Do Androids Dream of Magnetic Fields? Using Neural Networks to Interpret the Turbulent Interstellar Medium. (arXiv:1905.00918v1 [astro-ph.IM])
<a href="http://arxiv.org/find/astro-ph/1/au:+Peek_J/0/1/0/all/0/1">J. E. G. Peek</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Burkhart_B/0/1/0/all/0/1">Blakesley Burkhart</a>
The interstellar medium (ISM) of galaxies is composed of a turbulent
magnetized plasma. In order to quantitatively measure relevant turbulent
parameters of the ISM, a wide variety of statistical techniques and metrics
have been developed that are often tested using numerical simulations and
analytic formalism. These metrics are typically based on the Fourier power
spectrum, which does not capture the Fourier phase information that carries the
morphological characteristics of images. In this work we use density slices of
magnetohydrodyanmic turbulence simulations to demonstrate that a modern tool,
convolutional neural networks, can capture significant information encoded in
the Fourier phases. We train the neural network to distinguish between two
simulations with different levels of magnetization. We find that, even given a
tiny slice of simulation data, a relatively simple network can distinguish
sub-Alfv’enic (strong magnetic field) and super-Alfv’enic (weak magnetic
field) turbulence >98% of the time, even when all power spectral information is
stripped from the images. In order to better understand how the neural network
is picking out differences betweem the two classes of simulations we apply a
neural network analysis method called “saliency maps”. The saliency map
analysis shows that sharp ridge-like features are a distinguishing
morphological characteristic in such simulations. Our analysis provides a way
forward for deeper understanding of the relationship between
magnetohydrodyanmic turbulence and gas morphology and motivates further
applications of neural networks for studies of turbulence. We make publicly
available all data and software needed to reproduce our results.
The interstellar medium (ISM) of galaxies is composed of a turbulent
magnetized plasma. In order to quantitatively measure relevant turbulent
parameters of the ISM, a wide variety of statistical techniques and metrics
have been developed that are often tested using numerical simulations and
analytic formalism. These metrics are typically based on the Fourier power
spectrum, which does not capture the Fourier phase information that carries the
morphological characteristics of images. In this work we use density slices of
magnetohydrodyanmic turbulence simulations to demonstrate that a modern tool,
convolutional neural networks, can capture significant information encoded in
the Fourier phases. We train the neural network to distinguish between two
simulations with different levels of magnetization. We find that, even given a
tiny slice of simulation data, a relatively simple network can distinguish
sub-Alfv’enic (strong magnetic field) and super-Alfv’enic (weak magnetic
field) turbulence >98% of the time, even when all power spectral information is
stripped from the images. In order to better understand how the neural network
is picking out differences betweem the two classes of simulations we apply a
neural network analysis method called “saliency maps”. The saliency map
analysis shows that sharp ridge-like features are a distinguishing
morphological characteristic in such simulations. Our analysis provides a way
forward for deeper understanding of the relationship between
magnetohydrodyanmic turbulence and gas morphology and motivates further
applications of neural networks for studies of turbulence. We make publicly
available all data and software needed to reproduce our results.
http://arxiv.org/icons/sfx.gif
