Supplemental information Reducing weighting matrix The size of weighting matrix is important in the determination of reconstruction time and computing memory. When the number of voxel and pixel nodes are l×m×n (Nvox) and o×p (Npix), the size of weighting matrix is 3(Npix×projection number)×Nvox. The weighting matrix is really sparse, and a lot of zero components exist in the matrix. To remove the zero components for reducing the size of weighting matrix, the matrix is divided into many slices along z direction and the velocity in each slice is independently reconstructed. The vertical regions (r ranges) affected by the velocity information in each slice of the voxel are determined by the reconstructed voxel size, pixel size, and Dso. Then, the distance Dso can be used to make a single line in the projected plane correspond to each slice of the voxel. In this study, the weighting matrix can be reduced to 3(o×projection number)×(l×m). Using this reducing matrix, the reconstruction time is decreased, and the computing memory is effectively used. For example, if the camera has 1024 × 1024 pixels resolution and the size of interrogation window is 32 × 32 pixels with 50 % overlapping in the PIV analysis, the pixel nodes are 64 × 64. In addition, the voxel nodes and the number of projections are determined as 64 × 64 × 64 and 7, respectively. Then, the size of the reduced matrix (1344 × 262144, about 3.5 × 108) is much smaller than that of the original matrix (86016 × 262144, about 2.3 × 1010). The LSMR method and SMART algorithm The 3D velocity information is reconstructed using the least squares minimum residue (LSMR) method and simultaneous multiplicative algebraic reconstruction technique (SMART) algorithm. The LSMR method utilizes a conjugate-gradient type algorithm to solve sparse linear equations and sparse least-square problems based on the Golub–Kahan process. It computes a solution x for the problem of Ax = b. Here, A is a rectangular matrix of dimension m × n, and B is a vector of length m. If the system is inconsistent, it solves the least-square problem min ∥b - Ax∥2. In our system, the A and b correspond to the weighting matrix W and the measured velocity vector P, respectively. In the Golub-Kahan process, the solution is determined when backward error estimate (the Frobenius norm of residual vector, ∥rk∥ = ∥b - Axk∥) is smaller than the quantity depending on ATOL and BTOL which are involved in the practical stopping criteria of the computing process. The stopping rules of the computing for obtaining the solution are as follows: S1: Stop if ∥rk∥ ≤ BTOL∥b∥ + ATOL∥A∥∥xk∥ S2: Stop if ∥ATrk∥ ≤ ATOL∥A∥∥rk∥ If Ax = b ise consistent, S1 is applied. Otherwise, the rule S2 is applied. The accuracy of the LSMR method depends on the dimensionless quantities ATOL and BTOL. In order to reduce the effect of those quantities, the SMART algorithm is applied after finding the LSMR computation. Conventional tomographic PIV techniques commonly use the SMART algorithm. The solution is determined based on the product of the ratio of the recorded velocity to the projected velocity at each pixel node for each iteration k: V jk 1 Pi k V j k i n WinVn Ni Wij 1 / Ni where Ni is the total number of pixel nodes that observe a given voxel j. The relaxation parameter μ is typically chosen in the range of 0–2. Synthetic particle images and RMS error estimation Three different flow images were simulated. The first image was in response to a circular pipe flow with only one velocity component. The two others were axisymmetric and asymmetric flows similar with those used by Dubsky et al. The 3D displacement of the axisymmetric flow between two consecutive images is expressed by the following: x c1 ( y x 2 y y 3 ) y c2 ( x xy 2 x 3 ) z c3 (1 x 2 y 2 ) and the 3D displacement of the asymmetric flow is similarly expressed by: x c1 (1 x 2 y 2 ) y c2 (1 x 2 y 2 ) z c3 (2 x 3 x 5 xy 2 y 2 x 3 x) where c1, c2, and c3 are 5, 5, and 15 pixels in both cases. The reconstruction accuracy was estimated by the root mean square (RMS) error, obtained by the following formula RMS v 3n vexact 2 recon 3n where vrecon and vexact are the reconstructed and exact velocity at each voxel node. The digit n denotes the total number of voxel nodes. Scan angle and iteration number effect on the reconstruction accuracy The effects of the scan angle and iteration number of SMART method were investigated by simulation study. The variation in RMS error in relation to the scan angle for the asymmetric flow case is shown in Fig. S1(a). When the scan angle was larger than 30°, the results were acceptable qualitatively and quantitatively. Therefore, only the results for scan angles larger than 30° are depicted in Fig. S1(a). No peculiar trend in RMS error is shown. The minimum RMS errors for 11 and for 16 projections were obtained at the scan angles of 60° and 45°, respectively. The corresponding optimal angle increments are 6° and 3° for 11 and 16 projections. The effect of iteration number of SMART method on the RMS error method for asymmetric flow is illustrated in Fig. S1(b). The RMS error monotonically decreased as the iteration number increased. The RMS error remained constant when the iteration number exceeded 50. Image pre-processing Figure S2(a) shows a typical X-ray image captured without any objects. Given the honeycomb structure of the fiber optics plate (FOP) attached in front of the CCD camera, the corresponding unwanted image pattern was observed. In addition, X-ray illumination was uneven because of the anode heel effect. Therefore, flat field correction (FFC) pre-processing was adopted to improve the quality of the captured X-ray images in this study. The FFC is expressed as follows: FFC image object image offset image scale factor gain image offset image where the gain image is the image captured without test samples under the same experimental conditions, and the offset image is that taken without X-ray illumination. Compared with the raw image [Fig. S2(b)], the contrast of the tracer particles was much higher [Fig. S2(c)]. This implies that FFC pre-processing significantly improves the image contrast. The projected velocity field was obtained after FFC image correction. Thereafter, the 3D velocity fields were reconstructed from the projected velocity field information. List of Figures Figure S1. (a) Variations in RMS error with respect to scan angles at 11 projections (red circle) and 16 projections (black triangle) for the asymmetric flow case (iteration k = 50). (b) Effect of iteration number on 11 projections with a 60° scan angle (red circle) and on 16 projections with a 45° scan angle (black triangle) for asymmetric flow. (a) (b) Figure S2. Typical X-ray images (a) without test samples, (b) raw image, and (c) preprocessed by FFC. (a) (b) (c)