Inverse Rendering of Fusion Plasmas: Inferring Plasma Composition from Imaging Systems

TL;DR

The paper presents a differentiable rendering framework for inverse plasma diagnostics in tokamaks, enabling estimation of poloidal distributions of neutrals, electrons, and temperature directly from camera images. By coupling photon-emissivity-based emission, null-scattering light transport, and path-replay backpropagation, it computes gradients with respect to plasma maps and optimizes a multi-term loss that includes rendering, cross-section, midplane, and regularization components. The authors introduce a flux-surface parameterization of plasma quantities, a log-space encoding to enforce positivity, and a hierarchical optimization strategy validated on SOLPS-derived data with both full and cropped/quantised images. The approach yields plausible reconstructions of $n_{\mathrm{neutral}}$, $n_{\mathrm{electrons}}$, and $T_{\mathrm{electrons}}$, providing a potential new diagnostic pathway that complements simulations and can leverage multi-modal measurements. This differentiable, gradient-enabled framework scales to larger problems and can integrate additional diagnostics to constrain the inversions.

Abstract

In this work, we develop a differentiable rendering pipeline for visualising plasma emission within tokamaks, and estimating the gradients of the emission and estimating other physical quantities. Unlike prior work, we are able to leverage arbitrary representations of plasma quantities and easily incorporate them into a non-linear optimisation framework. The efficiency of our method enables not only estimation of a physically plausible image of plasma, but also recovery of the neutral Deuterium distribution from imaging and midplane measurements alone. We demonstrate our method with three different levels of complexity showing first that a poloidal neutrals density distribution can be recovered from imaging alone, second that the distributions of neutral Deuterium, electron density and electron temperature can be recovered jointly, and finally, that this can be done in the presence of realistic imaging systems that incorporate sensor cropping and quantisation.

Figures (8)

• Figure 1: Images from shots 30306, 30356 and 30419 in the MAST tokamak, obtained by the Photron APX-RS (RBB) camera mounted on port HM10. These images show the variety of plasma emission configurations.
• Figure 2: First Row: Cross-sections obtained from a SOLPS simulation of MAST shot 30356 Havlkov2015 as interpolated onto a regular 2D grid. Second Row: Parametric representation of the cross-sections using the formulation described in \ref{['subsec:syntheticplasma']}. Third Row: Residuals of the parametric reconstruction. The absolute $\log_{\mathrm{10}}$ difference between the maps in the first row and the second row.
• Figure 3: Rendered using Mitsuba 3 Mitsuba3 from SOLPS cross-sections using PEC coefficients obtained from Open-ADAS UniStrathclyde2021 to compute the volumetric emission.
• Figure 4: Experiment \ref{['experiment:neutrals-only']}: Neutrals reconstruction using $n_{\mathrm{electrons}}$ and $T_{\mathrm{electrons}}$ as known priors. The first row shows the reference images and cross-sections, the second row shows the reconstructions and the last row shows the residuals of these reconstructions. The renders are tonemapped using $\tau=0.109\mathrm{s}$ and $\gamma=0.625$.
• Figure 5: Experiment \ref{['experiment:all-maps-full-res']}: Reconstruction of total density $n$, ionisation fraction $\chi$ and electron temperature $T_{\mathrm{electrons}}$ using renders and midplane values as the ground truth. The first row shows the reference images and cross-sections while the second row shows the reconstructions. The renders are tonemapped using $\tau=0.109\mathrm{s}$ and $\gamma=0.625$.