<p>Accurately reconstructing the three-dimensional spatial structure and energy distribution of precipitating electrons in auroras is essential for understanding the magnetosphere–ionosphere coupling and related dynamic processes. Building from the recent advances in machine learning and computer vision, we propose a novel method that leverages implicit neural representations for the reconstruction of the auroral electron flux. We model the electron flux as a neural network mapping a continuous horizontal position at the top of the atmosphere to a histogram of the energy distribution at the given position. We use an atmospheric physics model to compute the emitted photon flux and electron density profiles from the neural representation. These predictions are matched against actual multi-angle images and radar measurements to optimize the network through gradient descent learning. To regularize this ill-posed problem, we enforce a smoothness constraint via a second-order derivative penalty on the reconstructed spectra. Our technique achieves down to four times smaller reconstruction error within a few minutes of training compared to previous Aurora Computed Tomography methods. Furthermore, it eliminates the need for explicit initial conditions and naturally scales to larger reconstruction volumes and additional sensor measurements without additional performance cost.</p> Graphical abstract <p></p>