4D imaging of sub-second dynamics in pore-scale processes using real time synchrotron x-ray tomography

A variable volume flow cell has been integrated with state-of-the-art ultra-high speed synchrotron x-ray tomography imaging. The combination allows the first real time (sub-second) capture of dynamic pore (micron) scale fluid transport processes in 4D (3D + time). With 3D data volumes acquired at up to 20 Hz, we perform in situ experiments that capture high frequency pore-scale dynamics in 5-25 mm diameter samples with voxel (3D equivalent of a pixel) resolution of 2.5 to 3.8 μm. The data are free from motion artefacts, can be spatially registered or collected in the same orientation making them 15 suitable for detailed quantitative analysis of the dynamic fluid distribution pathways and processes. The methods presented here are capable of capturing a wide range of high frequency non equilibrium pore-scale processes including wetting, dilution, mixing and reaction phenomena, without sacrificing significant spatial resolution. As well as fast streaming (continuous acquisition) at 20 Hz, it also allows larger-scale and longer term experimental runs to be sampled intermittently at lower frequency (time-lapse imaging); benefiting from fast image acquisition rates to prevent motion blur in highly dynamic systems. 20 This marks a major technical breakthrough for quantification of high frequency pore scale processes: processes that are critical for developing and validating more accurate multiscale flow models through spatially and temporally heterogeneous pore networks.


Introduction
Porosity, permeability and flow in geological systems are all highly dynamic. Changes in the confining conditions (e.g. pressure, temperature), flow volume, fluid chemistry/viscosity or suspension composition can drive mass transport (through processes such as precipitation, dissolution, deposition or erosion) and change the connectivity and tortuosity of the pore 5 network. The evolving porosity and permeability then cause further changes in both micro-and macro-scale flow. These mass transport processes and the passage of fluid-fluid and fluid-rock reaction fronts through heterogeneous geological systems occur at the pore scale. Developing a detailed appreciation of macroscopic processes as diverse as pollutant transport, hydrocarbon recovery, CO2 sequestration, storage of nuclear waste, aquifer management, nutrient bio-accessibility, building stone preservation and hydrothermal deposit formation, requires a quantitative understanding of the multi-scale effect of pore-10 scale processes.

Synchrotron imaging for dynamic geoscience applications
Laboratory and synchrotron X-ray computed micro tomography (usually XCT & sXCT respectively) are ideal approaches for imaging these processes as they allow in situ observation in a high resolution, non-destructive way. The general principles of X-ray tomography, reconstruction and data processing as relating to geoscience applications are covered in a series of recent 15 reviews Cnudde and Boone, 2013;Hess et al., 2011;Maire and Withers, 2014;Wildenschild and Sheppard, 2013). For the specific application of understanding pore scale processes, XCT and sXCT are now becoming widely used for both qualitative and quantitative imaging of complex natural pore networks, and the distribution of liquid(s) within them (Al-Raoush et al., 2011;Al-Raoush and Willson, 2005;Berg et al., 2013;Bhreasail et al., 2012;Boone et al., 2014;Cnudde and Boone, 2013;Dewanckele et al., 2012;Geraud et al., 2003;Herring et al., 2013;Iglauer et al., 20 2011;Katuwal et al., 2015;Ma et al., 2016;Naveed et al., 2013b;Olafuyi et al., 2010;Sakellariou et al., 2003;Sok et al., 2010;Wildenschild et al., 2002;Wildenschild and Sheppard, 2013). The data are also being used as a basis of, and validation for numerical simulations (Al-Raoush and Papadopoulos, 2010;Alhashmi et al., 2015;Bultreys et al., 2015;Degruyter et al., 2010;Menke et al., 2015;Naveed et al., 2013a;Raeini et al., 2015;Raeini et al., 2014;Walter Fourie et al., 2007). 25 Standard operation is to collect a set of 2D "projections" or "radiographs" at constant angular spacing (hereafter angular density) while the sample is rotated through 180° or 360°. Standard 3D tomographic image datasets require seconds (synchrotron), minutes (synchrotron and laboratory) or even hours (laboratory) to acquire (Maire and Withers, 2014). Until recently this has limited 3D experimental investigations of dynamic processes because the critical pore scale processes occur 3 over much shorter durations. Many key fluid-rock and fluid-fluid interactions therefore remain poorly constrained.
Furthermore, while numerical simulations can now consider multiple mass and thermal transport processes simultaneously and can incorporate realistic pore geometries (see  for in depth review), many generally still lack validation from experimental data from natural systems where observations are needed with a range of temporal and spatial resolutions.
In recent years, careful alignment of 3D datasets collected at fixed time points over moderate duration experiments has enabled 5 "time integrated" tomography as a tool for quantification of fluid dynamics and porosity evolution (Andrew et al., 2014;Andrew et al., 2015;Armstrong et al., 2014a;Berg et al., 2013;Blunt et al., 2013;Herring et al., 2014;Herring et al., 2013;Lin et al., 2016;Menke et al., 2015;Wildenschild and Sheppard, 2013). The increased image acquisition rates now available at 3rd generation synchrotron facilities has driven development of faster "continuous" imaging. In this mode, sample rotation is not stopped for each projection, instead data is acquired over a narrow arc (typically 0.1-0.5°). For most geological materials 10 collection of the projection data can be just a few seconds Berg et al., 2013;Bhreasail et al., 2012;Pistone et al., 2015;Pistone et al., 2013;Youssef et al., 2014). However, 2D imaging (the same projection acquisition rates but without rotation) is still needed to observe processes occurring on the millisecond-few second timescales; such as individual Haynes jumps and the subsequent relaxation dynamics (Armstrong et al., 2014b;Berg et al., 2013). 2D imaging cannot capture pore, and pore-fluid interface morphology, pore throat orientation and size, or the location of the events within the fluid volume. 15 Advances in temporal resolution through camera and beam line technology have been followed by highly parallelised iterative reconstruction methods which require far fewer projections, yet still achieve adequate reconstructed image quality (Batenburg and Sijbers, 2011;Brabant et al., 2014;Kaestner et al., 2015;Kazantsev et al., 2015a;Kazantsev et al., 2015b;Van Eyndhoven et al., 2015). However, direct in situ observation of sub-second events and processes, and those that cause rapid change over longer durations still remains challenging. Pushing time integrated tomography towards true long duration 4D imaging would 20 therefore have high impact on our understanding of pore scale processes.
Here we present a state-of-the-art 4D (3D + real time) imaging methodology that enables visualization and quantitative assessment of dynamic pore scale processes in real time (3D acquisition rates of up to 20 Hz) over variable experimental durations. We present two key methodology developments: sub-second real time imaging of fluid transport up to an order of magnitude faster than previously presented, and the extension of that imaging protocol to allow slower or variable flow 25 dynamics to be quantitatively assessed over longer time periods. We illustrate the potential of these developments for providing insight into pore-scale processes using three case studies showing preliminary data from experiments that quantify initial wetting, the evolution from dynamic to steady state flow, and the evolution of the reaction front during chemical dilution or fluid mixing.

2 Experimental set up
Disaggregated sandstone gravels (1-2, or 2-4 mm size fraction) were loaded into reusable cylindrical gravity fed flow cell and mounted on the JEEP i12 beamline (Diamond Light Source, Harwell, UK) ( Figure 1). Full beam specification can be found in Drakopoulos et al. (2015). The cell diameter for this system can be varied (up to 50 mm diameter) according to experimental need and x-ray transmission through the sample. Perspex sample chambers of 6 and 25 mm internal diameter were used here. 5 A circular piece of fine mesh was placed at the bottom of the cell above the outlet feed to prevent fine material being washed through to the outflow reservoir (maximum volume 125 ml in the configuration used). Particles were added to a minimum depth of 15 mm to ensure no effects from basal flow were observed in the imaging field of view. Particle diameters were sufficiently small that the influence of the cell geometry on particle packing was minimal. The random nature of the packing structure was checked before each experiment through the collection of a single tomography data set ( Figure 2). In the well 10 cemented host sandstones (prior to disaggregation) grain sizes ranged from 200-500 µm, with pore diameters of 10-150 µm.
Mercury and helium porosimetry and water absorption measurements on the disaggregated samples were 13-18%, 15-23%, and ~19% respectively, and the measured permeabilities ranged from 250-600mD. In the configuration used here, fluid was supplied at a constant flow rate by a 12 channel peristaltic pump operated from the beam line control room (outside the x-ray enclosure or "hutch") so that flow rates and compositions could be adjusted in real time. i.e., while 3D data are being acquired. 15 The inlet tube was attached centrally to the top of the cell using a slip ring to ensure a fixed drip point (approx. 5 mm above the centre of the rotating sample). Flow rates were varied from 40-200 μl/min, which corresponded to a drip every 2 and 10 seconds. The majority of experimental runs were performed using a 6 M KI solution to ensure high contrast between air, fluid and rock particles. Pilot scans showed un-doped H2O and some oils could be observed under the hard x-rays of the JEEP white beam (50-150 kV), although low contrast would make post-processing challenging. 20 The flow cell was designed to be easily modified; and can accommodate base fed laminar flows, simultaneous flow of multiple fluids (variable composition or temperature), or multi-phase suspensions. Thin walled (1 mm) aluminium sample chambers can be used for both variable thermal and pressure experiments. The cell can be loaded with disaggregated material (e.g. soil, sand, gravel, glass beads) or with more coherent cored soil/rock samples of any diameter. However, when designing ultra-high speed experiments or processes that cause subtle changes in attenuation/composition, some higher density or large volume 25 solid samples will increase the minimum exposure time and may also cause artefacts that reduce the ability to resolve features of interest.

Ultra-fast image acquisition
The maximum rotation speed of the JEEP tomography stage is 10 Hz (continuous rotation). As synchrotron tomography datasets are acquired over 180°, 3D volumes can be acquired at up to 20 Hz. The actual operational rotation speed for given 30 experiment will be controlled by the exposure time for each projection and the total number of projections collected per 3D data set. Exposure time is controlled by x-ray flux and sample density. Imaging at the JEEP beam line can be performed using a polychromatic (white, higher flux) or monochromatic x-ray beam (selectable energy range 53-150 kV). Images were acquired using a Vision Research Phantom Miro™ 310M camera (up to 5000 images per second). In monochromatic beam mode exposure times for the large diameter cell were 200 µs (3.8 μm pixel resolution), dropping to 90 µs for the smaller diameter 5 system (2.5 μm pixel resolution). The system can be in continuous rotation at this speed, in one direction, for as long as required. Beam hardening can occur under white beam in response to the energy dependence of the attenuation, but none was seen in our test data. While white beam would allow shorter exposure times (between 30-50% reduction in the test scans), there would be lower contrast between the low density phases and more x-ray scattering and associated noise in the images.
Imaging in monochromatic beam gave increased contrast between air and water in a three phase system, without requiring 10 long exposures. The diameter of the samples used here exceeds the field of view of the highest magnification module available at JEEP (1.25 μm pixel resolution), but region of interest images could be obtained with a 200 µs exposure. Qualitative analysis of the images at different rotation speeds showed no visible dependence of liquid distribution on acquisition (rotation) speed, suggesting the centrifugal effect on this sample and fluid was negligible.
In most "traditional" tomography experiments, over 1000 projections are collected for each 3D volume. We collected high 15 resolution data (1800 projections, 0.2 and 0.5 s per 3D data acquisition) for each dry sample prior to flow initiation to characterise the inter-and intra-grain porosity of each sample (Figure 2A, B). This dataset can also be used as a structural prior for iterative reconstruction methods (e.g.; Kazantsev et al. 2015 a, b;van Eyndhoven et al 2015). Reducing the angular density of the projections will shorten overall scan time. The reconstructed image quality does reduce as the projection density drops, but a significant increase in temporal resolution can be achieved. To achieve the maximum 20 Hz 3D image acquisition rate, 20 it is necessary to reduce the number of projections to 250 (200 µs exposure) and 550 (90 µs exposure) per scan.
Under-sampling (using fewer projections) to this degree does cause image quality to degrade, making differences between phases of similar attenuation difficult to observe, and reducing the precision with which phase boundaries and volumes can be defined. When features of interest are relatively small, or have complex geometries then under-sampling can prevent quantitative analysis. However, here the contrast between the saline solution, rock and air is high, and the primary target is 25 the sub-second distribution of the fluid phase. Using the routine filtered back projection reconstruction algorithms, testing with as few as 360 ( Figure 2C) and 180 projections ( Figure 2D) gave adequate images for basic quantification ( Figure 2E).
Comparison of the two under-sampled datasets shows little increase in quality and little overall change between the 360 and 180 projection scans. For the smaller cells, it is possible to collect 360 projections at maximum image acquisition frequency.
The Miro camera has fixed on board data storage, which for the field of view needed in this experiment allowed a maximum 30 of approximately 20000 projections to be collected (in this set up, the number of projections is dependent on the field of view 6 needed). The under-sampling allowed between 53 and 141 3D datasets to be acquired in a single experimental run (2. [5][6][7] seconds of acquisition at maximum frequency). Capturing a specific event with this short a duration of collection is challenging, especially as the high rotation speeds prevents observation of small volume changes in the live view radiographs.
During collection the camera is armed and then triggered. Arming the camera starts projection acquisition, but that data is automatically overwritten until the trigger signal is received. The user can define the trigger to mark a) the last image; all 5 existing projections on the card are saved, b) the first image; projections are collected until the card volume is full, or c) an event; with a user defined number of projections before and after the trigger are saved. The latter option allows a small amount of buffer time for the user to capture a specific event, such as a drip entering the sample, or a flow volume/composition change.
Automatic triggering mechanisms such as light gates can be incorporated although in this experiment a webcam was used to observe the delivery of each drip. 10 To clarify the real-time nature of the data, the following terminology is used in all subsequent discussion: one or more scans are collected during each experiment, each scan contains a number of 3D "frames" which were acquired at a known speed, and with a given frequency (defined by the time between the first projections of sequential frames). Data are offloaded from the camera between scans (download takes between 5 and 7 minutes). Several scans can be collected during an experiment.
To highlight specific processes and features, most figures presented below show only 2D slices (either perpendicular or parallel 15 to flow) or a sub-volume of the reconstructed data set. The post processing and image segmentation are the same for all examples, and were performed in Avizo™ and MATLAB. The methods are outlined in Appendix 1.

Case study: Observing sub-second dynamics using ultra-fast imaging
Wetting processes and the formation of rivulet networks are poorly understood in heterogeneous media. The processes operate on short times scales and are strongly influenced by sample morphology. Non-invasive 4D tomography is ideal for capturing 20 the dynamics of these behaviours. Figure 3 shows the evolution of the fluid distribution in a sub volume of the sample immediately below the drip point. The data were collected under continuous acquisition mode, with a 3D frame acquisition frequency of 0.07s (system operating at 7Hz rotation speed).
The first drip of KI solution (shown in blue) enters the dry sandstone bed (rendered in translucent grey) between Frame 0 and Frame 1 (Figure 3). Initially (t = 0.07s) fills the inter-grain pore space to the front of the sub volume (A), and immediately 25 begins lateral transport into the intra-grain porosity in the neighbouring grains (B, Figure 3). By the second frame (t = 0.14s) the inter-grain volume has largely emptied, and the majority of the fluid is now within the grains. Intra-grain transport has also occurred with capillary driven wetting of grains not in direct contact with inter-grain fluid (e.g. C), and from other viewing angles drainage of previously saturated intra-grain pores can also be observed. Between F1 and F2 most grains in contact with the inter-grain fluid have become partially or fully saturated. The remaining intra-grain fluid is generally limited to a few small 7 pores, and surface films (D). The figure only shows the relaxation after the first drip. As the experiment continues (F3 to F5), the fluid moves away from the initial location, with grain saturation increasing by radial and vertical redistribution, with oblique upward propagation of the saturation front in some instances (E). Most grains show an expansion and coalescence of grain-surface fluid films after initial intra-grain saturation (F), although grain surface wetting does occur with no, or prior to intra-grain saturation (G). Capillary forces continue to redistribute the intra-grain pore fluid over the remainder of the 5 experiment, with saturation of many grains decreasing as the system moves towards equilibrium. Grains at both the upper and lower extent of the wetted region (H) are almost completely saturated and then drained again within 1.54 seconds.
Vertical slices taken through the same sub-volume of the sample in the second scan of the experiment shows the arrival of the third drip (Figure 4). In this experiment the fluid feed was stopped during data download. When the pump was restarted, the first drip was used as the trigger for the scan acquisition. The initial frames of the second scan were before the arrival of the 10 drip, and act as a wet reference. Differences in the fluid volume across the download period result from the fall of a drip at the very end of the experiment, immediately after fluid feed was stopped. Figure  Although the focus of this contribution is on method development, we include a limited selection of the quantitative analysis performed on these samples, as an illustration of the adequacy of the data quality. Quantitative analysis of inter-and intragrain fluid volume shows the strong fluctuation over the course of two drips ( Figure 5) and the general instability of the fluid volume as it is redistributed. After the first drip, the saturation of the intra-grain porosity becomes more gradual, with significant increases on the arrival of subsequent drips. The inter-grain distribution equilibrates more slowly. 25

Ultra-high speed imaging for longer duration experiments
The experiments presented above allow capture of the high speed dynamics, but have very short overall durations (< 10 seconds). Capturing longer duration processes with the same 3D frame acquisition rates requires lower frame acquisition frequencies. The Miro camera allows frames to be acquired either continuously (as above) or at user defined intervals (hereafter 8 called gapped acquisition). Gapped acquisition decreases frame acquisition frequency, while maintaining frame acquisition speeds, and the resultant data therefore benefits from a lack of motion induced blurring in the reconstructed data.
Although any time or angular gap can be defined, in these experiments they were limited to odd multiples of 180°. This ensures that the initial projection of every frame was collected at the same angle. Registration of the data in this manner saves significant time aligning data sets during post processing. A gapped acquisition with a spacing of 1 (G1) collects over the 0-5 180° sector of every rotation (1 full rotation between acquisition start points); a gapped acquisition with spacing of 5 (G5) acquires from 0-180° on every third rotation (3 full rotations between acquisition start points).
Although not used in this experiment, the camera memory can also be partitioned allowing for a fixed and constant number of projections/frames in each partition. This could enable more efficient imaging across different temporal scales in one scan, or the inclusion of much longer delays between high speed frame acquisitions. 10

Case study: Capturing multi-scale dynamics in pore-scale processes
To show the suitability of this method for capturing the variability of pore-scale flow dynamics, wetting experiments were scanned repeatedly at different gap lengths (continuous through to G25). Between 53 and 106 frames were collected in each scan. High frequency data acquisition (continuous or G1) was used to capture the first few drips, with longer acquisition intervals (G3-G25) used to capture the slower processes as the flow network evolved towards steady state through time ( Figure  15 6). In some experimental runs, variability in the high speed processes through time was also assessed, by repeating continuous, G1 or G3 acquisition at intervals between G5 or G7 acquisitions.
The larger fluid volume in these images for longer durations experiments makes visualisation of the 3D fluid distribution challenging to show on the page (as in Figure 3), and so a semi-transparent render of a distance map is used to show the variability in the local fluid volume ( Figure 6). Areas of inter-grain fluid appear as red regions, while the majority of the intra-20 grain fluid is shown in blue and green. In all scans at all frame acquisition frequencies, we see significant variability in the distribution of the KI solution ( Figure 6) over the 35-60 minute experiments.
From the same data it is possible to track the local changes in saturation at the grain-scale. Well-established network analysis and morphological quantification algorithms can be applied to extract details of saturation on the pore scale (Figure 7) or on the geometry of individual phase interfaces and contact angles (Andrew et al. 2014. For under-sampled data higher 25 levels of image noise in filtered back projection reconstructions may limit quantitative surface curvature and contact angle analysis, but optimised iterative reconstruction methods may give more suitable data (see Section 5 below).

Case Study: In situ observation of mixing dynamics
One key application for ultra-high speed in situ imaging is to track the interaction between multiple phases: capturing both spatial and temporal heterogeneity of replacement, dilution and chemical reactions. After 30 mins under a 6M saline feed the system was flushed with H2O. Scanning was performed at G1 (0.07 s scan acquisition time, 7 Hz acquisition frequency), the scan started on arrival of the first H2O drip. Other experimental runs tested replacing H2O with KI solution, and replacing oil 5 with KI solution. The need for ultra-high speed imaging to capture the complex and spatially variable mixing and dilution processes can be seen in Figure 8. Dilution of the uppermost part of the local inter-grain network is instantaneous ( Figure 8A), but percolation of the dilution/replacement front through the image volume takes approximately 5 seconds ( Figure 8A-F Some grains show variability in the KI concentration within the porosity, implying that although the bulk porosity of this sandstone is approx. 20%, and the connectivity is high, individual grains have discrete pore networks within them (blue boxes F45). Using slower data acquisition methods would not capture the dynamics of replacement. 15

Moving beyond the state-of-the art
The ultra-high imaging presented here allows capture of wetting and transport behaviour at the moderate-to-high spatial resolution necessary for quantitative understanding of the processes involved. The 3D frame acquisition speed and frequency are substantially faster that has been previously achieved. Events such as Hayne jumps (millisecond time-scale), and the subsequent relaxation (a few seconds) have been observed in 2D imaging (acquisition of projections without rotation) 20 (Armstrong et al., 2014b). The method presented here can provide data on the 3D nature of those changes, at comparable timescales. Such quantitative data can be used to develop or validate larger scale flow models and to better incorporate pore scale processes.
The data presented here are reconstructed using standard filtered back projection reconstruction algorithms. However, these methods are not optimal for under sampled data. Recent advances in iterative reconstruction algorithms use information about 25 the different phases contained within the data, to enhance the quality of the reconstruction, with great effect (Figure 9). Visual comparison of this type of method (Figure 9 B, D) with reconstructions using filtered back projection (Figure 9 A, C) shows the improvement in phase resolution that can be achieved on fewer projection data. This example applies an iterative optimisation solver, and along total variation penalisation is incorporated as a regularisation step (Rudin et.al.,1992;Little and Jones, 2010) that takes the sample porosity into account (i.e. uses the data itself as prior information), with refinement at each 30 iteration. The iterative data have fewer line artefacts and lower noise, and permits accurate phase segmentation from fewer than 45 projections.
Iterative methods have significant potential to develop imaging dynamic pore-scale processes still further. Of specific interest are those methods that use high quality dry scan data as a structural prior (Kazantsev et al., 2014;Kazantsev et al., 2015a;Kazantsev et al., 2015c;Van Eyndhoven et al., 2015), and those that divide the data volume into static and dynamic regions 5 using the static information from sequential scans to better define the reconstruction volume Kazantsev et al., 2014;Kazantsev et al., 2015a;Kazantsev et al., 2015c;Van Eyndhoven et al., 2015). Both allow reconstruction of better quality images from under-sampled data (data with fewer projections) and also allow data of comparable or better quality to that displayed here to be produced from datasets with as few as 18 projections (Van Eyndhoven et al., 2015). These methods can improve the quality of the quantitative data, especially on the key phase interfaces, however the key advance is likely to come from working with 10 still higher degrees of under-sampling. However, iterative methods are computationally expensive and can take an order of magnitude longer to reconstruct. The methods applied here took between 15 ( Figure 9D) and 45 ( Figure 9B) minutes per volume, depending on the number of projections in being used. As such, their use in processing every frame of long duration high frame rate data for near real time assessment remains unrealistic at present. This may change as computational efficiency increases, but we envisage that iterative methods are more practically applied retrospectively to achieve higher spatial 15 resolution over short periods of critical interest.
At the 3D volume acquisition rates achieved here, the projection exposure times and mechanical limit of the rotation stage means that imaging faster is not currently possible for the current samples. For higher density materials, or higher resolution cameras where longer exposure times mean slower overall volume acquisition, iterative methods may improve temporal resolution. However, on the JEEP experimental set up, the main advantage of iterative methods is the ability to acquire fewer 20 projections per scan and so collect more scans per experiment without sacrificing image quality. This could enable experimental run times to be extended by up to an order of magnitude.
The methods presented here provide a workflow for visualization and quantification of sub-second dynamics in porous media ( Figure 10). We used the disaggregated sample to capture two spatial scales of fluid-pore interaction, but the overall experimental set up can equally be used for solid core samples at the same spatial and temporal resolution. This experiment 25 was performed at below maximum resolution, and imaging finer structures (grains or pores) at resolutions of 1.2 µm is possible.
For higher resolution or higher density samples slightly slower projection acquisition times may be needed, and in extreme circumstances this will reduce overall frame acquisition frequency. The ability of the experimental set up to acquire large numbers of projections at constant angular density means that even higher temporal resolution could be achieved.
Reconstruction of volumes from datasets with small rotational offsets (typically about 10-20 degrees, i.e. Scan 1 reconstruction 30 from 0-180°, Scan 2 reconstruction from 10-190° etc.) could increase temporal resolution by over an order of magnitude (Van 11 Eyndhoven et al. 2015), especially when coupled to iterative methods that can reduce image noise, and enhance phase boundary definition in the reconstructed data.
Complex multi-phase, spatially heterogeneous micro-and pore-scale processes control many key macro-scale geological system responses. However, the method is not limited to the simple drip-fed experiments on geological materials presented here, and can as be easily applied to high speed processes operating in any material (metals, ceramics, plastics, biological 5 materials, foodstuffs and any other) suitable for synchrotron x-ray tomography at JEEP/i12. It can also be applied in other dynamic systems, for samples and materials undergoing modification through treatment at temperature (hot or cold) or experiencing deformation (tension, compression, cyclic fatigue, fracture or shear). The development of the real-time 4D imaging methods we have presented means that quantitative understanding of all of these processes is now possible, in porous media and beyond. 10 12

Data availability
The projection data and the reconstructed volumes of these experiments are available from the authors. 13 6 Appendix 1

Post processing and image analysis
Visualisation and quantification was performed using Avizo™ and MATLAB. The greyscale volumes were cropped to the sample boundaries to reduce all subsequent processing times. When processing continuous acquisition data the frames were first aligned to allow direct comparison and quantitative analysis. For the gapped scans, all data are acquired over the same 5 angle and no additional alignment is required. The italicised steps in the following refer to built-in algorithms in Avizo™, and the reader is directed to the associated documentation for further details of the implementation.
An edge preserving Bilateral Filter was applied to the greyscale filtered back projection reconstructions to reduce image noise.
Water, air and rock phases were segmented using the Interactive Thresholding, and the resultant binary data volumes were used in subsequent analysis. The inter-and intra-grain porosity was defined by creating two separate binary masks: one by 10 segmentation of the wet and dry grains, and one by segmentation of the fluid (inter-and intra-granular). Logical combination of these allows calculation of the inter-and intra-granular fluid volumes. Fluid volumes were calculated using in-built image measures, (Volume3D).
Additional quantitative analysis of the intra-grain porosity was performed on single grains isolated using the MATLAB command imfill followed by Watershed segmentation. The pore network within the grain was displayed using the Centerline 15 Tree module which applies the TEASAR algorithm (Sato et al. 2000) defining local Euclidian distance to the nearest object boundary. Fluid thickness was calculated by applying a Chamfer 3D distance map to the fluid label volume to calculate the distance to the nearest surface. The exchange of KI for H2O was tracked by calculating a differential image between the initial scan and each subsequent frame.       Figure 6: 3D renders of a gapped experimental run. For each scan series the first, middle and last frame of each acquisition period is shown. Only the fluid volume has been rendered, with the colour representing the distance to the fluid interface. Cool colours represent small fluid volumes; hot colours represent fluid more than 20 µm from the interface. The field of view is 1.5 mm x 1.5 mm x 3 mm. Data from sample LD_1, scans 1 to 5.   H2O at t=0. A-H, 3 D renders of the differential density (greyscale) between the initial saline fluid volume and the saline + water mixture at the times shown. The frame acquisition rate was 0.07s and acquisition frequency was 7 Hz (G1). Data is from sample LD_3, scan 1. Yellow represents complete replacement with H2O. Cooler colours represent volumes that are undergoing dilution. Volumes remaining at 6M KI are not shown, as they have undergone no change.