Streaming Collection and Real-Time Analysis of Tomographic Data at the APS

Introduction Brilliant synchrotron light sources are able to perform continuous tomographic data acquisition at more than 7.7 GB/s rate [1, 2] generating terabytes of data in a very short time, opening the possibility of studying very fast processes at unprecedented high temporal resolution. For example, scientists at the TOMCAT beamline of the PSI and their collaborators have recently set a new world record by demonstrating 1000 tomograms per second (3D image from 40 projections per millisecond) acquisition speed using a new high-speed camera and highnumerical-aperture macroscope.1 The majority of today’s high-speed tomographic equipment captures events in a predefined area of the sample and track sample evolution only through projection data. In many circumstances, this semi-blind traditional technique misses the dynamic phenomena since the location and timing of its origination are not known in advance. Another challenge in studying rapid processes is determining a representative region of interest for scanning, i.e., the region where the dynamic process begins and evolves over time. Most of the time, the dynamic phenomenon is missed because it happens in a location not under observation, evolves quicker than predicted, or demands a different spatial or temporal resolution than the instrument is set to. The ideal environmental control system parameters are another challenge for in-situ research of constantly evolving samples. Without real-time 3D imaging input, it is practically impossible to determine appropriate environmental parameters, such as cooling/heating rates, pressure, or loading forces. There are many studies that would greatly benefit from fast 3D imaging optimized by using real-time image reconstruction for feedback and control. In material engineering and geomechanics, it is important to understand the mechanisms of failure origination [3]. These processes are very challenging for 3D imaging because a crack may start in different parts of the sample. The authors in [2] conducted experiments on 3D imaging of ultrafast formation of dendrites during the solidification of casting alloys or the growth and coalescence of bubbles in a liquid metal foam. Such metal foams based on aluminum alloys are being investigated as lightweight materials, for example for the construction of electric cars. An important topic in Geosciences is to study fast non-equilibrium pore-scale processes including wetting, dilution, mixing, and reaction phenomena, without significantly sacrificing spatial resolution, for example in fast pore-scale fluid dynamics – an incremental capillary-water movement known as the Haines jumps [4]. In [5] the authors used dynamic in-situ imaging to study the process of methane hydrate formation in porous samples. Besides the fact that the methane hydrate dissociation process is very fast, it also occurs at different sample regions, making representative dynamic 3D even more challenging. A conventional approach for data acquisition in tomographic experiments is based on real-time visualization of 2D projections streamed from the detector. These projections are typically used to align the sample on the rotation stage and adjust the detector exposure time. Further tomographic scanning in fly-scan mode involves saving a series of projections while the sample is continuously rotated. After scanning, the acquired data are transferred from the detector computer to a processing and visualization workstation where the reconstruction procedure and the 3D rendering are performed. Data acquisition/transfer and reconstruction become time-consuming, especially in the case of dynamic tomography experiments. Here, we propose to completely change the approach of doing tomography, see Figure 1. Instead of working with 2D projections coming from the detector, we adopt the streaming approach and work with real-time reconstructions. The streaming approach allows for faster adjustment of acquisition parameters, more convenient alignment, easier selection of the region of interest, saved data reduction, much better control of dynamic experiments, and more. In what follows we will briefly discuss most of the details about the proposed streaming acquisition model and demonstrate how we use it at sector 2-BM of the Advanced Photon Source. More details about the model can be found in [6].