Super time-resolved tomography

Abstract

Understanding 3D fundamental processes is crucial for academic and industrial applications. Nowadays, X-ray time-resolved tomography, or tomoscopy, is a leading technique for in-situ and operando 4D (3D+time) characterization. Despite its ability to achieve 1000 tomograms per second at large-scale X-ray facilities, its applicability is limited by the centrifugal forces exerted on samples and the challenges of developing suitable environments for such high-speed studies. Here, we introduce STRT, an approach that has the potential to enhance the temporal resolution of tomoscopy by at least an order of magnitude while preserving spatial resolution. STRT exploits a 4D DL reconstruction algorithm to produce high-fidelity 3D reconstructions at each time point, retrieved from a significantly reduced angular range of a few degrees compared to the 0-180 degrees of traditional tomoscopy. Thus, STRT enhances the temporal resolution compared to tomoscopy by a factor equal to the ratio between 180 degrees and the angular ranges used by STRT. In this work, we validate the 4D capabilities of STRT through simulations and experiments on droplet collision simulations and additive manufacturing processes. We anticipate that STRT will significantly expand the capabilities of 4D X-ray imaging, enabling previously unattainable studies in both academic and industrial contexts, such as materials formation and mechanical testing.

Figures (5)

• Figure 1: STRT concept. (a) Data acquisition of STRT vs. tomoscopy. Tomoscopy requires a full 180-degree scan to capture a single time point, while STRT reduces the required scan angle to much less than $180\degree$. (b) X-Hexplane tensorial model. X-Hexplane contains six feature planes spanning each pair of coordinate axes (e.g., XY, ZT). Points in spacetime are projected to each plane. The features extracted from the six planes are fused and sent to a tiny MLP to get $n(\textbf{x},t)$ at a specific spatiotemporal point. (c) X-Hexplane rendering and cost function. 2D projections from a given experiment angle and time point can be rendered by integrating $n(x,t)$ along the ray direction at that specific time point. The parameters of X-Hexplane are updated by minimizing the MSE loss between the predicted results and the measured projections.
• Figure 2: STRT reconstructions for droplet collisions compared to the ground truth. The left column depicts the 4D ground truth for different time points (frames) from two orthogonal views: one along the projection axis (side view) and one perpendicular to this plane (top view). The other columns show the corresponding STRT reconstructions for 60$\times$, 20$\times$, and 10$\times$ temporal enhancement (TE).
• Figure 3: Spatial resolution determined by FSC as a function of time. (a) illustrates the resolution changes during droplet collision with temporal enhancements (TE) of 60$\times$, 20$\times$, and 10$\times$. (b) presents the evolving revolution in additive manufacturing over time with temporal enhancements of 200$\times$, 20$\times$, and 10$\times$.
• Figure 4: STRT reconstructions for additive manufacturing and ground truth. The ground truth and STRT results for 200$\times$, 20$\times$, 10$\times$ temporal enhancement (TE) are displayed in separate columns, while different rows represent distinct time frames of the printing process. The dynamics of the printing processes are marked by solid squares.
• Figure 5: Comparison of the reconstruction performed with 10$\times$ temporal enhancement (TE) with the ground truth (a) volume rendering, (b) volume rendering clipped by slices reconstructed using corresponding approaches (ground truth and 10$\times$ TE), (c) results of material segmentation highlighting liquid surfaces, (d) cross sections along the orange line indicated in (c).