Overview of the SuNeRF-CME method for tomographic reconstruction of coronal mass ejections (CMEs) using physics-informed neural radiance fields (NeRFs). The approach leverages multiviewpoint observations of total tB and polarized pB brightness to estimate the electron density in the heliosphere. The tomographic reconstruction is performed via a ray tracing approach (red): (1) Rays are traced per pixel into the reconstruction volume, sampling spatial-temporal coordinates x, y, z, t along the line of sight. (2) A neural network maps these coordinates to local electron density ρ and plasma velocity v = (v_x, v_y, v_z). (3) The electron density and coordinate positions are used to compute Thomson-scattered radiance, integrating along the line of sight. (4) The resulting estimated total and polarized brightness values are compared to the observations, and the model is updated to minimize the deviation. To address challenges posed by sparse viewpoint coverage, additional physical constraints are incorporated (blue): (1) Points are randomly sampled within the 3D domain. (2) The neural network maps the coordinate points to electron density and velocity. (3) Using automatic differentiation, the derivatives of the output values with respect to the input coordinates are constructed. (4) In combination with the ray tracing loss, the model is constrained by minimizing the residuals of the continuity equation, regularizing for an approximately radial solar wind, and enforcing consistency with an expected ambient solar-wind speed range.