Neural networks swap places with black hole image simulations, both directions

The work

Technical · from the record

The authors extended a variational autoencoder with a second decoder branch that maps latent variables to the physical parameters of an image, calling the result ALINet, and trained a separate network, InvNet, to map physical parameters back to latent distributions so images can be produced from parameters. The networks were trained on a library of ray-traced radiatively inefficient accretion flow (RIAF) black hole images, using 160,000 images for training, 20,000 for validation and 20,000 for testing, with five varying physical parameters and a five-dimensional latent space. On the test data the paper reports 1-sigma errors of 2.5% or less for all parameters, for both ALINet and ALINet+InvNet, and compares these with spin prediction errors of around 20% for visibility amplitude analysis and 5% with closure phases reported elsewhere. The paper reports image generation taking about 1 ms with one Tesla T4 GPU, against a few minutes for the ray-tracing simulations.

How AI was used

A library of RIAF black hole images was first produced by conventional ray tracing over a grid of ten values in each of five physical parameters, and the parameters were rescaled to the unit interval before the library was split into training, validation and test sets. A variational autoencoder was modified by attaching a second decoder branch of fully connected layers that maps the latent variables to the image's physical labels, with a loss function combining image reconstruction, a beta-weighted Kullback-Leibler regularisation term and a weighted sum-squared-error term on the parameters (alpha = 4 x 10^4). The encoder was an augmented AlexNet-style convolutional network and the first decoder branch its mirror, with sigmoid outputs on both branches. This network was trained for 45 epochs at batch size 64 with the learning rate stepped from 10^-3 to 10^-4 to 10^-5 in blocks of 15 epochs. A second network, InvNet, was then trained on the frozen ALINet encoder's latent means and standard deviations to invert the mapping from physical parameters to latent distributions, for four epochs in total with alpha = 10. The trained ALINet was run over the held-out images to recover parameters, and InvNet followed by the ALINet decoder was run over held-out parameter sets to generate images, with errors compared against the simulation truth. The same architecture, in a simpler form, had first been trained on MNIST with a ten-dimensional latent space to test the design.