T$ ), and High-Latitude Region ( $H$ ) based on baseline elevation density.
Correction Coefficients:
- Central Region: A correction coefficient is calculated as the ratio of average baselines in non-central regions to those in the central region.
- Transition Region: Coefficients are derived from the ratio of mean brightness temperatures between high-latitude and transition regions.
Adaptive Calibration: A pipeline applies a large-scale median filter to isolate background components, computes a latitudinal gain sequence, and applies Gaussian smoothing and a damping factor ( $\alpha=0.5$ ) to generate a pixel-wise gain field. This field is used to differentially compensate background and structural components, followed by global intensity shifting and outlier clipping.

3. Key Contributions

Heterogeneous Framework: Development of a scalable simulation framework that strategically assigns tasks to CPUs (for massive data aggregation/statistics) and GPUs (for compute-intensive Fourier transforms and baseline generation).
3D Full-Sky Imaging: Adaptation of 3D Fourier transform methods specifically for the non-coplanar, wide-field geometry of a lunar-orbiting interferometer, moving beyond ground-based 2D approximations.
Baseline-Driven Correction: Proposal of a novel image brightness temperature correction method that utilizes the specific distribution of baselines to mitigate artifacts caused by orbital inclination, effectively suppressing latitude-dependent non-uniformities.
Performance Optimization: Implementation of specific CUDA optimizations, including shared memory usage for reusable data, loop unrolling, and the use of the Thrust library for efficient sorting and deduplication.

4. Results

Experiments were conducted using a cluster with 2 Intel Xeon Gold 8336C CPUs and 3 NVIDIA GeForce RTX 4090 GPUs.

Efficiency:
- At 1 MHz, the full pipeline (observation simulation + image reconstruction) completed in approximately 52 minutes (25 min + 27 min).
- At 10 MHz, the framework successfully reconstructed images in 344 hours (167 h + 177 h) using 3 GPUs. This demonstrates the feasibility of processing high-resolution data (Nside=4096, ~200M pixels) which would be intractable with serial methods.
Image Quality:
- Quantitative Metrics: The image correction method significantly improved fidelity.
  - SSIM: Increased from 0.7847 to 0.9369 at 1 MHz and from 0.8195 to 0.9290 at 10 MHz.
  - PSNR: Improved from ~20.2 dB to 29.5 dB (1 MHz) and ~20.9 dB to 28.1 dB (10 MHz).
  - MSE: Reduced by nearly an order of magnitude.
- Visual Analysis: The correction effectively suppressed the elevated brightness temperatures in the central region ( $\pm 30^\circ$ ). While faint residual artifacts remain near the boundaries of the central and transition regions, the overall morphological fidelity was restored.

5. Significance and Claims

The paper claims that this work provides a critical technical foundation for the Hongmeng Plan mission. By demonstrating a scalable parallel framework capable of handling the massive data volumes and complex geometric constraints of a lunar-based array, the authors establish a pathway for:

Real-time and High-Precision Processing: Enabling the processing of petabyte-scale data required for future exascale radio astronomy projects.
Scientific Feasibility: Validating that high-resolution, full-sky imaging of the Cosmic Dark Ages is computationally achievable using current heterogeneous hardware, provided that specific corrections for orbital geometry are applied.
Methodological Reference: Offering a reference architecture for future lunar interferometric missions, particularly in handling non-coplanar baselines and large-scale data statistics.

The authors acknowledge limitations, noting that the current simulation does not yet incorporate lunar occultation effects and that full-precision simulation at 30 MHz is currently constrained by GPU memory, requiring future work in distributed computing and kernel customization.