Geologists date and match rock layers across distant sites by the fossils they contain. If a species first appears at one depth in one borehole and at another depth in a second borehole, those two depths are assumed to be the same moment in time. This rests on an untested assumption: that a fossil lands where the currents of its own time put it, rather than being sorted, carried downstream, or winnowed away by the same flows that deposit the sediment. Automated tools for matching layers have been built and scored only against human-drawn correlations of field data, where the true timelines are never known. We propose to build, in a laboratory flume, sedimentary sequences whose true timelines are recorded on camera. Each sediment release carries its own set of "fossil" tracers, so every layer has a known age. These sequences then let us measure how much transport shifts apparent fossil boundaries and test which correlation methods actually recover the true time lines.
In plain terms, we want to find out whether fossil tie points drift with transport distance by an amount we can predict from how fast each particle type settles. If they do, a correlation method that loosens its trust in each tie point according to that particle's settling behavior should match layers better than methods that treat fossil boundaries as exact. The flume supplies what field data cannot, which is a true answer to score against.
Core Claim
Our hypothesis is that, in gravity-current deposits, the depth where a tracer "zone" begins or ends is not the same moment everywhere along the flow path. The offset depends on which transport regime a site sits in and on the tracer's settling velocity. Buoyant, low-density organic particles such as pollen should run ahead of their host sediment into the distal, suspension-dominated reaches. Denser or larger fragments should lag in the proximal bed-load reaches. Two consequences follow. Correlation methods that treat biozone boundaries as fixed anchors should show systematic, position-dependent timing errors. And a correlation method that models this offset should reduce those errors. A second claim, which may be the more useful one, concerns how alignment methods are currently judged. High signal similarity after alignment, measured with Pearson correlation, is the standard score. We expect it to be a poor stand-in for timing accuracy when the true isochrons (surfaces of equal time) are known.
Grounding and Prior Work
The physical setting comes directly from the flume study of lacustrine hyperpycnal flows. That study ran 20 sand–silt–clay gravity-current experiments in a 27 m, 3° flume monitored by eight cameras. It showed that transport passes through three stages: bed load over 0–9 m, suspended load over 9–13 m, and beyond 13 m a lift-off or "hydroplaning" stage that forms a detached new head. It then mapped four facies zones. It added no biological particles, which leaves open exactly the question we ask. We reuse its apparatus design and regime map as the method. We add tracers, timed releases, and coring.
The algorithmic side is Deep Hierarchical Graph Correlation (DHGC), a two-stage method. A one-dimensional convolutional network embeds windows of Gamma-Ray logs, and a constrained graph search (Dijkstra with monotonicity and slope limits) finds the matching path between two wells. On 89 unseen North Sea well pairs it lifted mean Pearson correlation from 0.35 to 0.91, with 98.9% success. Unconstrained dynamic time warping (DTW), a classical method that stretches and compresses one sequence to fit another, reached a higher Pearson of 0.96. The authors rejected DTW anyway because it collapses long intervals onto single points, which they call singularities. This paper motivates our proposal in two ways. Its ground truth is human interpretation, and it uses no fossil tie points. It also leaves unresolved whether DTW's singularities are artifacts or real gaps in deposition. The flume can settle that question because erosion surfaces are filmed.
On the fossil side, two field studies show that biozones are treated as definitive horizons. One is the graptolite zonation R1–R9 of the Renheqiao Formation study. The other is the palynomorph zones of borehole ZKY2-1. The Renheqiao work also reports a negative correlation between graptolite abundance and organic carbon, which it reads ecologically. The research file notes that differential hydrodynamic sorting is an untested competing explanation. We treat that finding as a secondary target our experiments can inform. The Namigata mollusc re-evaluation shows what is at stake: a wrong fossil-based correlation misplaced a formation by about 20 million years.
Research Program
We would build a flume matching the published geometry and run sequences of sediment releases. Each release carries a unique combination of modern tracer particles chosen to span a range of hydraulic behavior. These would include small spores, saccate pollen, and denser fragments standing in for graptolite periderm. The release sequence defines "biozones" whose true first and last appearances are known in clock time. Camera records of the bed surface, together with thin dyed marker releases, give the true isochrons along the whole flume.
After each sequence we take cores at stations inside each of the three transport regimes. Every core yields two records. One is a continuous log of grain size from laser particle analysis, which acts as a stand-in for a wireline log. The other is a discrete series of tracer counts from chemical extraction and counting, in the same way palynomorphs are processed in the ZKY2-1 study. This produces many core pairs with exact timing truth. It also covers conditions that broke DHGC: short segments, high-frequency layering, and filmed erosional gaps.
We would then build one correlation method. It extends constrained graph search with fossil tie points whose allowed timing offset is learned as a function of transport regime and tracer settling velocity, calibrated on held-out runs. Settling velocity is measured independently in a settling column, so the model is not tuned to the answer. Finally we fit a simple numerical advection–settling model to the same runs. This tests whether measured offsets can be predicted without the flume.
Baselines
The naive reference point treats tracer zone boundaries as exact time lines and interpolates depth linearly between them. This is how a biozone scheme is usually read, and our hypothesis predicts it will show errors that grow with distance between stations and across regime boundaries.
Sliding cross-correlation of the grain-size logs is the simplest signal-matching method. It is known to fail under nonlinear stretching, as the DHGC paper reports. We include it as a floor.
Unconstrained DTW on the grain-size logs is the strongest method on the metric the field currently uses: 0.96 mean Pearson on the North Sea set, as reported in the DHGC paper. We run it with the same inputs and no training, and we score both Pearson and timing error. We also check every DTW singularity against filmed erosion or non-deposition intervals.
DHGC is the strongest learned method, with a reported 0.91 mean Pearson and 98.9% success on 89 North Sea pairs. We run it in two ways: the published architecture trained on North Sea logs as the file describes them, and the same architecture retrained on flume runs. Both use the same run split and the same hyperparameter search budget as our method.
The key ablation is our fused method with fossil tie points but with the offset term removed, so tie points are treated as exact. Any gain of the full method over this ablation isolates the settling-aware ingredient rather than simply the addition of fossils.
Our method counts as beating a baseline if its mean isochron depth error on held-out runs is lower in every transport regime and the difference exceeds run-to-run variation across repeated sequences. We commit to reporting cases where DTW or DHGC wins.
Why not just…?
Why not just use DTW, which already scores highest? DTW's 0.96 Pearson in the DHGC study measures how similar the two logs look after warping, not whether matched depths are the same age. In the flume we can measure both. If DTW also has the lowest timing error, and its singularities line up with filmed gaps, then DTW is right and the geological objection to it is wrong. We would report that as a finding.
Why not just validate on numerical synthetic stratigraphy? This is a real threat, because simulation is far cheaper. The flume work shows effects that simple transport models do not assume, such as hydroplaning heads and clay sustaining turbulence (flume study). We test this threat directly. If an advection–settling model calibrated on some runs predicts tracer offsets on held-out runs within their measured spread, synthetic benchmarks are adequate and the flume's main value becomes calibration. If it fails, specifically in the lift-off zone, physical benchmarks are needed.
Why not just use absolute dating or marker beds? Radiometric and orbital-tuning clocks are often unavailable or overprinted in sedimentary rock. The Caiwei Guyot work, for example, still needed a magnetic-reversal anchor before it could tune its record. Our dyed marker layers serve only as ground truth. They are withheld from every method, so they cannot substitute for the thing being tested.
Why not just trust expert correlation? Experts are exactly what DHGC was scored against, and the Namigata case shows experts can be confidently wrong. We would ask practitioners to correlate a subset of flume cores blind, and score them on the same timing error.
Decisive Checks
The central check is whether tracer boundary offsets differ significantly across the three transport regimes and rank by independently measured settling velocity. If they do not, the core claim fails, and the settling-aware method has no reason to exist. We would then publish the benchmark alone. The second check is whether the ranking of methods by Pearson agrees with their ranking by isochron error. A disagreement would show that current benchmark practice rewards the wrong thing. The third check is whether flume-retrained DHGC and our method transfer across runs with different sediment mixes. This limits how far any conclusion can be pushed toward real basins. Scaling from a 27 m flume to kilometer-scale systems remains outside what we can claim. The results would instead show which algorithmic assumptions survive a physical system where the answer is known.
References
- Sedimentary processes and patterns in deposits corresponding to freshwater lake-facies of hyperpycnal flow – An experimental study based on flume depositional simulations
- Deep Hierarchical Graph Correlation: A Two-Stage Approach to Well-Log Alignment Using CNNs and Dynamic Programming
- Hydrocarbon-Generating Assemblages and Organic Matter Accumulation Patterns from the Basal Silurian Renheqiao Formation in Western Yunnan, China
- Late Cretaceous (Santonian to Campanian) Palynological Records and Paleoclimatic Significance from Borehole ZKY2-1, Songliao Basin
- Molluscan fauna of the “Miocene" Namigata Formation in the Namigata area, Okayama Prefecture, southwest Japan
- Geochronology and Geochemical Properties of Mid-Pleistocene Sediments on the Caiwei Guyot in the Northwest Pacific Imply a Surface-to-Deep Linkage