Measuring the spectral index of turbulent gas with deep learning from projected density maps

TL;DR

This paper demonstrates that deep learning can accurately infer the spectral index of turbulent gas directly from column density maps, bypassing traditional spectral analysis methods, with potential applications in astrophysics.

Contribution

The study introduces a CNN-based approach to estimate turbulence spectral index from density maps, showing promising accuracy and robustness on simulated data.

Findings

CNN predicts spectral index with mean squared error of 0.024

Model performs well across spectral indexes from 3 to 4.5

Robust to different simulation resolutions and altered images

Abstract

Turbulence plays a key role in star formation in molecular clouds, affecting star cluster primordial properties. As modelling present-day objects hinges on our understanding of their initial conditions, better constraints on turbulence can result in windfalls in Galactic archaeology, star cluster dynamics and star formation. Observationally, constraining the spectral index of turbulent gas usually involves computing spectra from velocity maps. Here we suggest that information on the spectral index might be directly inferred from column density maps (possibly obtained by dust emission/absorption) through deep learning. We generate mock density maps from a large set of adaptive mesh refinement turbulent gas simulations using the hydro-simulation code RAMSES. We train a convolutional neural network (CNN) on the resulting images to predict the turbulence index, optimize hyper-parameters in validation and test on a holdout set. Our adopted CNN model achieves a mean squared error of 0.024 in its predictions on our holdout set, over underlying spectral indexes ranging from 3 to 4.5. We also perform robustness tests by applying our model to altered holdout set images, and to images obtained by running simulations at different resolutions. This preliminary result on simulated density maps encourages further developments on real data, where observational biases and other issues need to be taken into account.