Abstract Time-resolved X-ray computed tomography (4D-CT) enables dynamic processes within objects to be tracked over time. A key application of 4D-CT is the scientifically and societally important study of multiphase flow in porous media. Obtaining high-quality 4D-CT images is challenging because a high data acquisition rate must be combined with bespoke algorithms for faithful reconstruction. Here, we introduce NeCT, a physics-based deep learning model achieving state-of-the-art sparse-view and 4D-CT reconstructions. NeCT enables reconstruction of 4D objects using an implicit neural representation in space and time. With standard micro-CT instruments, NeCT achieves a temporal resolution approaching a few seconds. Reconstructions of benchmark static tomography datasets show that NeCT also outperforms established algorithms in 3D. We demonstrate fast imaging of liquid imbibition in sandstone, with the high spatiotemporal resolution allowing salt dissolution and pore-filling events to be directly observed. Finally, opportunities beyond classic CT with our dynamic continuous approach are discussed.
Friis et al. (Mon,) studied this question.