🔭 astrophysics

Standard Reconstruction Shifts the Optimal Input Scale for CNN-Based Density-Field Reconstruction

This study demonstrates that applying standard first-order reconstruction before CNN-based density-field reconstruction shifts the optimal input scale to smaller physical sizes ( $\sim38$ – $114~h^{-1}\mathrm{Mpc}$ ), revealing a synergistic approach where perturbative methods effectively recover large-scale displacements while neural networks excel at modeling the remaining quasi-linear and non-linear evolution on smaller scales.

Original authors: Koichiro Nakashima, Kiyotomo Ichiki, Atsushi J. Nishizawa

Published 2026-07-20

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Koichiro Nakashima, Kiyotomo Ichiki, Atsushi J. Nishizawa

Original paper licensed under CC BY 4.0 (http://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

Imagine the universe as a giant, invisible ocean of dark matter. Billions of years ago, this ocean was almost perfectly smooth, like a calm lake on a windless day. But over time, gravity acted like a relentless wind, creating waves, whirlpools, and massive currents that pulled matter together into the cosmic web of galaxies we see today. Astronomers are obsessed with figuring out what that calm, early "lake" looked like. Why? Because the secrets of the universe's birth, its expansion, and the mysterious forces driving it are hidden in those initial ripples. The problem is, looking at the universe today is like trying to guess the shape of a calm pond by staring at a churning, stormy sea; the waves have mixed everything up, making it incredibly hard to work backward to the beginning.

To solve this puzzle, scientists use two main tools. The first is "standard reconstruction," a mathematical technique based on known physics that tries to reverse the biggest, most obvious currents (the bulk flows) to smooth the sea back out a bit. The second is a "Convolutional Neural Network" (CNN), a type of artificial intelligence that acts like a super-smart detective, looking at local patterns to guess what the original picture looked like. For a while, researchers thought the AI needed to see a huge, wide view of the universe to do its job well, hoping that a bigger picture would help it understand the big waves.

However, a new study by Koichiro Nakashima and colleagues suggests that this intuition might be backwards when you use the mathematical tool first. They ran massive computer simulations of the universe to test how these tools work together. They found that if you let the AI look at the raw, messy universe, it does best when it sees a large area, about 150 to 200 units of distance (specifically $h^{-1}$ Mpc). But, if you first use the mathematical tool to smooth out the big waves, the AI's sweet spot changes dramatically. Suddenly, it works best when it looks at a much smaller, more detailed patch of the universe, roughly 38 to 114 units away.

The paper reveals that the two methods are actually a perfect team, but they need to split the work. The mathematical tool is the master of the big picture; it efficiently handles the massive, long-distance movements of matter that the AI struggles to see from a small window. Once the math tool has done its job of reversing those giant flows, the AI doesn't need to look at the whole ocean anymore. Instead, it can zoom in on a smaller, high-resolution area to fix the remaining, trickier details—the smaller waves and the complex swirls that the math tool missed. The study shows that combining them this way creates a much clearer picture of the early universe than using either method alone, proving that sometimes, to see the whole forest, you first need to clear the biggest trees, and then let the detective look closely at the leaves.

Technical Summary: Standard Reconstruction Shifts the Optimal Input Scale for CNN-Based Density-Field Reconstruction

Problem Statement
Reconstructing the high-redshift initial density field from late-time large-scale structure (LSS) is essential for improving constraints on early-Universe physics, such as primordial non-Gaussianity, and for sharpening baryon acoustic oscillation (BAO) measurements. While perturbative methods like the Zel'dovich approximation (standard reconstruction) effectively reverse large-scale bulk flows, they struggle with non-linear evolution and shell crossing. Conversely, Convolutional Neural Networks (CNNs) have shown promise in learning non-linear mappings but face a fundamental tension: gravitational evolution is non-local, yet CNNs possess limited receptive fields. Previous work (e.g., Shallue & Eisenstein 2023; Nakashima et al. 2025) suggested that hybrid approaches—combining standard reconstruction with CNNs—could mitigate this, but the specific dependence of CNN performance on input physical scale after standard reconstruction remained unquantified.

Methodology
The authors utilize dark-matter-only $N$ -body simulations (Indra suite, $L_{box} = 1 \, h^{-1}$ Gpc, WMAP7 cosmology) to reconstruct the dark matter density field at redshift $z=10$ from the evolved field at $z=0$ . They compare three distinct reconstruction strategies:

CNN only: A single-input CNN applied directly to the $z=0$ density field.
Dual-input CNN: A network combining two physical scales (small and large) without prior standard reconstruction.
Stdrec.+CNN: A hybrid approach where standard first-order reconstruction (using the Lagrangian Growth-Shift scheme in nbodykit) is applied to the $z=0$ field, and the resulting reconstructed field is fed into a single-input CNN.

A critical experimental variable is the physical side length of the CNN input sub-box ( $L_{sub}$ ), which is varied over the range $38 \text{--} 380 \, h^{-1}$ Mpc while keeping the numerical resolution fixed at $39^3$ voxels. This setup isolates the trade-off between spatial context (physical volume) and spatial resolution (cell size). The network architecture follows Mao et al. (2021), predicting the central voxel density at $z=10$ from the input sub-box.

Key Results
The study reveals a systematic shift in the optimal input scale for the CNN depending on whether standard reconstruction is applied:

Optimal Scale Shift: For the "CNN only" approach, reconstruction accuracy (measured by normalized loss, KL divergence, and Fourier correlation) peaks at large input scales of $L_{sub} \sim 150\text{--}200 \, h^{-1}$ Mpc. However, for the "stdrec.+CNN" approach, the optimal scale shifts significantly to smaller volumes, $L_{sub} \sim 38\text{--}114 \, h^{-1}$ Mpc.
Performance Superiority: The "stdrec.+CNN" model consistently outperforms both the single-input and dual-input CNNs applied without standard reconstruction. Specifically, at $L_{sub} \sim 76 \, h^{-1}$ Mpc, the hybrid model achieves a validation loss approximately 80% lower than the "CNN only" model and 50% lower than the dual-input model.
Redundancy of Dual-Input after Reconstruction: Applying the dual-input architecture after standard reconstruction yields no measurable improvement over the single-input hybrid model. This suggests that once standard reconstruction has corrected the dominant large-scale displacements, the additional large-scale input branch of the dual-input CNN becomes redundant.
Statistical Metrics: The hybrid model produces the tightest cell-by-cell correlation, the most homogeneous residual fields (indicating the removal of coherent large-scale errors), and the closest match to the target one-point density probability distribution function (PDF).

Significance and Claims
The paper argues that these findings support a physically motivated separation of scales between analytic and data-driven reconstruction methods:

Division of Labor: Standard reconstruction (based on the Zel'dovich approximation) is more efficient at recovering coherent large-scale displacements driven by long-wavelength modes ( $k \sim 0.01 \, h$ Mpc $^{-1}$ ). Once these bulk flows are reversed, the CNN is better suited to modeling the remaining quasi-linear and non-linear evolution on smaller scales.
Validation of Hybrid Strategies: The results provide an independent explanation for the success of hybrid frameworks (e.g., Shallue & Eisenstein 2023). The effective receptive scale of $\sim 60 \, h^{-1}$ Mpc used in previous hybrid studies falls squarely within the newly identified optimal range ( $38\text{--}114 \, h^{-1}$ Mpc) for post-reconstruction CNN inputs.
Efficiency: The study demonstrates that combining analytic reconstruction with machine learning allows each method to address the scales for which it is best suited, rather than relying solely on increasing CNN receptive fields or network complexity.

The authors note that these results apply to idealized real-space dark matter fields and that future work must address redshift-space distortions, tracer bias, and observational effects before application to galaxy surveys.

More like this