Equation vs. AI: Predict Density and Measure Width of molecular clouds by Multiscale Decomposition

TL;DR

The paper tackles the challenge of inferring 3D density structure from 2D column density by introducing Multi-scale Decomposition Reconstruction (MDR), a physics-based method that extracts a local characteristic width $l_c$ via Constrained Diffusion and, under a statistical isotropy assumption, uses $l_t \approx l_c$ to estimate the volume density as $\langle n(H)\rangle = \Sigma / l_t$. MDR is validated against FLASH and Enzo MHD simulations and applied to real observations (Orion A, NGC 1333, Cygnus-X), achieving ~0.25 dex accuracy in density predictions and showing robust width-density coupling across scales. When compared with the AI-based DDPM model, MDR demonstrates stronger interpretability, lower computational cost, and competitive performance, particularly in diffuse or extended structures where DDPM struggles with priors and biases. The results highlight MDR as a transparent, efficient complement to AI methods, offering physical insights, reliable benchmarks, and practical tools for analyzing multi-scale ISM structure and star formation potential. Key outputs include width maps $l_c$, thickness estimates $l_t$, and predicted $\langle n(H)\rangle$ maps, with validation showing consistent high-density tails in the density PDFs and agreement with independent tracers such as HC$_3$N.

Abstract

Interstellar medium widely exists in the universe at multi-scales. In this study, we introduce the {\it Multi-scale Decomposition Reconstruction} method, an equation-based model designed to derive width maps of interstellar medium structures and predict their volume density distribution in the plane of the sky from input column density data. This approach applies the {\it Constrained Diffusion Algorithm}, based on a simple yet common physical picture: as molecular clouds evolve to form stars, the density of interstellar medium increases while their scale decreases. Extensive testing on simulations confirms that this method accurately predicts volume density with minimal error. Notably, the equation-based model performs comparably or even more accurately than the AI-based DDPM model(Denoising Diffusion Probabilistic Models), which relies on numerous parameters and high computational resources. Unlike the "black-box" nature of AI, our equation-based model offers full transparency, making it easier to interpret, debug, and validate. Their simplicity, interpretability, and computational efficiency make them indispensable not only for understanding complex astrophysical phenomena but also for complementing and enhancing AI-based methods.

Figures (14)

• Figure 1: Pipeline of Multi-scale decomposition reconstruction to measure width map and predict volume density distribution. The top panel shows the raw data of input, which is the column density of the molecular cloud, OMC-1. The middle panels present the calculated process of the Multi-scale Decomposition Reconstruction. The bottom panels are the output productions, width map and volume density distribution.
• Figure 2: Width map of molecular clouds NGC 1333 and OMC-1. The panels present the width distribution at the plane of sky in molecular clouds NCG 1333 (left) and OMC-1 (right), which is equal to twice the distance from the local position to the structure center. The black contours show the distribution of raw data, column density from Herschel Gould Belt Survey, whose levels are from 10$^{21.5}$ to 10$^{23.5}$ cm$^{-2}$ with steps as 10$^{0.25}$. Scale annotations (circles) show representative large (a visual guide, e.g. 0.15 pc) and small (spatial resolution) scales for comparing the derived widths with physical ISM structures.
• Figure 3: Distribution between the characteristic scale, predicted volume density, and true volume density in FLASH simulation filament clouds, M3. The top panels show the H$_2$ column density and true volume density computed by 3D density structure of MHD simulation. The middle panels display the characteristic scale r derived by H$_2$ column density (see Eq.\ref{['eqcr']}), predict volume density (see Eq. \ref{['eqdensity']}), and the distribution between pseudo volume density and true volume density. The red line shows the n(H)$_{\rm predict}$ = n(H)$_{\rm true}$. The bottom panels present the predicted volume density map and the distribution between predicted volume density and true volume density with a binning size of 50. The red line means n(H)$_{\rm predict}$ = n(H)$_{\rm true}$. The histograms are constructed with a constant bin width of $\Delta \log_{10} (n) = 0.067$ dex This figure compares predicted volume density (MDR) and projected volume density along the z-axis, which the result of other projected directions is shown in Fig. \ref{['figM3xy']}.
• Figure 4: Distribution between the characteristic scale, predicted volume density, and true volume density in Flash simulation, extended clouds, M4. They are the same as Fig. \ref{['figM3']} but for cloud, M4 with an extended structure. The comparison result along with other projections is shown in Fig. \ref{['figM4xy']}.
• Figure 5: Effective thickness vs. density profile along the LOS in MHD simulation. The blue lines show the density profiles along the LOS ($z$-axis) direction for the M4 cloud, which exhibits complex structure containing mixed clumps and filaments. Red arrows present the effective thickness $l_{\rm t}$ derived from column density using Multi-scale Decomposition Reconstruction. Dashed red curves show Gaussian fits to the density profiles, from which FWHM sizes are extracted.