Grain Activity Drives Shape Dependence of Bedload Flux

AGU Fall Meeting 2026 · Poster EP41C · 10 December 2026
Shaalin Sehra (MIT EAPS) · shaalin@mit.edu

We tracked glass spheres and natural gravel in the ~1 cm wide flume of Deal et al. (2023) with high-speed side-view video (~4.9 million grain detections). At the same side-wall-corrected Shields number τ*, spheres carry ~1.6× the flux of gravel because ~2.5× as many grains are moving per bed area (from power-law fits), while moving spheres travel at ~0.9× the speed of gravel. Speed at a fixed height above the bed barely changes with τ*; mean speed rises because more grains saltate higher. Gravel stays closer to the bed.

Videos

Side view through the flume wall, mirrored so flow runs left to right; both shapes at the same scale; every clip slowed the same 8.7× (the clock shows real time). Dots mark moving grains, colored by the height of the grain center above the bed: ● rolling/sliding (< 1 D), ● saltating 1–2 D, ● saltating > 2 D (D = grain diameter). τ* is side-wall corrected (Guo, 2015).

Spheres vs. gravel at the same flow strength

Glass spheres on top, natural gravel below, at three matched τ*. Short trails show the last 80 ms of each path; the counters show how many grains are moving right now.

Low τ* (≈ 0.12)

Near the threshold of motion only a few grains move; gravel has ~2 moving grains in view at a time.

Medium τ* (≈ 0.2)

About 49 spheres vs. 28 gravel grains are moving at any moment, and more of the spheres are high above the bed.

High τ* (≈ 0.3)

Many grains move for both shapes and most saltate; spheres still have more grains moving.

How activity and velocity are measured

Counting grains, step by step

One frame is frozen and every moving grain is counted (N). Dividing by the bed area in view (11.5 cm × 1 cm flume width) gives activity n, and n* = (π/6) n D² makes it dimensionless. Arrows show each grain’s downstream speed; their mean is û. Over 2 s at τ* ≈ 0.2: spheres n ≈ 3.7 vs. gravel ≈ 2.0 grains/cm², while û ≈ 25 vs. 23 cm/s.

Trajectories pile up

Each moving grain leaves a permanent trail for about 1 s of real time; the video then fades so only the paths remain. Dashed lines: 1 D and 2 D above the bed.

Low τ*

Medium τ*

High τ*

Extended figures

Fig. S1. Video tracking recovers the weighed flux

Does the video measure the same sediment transport that the trap weighs?

Figure S1

Each point is one flume run: the flux we get from the video (grains moving per area x their speed x grain volume) on the vertical axis, and the flux weighed in the sediment trap on the horizontal axis. Points on the black line agree exactly; the grey band is +/-25%. For glass spheres every run falls on the line (median video/weighed ratio 1.00); for natural gravel the video is a little low (median 0.73), so the gravel activity and flux from video may be somewhat underestimated.

Fig. S2. Experiments spread out along activity, not speed

Why does flux change so much while the average grain speed barely does?

Figure S2

Every point is one experiment, placed by its mean grain speed (horizontal) and its activity, the number of moving grains per square centimeter of bed (vertical). The grey lines connect combinations that give the same flux, because flux = activity x speed; moving up crosses many flux lines, moving sideways only a few. The points climb almost straight up, so running the flume harder mainly puts more grains in motion rather than making each one faster.

Fig. S3. Side-wall correction (Guo)

Why do we use a corrected bed stress?

Figure S3

The flume is only about 1 cm wide but much deeper than that, so the glass side walls take up much of the flow's friction and the usual depth-based estimate misstates the stress the bed actually feels. The Guo (2015) correction separates bed and wall friction using the measured depth, slope and discharge. Left: each run moves straight up from the 1:1 line, by a factor of about 3.3 for spheres and 3.0 for gravel; right: that factor is nearly the same at all stresses, so the correction shifts the runs without reordering them. Open symbols are gravel runs without a discharge measurement, which use the median factor.

Extended methods

  • Experiments. Re-analysis of the Deal et al. (2023) experiments: glass-walled flume at Simon Fraser University, ~1 cm wide. Glass spheres D = 4.9 mm; natural gravel D = 4.1 mm.
  • Video. High-speed side view through the glass wall at 260–520 frames/s (a few sphere runs up to 2080); 20 experiments (14 spheres, 6 gravel).
  • Tracking. Grain centers found in every frame (spheres: ring finder; gravel: YOLOv2 neural network) and linked into tracks; ~4.9 million detections.
  • Moving grains. Speed above 2.5 mm/s and travels more than 8 mm.
  • Activity, velocity, flux. Activity = moving grains per bed area, averaged over all frames; velocity = mean speed of moving grains; flux = activity × velocity × grain volume. Video flux matches the sediment trap for spheres and is ~0.75 of it for gravel (Fig. S1), so flux values use the trap.
  • Shields number. Side-wall corrected (Guo, 2015), which separates bed and wall friction and raises τ* about 3× (Fig. S3).
  • Height and saltation. Height of the grain center above the bed, in grain diameters; saltating = at least 1 D above the bed.
  • Ratios. 1.6×, 2.5× and 0.9× compare power-law fits at the same τ*.

Poster figures, full size

Fig. 1. Video frames: (A) glass spheres, (B) natural gravel. Green: moving-grain center points and velocity vectors; black: bed surface.

Methods sketch: activity n = number of moving grains ÷ bed area (e.g. 19 grains in 13 cm × 1 cm ≈ 1.5 grains/cm²); flux = activity × velocity × grain volume.

Fig. 2. Experiment means vs. side-wall-corrected τ*; flux from the sediment trap. q: volume flux per unit width; n: moving grains per bed area; û: mean streamwise grain velocity; R: submerged specific gravity; g: gravity; D: grain diameter.

Fig. 3. Sample grain paths from sphere and gravel experiments at matching τ*; bold lines: representative paths.

Fig. 4. Grain activity (A, B) and mean velocity (C, D) by height above the bed.

Just for fun

Ultra slow motion

Glass spheres filmed at 2080 frames per second, played 69× slower than real time.

Just the grains

Spheres (top) and gravel (bottom) at high τ*, no overlays.

Under the hood

A wipe reveals the tracking drawn on top of the raw video.

References

Deal, E., et al. (2023). Grain shape effects in bed load sediment transport. Nature, 613, 298–302. https://doi.org/10.1038/s41586-022-05564-6
Guo, J. (2015). Sidewall and non-uniformity corrections for flume experiments. Journal of Hydraulic Research, 53(2), 218–229.
Redmon, J., & Farhadi, A. (2017). YOLO9000: Better, faster, stronger. CVPR 2017.
Rushlow, M. R. (2020). Using machine learning, particle tracking, and grain shape modeling to characterize bed-load sediment transport. B.S. thesis, MIT.

Poster (PDF)

Click the poster to open the full-size PDF.

Poster (44 √ó 36 in)

Poster Poster (44 √ó 36 in)

Open PDF