X-ray computed tomography (CT) is an important technique for characterizing the 3D structure of materials, and, has found broad applications in medical imaging, biological research, and material science. Recently, due to the advances in x-ray source technology which provides very high beam intensities, there is great interest in applying CT to track 3D dynamics of samples, sometimes referred as 4D imaging. However, conventional CT reconstruction algorithms require that the sample does not change over the course of the CT data acquisition time, which limits the achievable temporal resolution of CT. The timescale of the sample dynamics being investigated must be much longer than the CT data acquisition time, in order to fulfill the stable sample requirement during data acquisition. At the same time, there is increasing concern about x-ray induced sample damage and undesired x-ray induced interactions. Thus, there is great interest and research to significantly reduce the CT data acquisition time and at the same time, reduce the sample dose. Here, we describe a new machine-learning (ML) based algorithm that significantly reduces CT data acquisition time as well as overall x-ray dose. With this approach, we achieved an ultrafast nano-tomography with sub-10 s data acquisition time and sub-50 nm spatial resolution with a Transmission X-ray Microscope. We applied our new algorithm to study the dynamic morphology changes in a Li-ion battery cathode material under a heating rate of 50 oC/s and found a self-healing of cracks during the sintering process. The proposed reconstruction protocol can be applied to other tomography modalities.