Machine-Learning Estimation of Energy Fractions in MHD Turbulence Modes

TL;DR

The paper tackles the problem of decomposing MHD turbulence energy into the Alfvén, slow, and fast modes in the ISM. It introduces a FiLM-conditioned residual neural network trained on three-dimensional isothermal and multiphase MHD simulations to infer mode fractions directly from spectroscopic maps (intensity, centroid, and velocity-channel maps). The key result is that Alfvén and slow modes dominate the energy budget in multiphase gas, while the fast mode contributes a smaller fraction, with multiphase conditions increasing the fast fraction relative to purely isothermal cases. The approach achieves mean absolute errors of about 0.05 for seen data and 0.1 for unseen data, offering a practical pathway to observationally constrain turbulence composition and refine models of cosmic-ray transport and star formation.

Abstract

Magnetohydrodynamic (MHD) turbulence plays a central role in many astrophysical processes in the interstellar medium (ISM), including star formation, heat conduction, and cosmic-ray scattering. MHD turbulence can be decomposed into three fundamental modes-fast, slow, and Alfvén-each contributing differently to the dynamics of the medium. However, characterizing and separating the energy fractions of these modes was challenging due to the limited information available from observations. To address this difficulty, we use 3D isothermal and multiphase MHD turbulence simulations to examine how mode energy fractions vary under different physical conditions. Overall, we find that the Alfvén and slow modes carry comparable kinetic-energy fractions and together dominate the turbulent energy budget in multiphase media, while the fast mode contributes the smallest fraction. Relative to isothermal conditions, multiphase simulations exhibit an enhanced fast-mode energy fraction. We further introduce a machine-learning-based approach that employs a conditional Residual Neural Network to infer these fractions directly from spectroscopic data. The method leverages the fact that the three MHD modes imprint distinct morphological signatures in spectroscopic maps owing to their differing anisotropies and compressibilities. Our model is trained on a suite of isothermal and multiphase simulations covering typical ISM conditions. We further demonstrate that our machine learning model can robustly recover the mode fractions from spectroscopic observables, achieving mean absolute errors of approximately 0.05 for seen data and 0.1 for unseen data.

Figures (6)

• Figure 1: Mode decomposition method. We separate Alfvén, slow, and fast modes in Fourier space by projecting the velocity Fourier component $\pmb{v_k}$ onto bases $\hat{\xi}_{\rm a}$, $\hat{\xi}_{\rm s}$, and $\hat{\xi}_{\rm f}$, respectively. Note that $\hat{\xi}_{\rm s}$ and $\hat{\xi}_{\rm f}$ lie in the plane defined by $\mathbf{B_0}$ and $\hat{\mathbf{k}}$. Slow basis $\hat{\xi}_{\rm s}$ lies between $-\hat{\theta}$ and $\hat{\mathbf{k}}_\parallel$. Fast basis $\hat{\xi}_{\rm f}$ lies between $\hat{\mathbf{k}}$ and $\hat{\mathbf{k}}_\bot$. From 2003MNRAS.345..325C.
• Figure 2: Illustration of three MHD turbulence modes after mode decomposition. Panel (a): isothermal simulation min05 ($\beta = 0.02$). Panel (b): multi-phase simulation B6v5 ($\beta = 0.0061$).
• Figure 3: Dependence of energy fractions of the three MHD modes on plasma $\beta$. The colors represent different modes — red for Fast Mode, yellow for Slow Mode, and green for Alfvén Mode. Solid lines denote isothermal simulation, while dotted lines correspond to multi-phase simulation.
• Figure 4: Boxplot of energy fraction of Alfvén mode, Slow mode, and Fast mode in three separate Multi-phase simulations and one isothermal simulation. Error bars indicate 1 sigma variance. From left to right: CNM, UNM, and WNM phase. Color differs from simulation. The red box refers to the isothermal simulation, which is consistent between phases.
• Figure 5: A comparison of the predicted mode fraction and the ground truth using simulation B3v2.5. Panel (a): the input feature maps. Panel (b): the ground truth. Panel (c)-(e): the predicted mode fraction with different feature maps as training inputs.