Reconstructing Richtmyer-Meshkov instabilities from noisy radiographs using low dimensional features and attention-based neural networks

TL;DR

The paper addresses reconstructing Richtmyer–Meshkov instabilities from noisy radiographs by leveraging low-dimensional, robust dynamic features and attention-based neural networks. It introduces two architectures, ShockDecoderViT (a conditional variational autoencoder with a Vision Transformer) and Mass-Conserving Transformer (a deterministic, mass-conserving transformer), operating on time-series shock/edge features extracted from radiographs to predict density fields in a dynamic radiography pipeline. The authors generate a large ICF double-shell dataset with varied MG EOS parameters and perturbations, and they incorporate a noise model for feature extraction to test robustness. Results show both methods achieve accurate density reconstructions, with Mass-Conserving Transformer offering higher structural similarity and robustness, and the work demonstrates the feasibility of extracting RMI growth rates from reconstructed density fields, potentially enabling experimental validation in spherical geometries.

Abstract

A trained attention-based transformer network can robustly recover the complex topologies given by the Richtmyer-Meshkoff instability from a sequence of hydrodynamic features derived from radiographic images corrupted with blur, scatter, and noise. This approach is demonstrated on ICF-like double shell hydrodynamic simulations. The key component of this network is a transformer encoder that acts on a sequence of features extracted from noisy radiographs. This encoder includes numerous self-attention layers that act to learn temporal dependencies in the input sequences and increase the expressiveness of the model. This approach is demonstrated to exhibit an excellent ability to accurately recover the Richtmyer-Meshkov instability growth rates, even despite the gas-metal interface being greatly obscured by radiographic noise.

Figures (16)

• Figure 1: Example double shell capsule specification based on the 1.06 MJ yield design from Ref. merritt19.
• Figure 2: Sample $(r, z)$ projection of the density (1$^{\rm st}$ column), zoomed-in view of the Richtmyer–Meshkov interface (2$^{\rm nd}$ column), synthetic radiograph (3$^{\rm rd}$ columns), and a zoomed-in view of the radiograph (4$^{\rm th}$ columns) labeled with the RMI interface (left half) and Canny edge labels (right half).
• Figure 3: Left: 3D mock-up of a Tantalum shell (green) with a perturbation on the interior surface and an outer ablator layer (gray). Middle: projection of the Tantalum shell onto $(r, z)$ coordinates. The inner radius is parameterized by the angle, $u$, between the white dotted line and the $r$ axis. The drive from the ablator is modelled as an initial velocity on the Tantalum shell ($v_{\rm impl}$). Right: Plot of the 20 separate profiles for radius of the perturbed inner surface verses angle $u$.
• Figure 4: Example plots of the density evolution (a) and the various RMI profiles representing each inner surface perturbation profile (b).
• Figure 5: Upper left: radiograph before applying noise. Upper right: radiograph after applying noise. Bottom left: horizontal line out through the center of the radiographs. Bottom right: vertical line out through the center of the radiographs. In both line plots, the orange line corresponds the noisy radiograph and the blue line corresponds to the radiograph without noise.