Engineers who design pneumatic conveyors, slurry lines and fracturing tools need to know where a pipe wall will wear through and how that changes with particle size and flow speed. Today each new geometry or operating point requires its own particle-tracking fluid simulation: a computational fluid dynamics run in which millions of particles are pushed through a solved airflow and every wall strike is converted into mass loss by an empirical wear formula. Each run takes hours, and a design sweep multiplies that cost. Computer vision faced the same problem with 3D scene reconstruction, where every scene was once fitted from scratch. Large feed-forward networks now replace that per-scene fitting with one network pass, and a few cheap correction steps recover the detail a single pass misses. We propose to apply that recipe to erosion, with one twist that comes from the physics.
Erosion depends non-monotonically on particle size. In a coal riser it peaks for particles of 300–350 µm, even though heavier particles hit harder and lighter particles hit more often. Taken literally, this "energy versus frequency" trade-off cancels exactly: at fixed mass flow, the number of impacts scales as $1/d^3$ and the energy per impact scales as $d^3$. The peak must therefore come from two other effects. The first is the chance that a particle reaches the wall at all, which small, streamline-following particles rarely do. The second is the speed and angle at which particles arrive, since large particles lag the carrier flow. Our core claim is that a feed-forward model should learn only these two size-dependent hydrodynamic quantities. The mass bookkeeping and the wear law should be applied exactly, outside the network. A model built this way should place the erosion peak correctly at particle sizes and geometries it never saw. It should also let the user swap wear laws without retraining, both of which a network that regresses erosion directly cannot promise.
Core Claim
For each point $x$ on the wall surface, the network predicts two quantities. The first is $p(x \mid d, U, G)$, the expected number of impacts per unit area per injected particle, given particle diameter $d$, inlet gas speed $U$ and geometry $G$. The second is a conditional distribution $q(v, \theta \mid x, d, U, G)$ over impact speed $v$ and impact angle $\theta$. The erosion rate then follows in closed form:
$$\mathrm{ER}(x) = \frac{M}{m(d)}\, p(x)\, \mathbb{E}_{q}\big[\, m(d)\, K v^{k} f(\theta) \,\big] = M\, p(x)\, \mathbb{E}_{q}\big[ K v^{k} f(\theta) \big],$$
Here $M$ is the solids mass flow, $m(d)\propto d^3$ is the particle mass, and $K v^{k} f(\theta)$ is any standard per-unit-mass wear law (Finnie, Oka, McLaury or the Generic model).
The cancellation of $m(d)$ is the point of the design. The cubic scalings sit in fixed arithmetic. The network spends its capacity on how particle inertia changes wall reach ($p$) and arrival kinematics ($q$), which is where the physics of the peak actually lives.
The second ingredient addresses the blur of single-pass prediction. We borrow the recurrent error-feedback idea from ReSplat, which refines Gaussians over four steps using maps of rendering error and beats DepthSplat by 3.5 dB PSNR on DL3DV at a fraction of per-scene optimization cost. In our setting, the analogue of a rendering error map is the mismatch between the predicted impact field and a few hundred particle parcels traced through a coarse, frozen flow field. Tracing those parcels takes seconds. They are far too few to estimate erosion directly, but they are enough to tell the model where it is wrong.
Relation to Prior Work
The physics papers motivate the representation. Ogunlela et al. and their companion study map wear across 1–1000 µm in a 1.2 m riser. They find a steady low regime below the Stokes transition, a sharp rise near 200 µm, a peak at 300–350 µm and a decline beyond 600 µm. They also report that four wear laws agree on the shape of this curve but differ substantially in magnitude. That agreement in shape and disagreement in scale is exactly what our factorization predicts: shape comes from $p$ and $q$, scale from the wear law. Sheng et al. show that larger proppant can reduce erosion through velocity slip, and that erosion leaves geometry-specific fingerprints: a "rabbit's ear" pattern on upper nozzles and a "half bottom loop" on lower ones. This is evidence that the spatial impact distribution, not a scalar rate, is the right prediction target.
Three further papers fix the operating window and the cost of per-case simulation. Zhou needed a mesh of more than 12 million cells and 88 runs for a single 5 m pipe. Zhou also found that lift forces cannot be neglected below 15 µm, so our simulations must include Saffman lift. Liu et al. give the elbow regime: bend ratios $R/d$ of 1, 3 and 5, and air speeds of 5–15 m/s. Jin et al. show a similar threshold behaviour in air curtains, but we leave that application aside.
The vision papers supply the method. GS-LRM shows that a transformer can map inputs directly to explicit primitives in 0.23 s. DGS-LRM extends this to dynamic scenes, replacing 1–3 hour per-scene fits. Both regress quantities with no notion of mass or momentum. We keep their amortization strategy but make the network's output a physically typed quantity, which they do not.
Several papers constrain the design. FluidFormer shows that purely local neural fluid models drift and leak through unseen boundaries, and that global attention plus explicit conservation fixes this. It also shows that all-to-all attention over particles does not scale. We therefore attend over wall-surface points, not particles. PIVNO gets per-case-free velocimetry by enforcing divergence-free flow, which is invalid for the dispersed solid phase. Our physics prior is mass bookkeeping, not incompressibility. Fracture-GS couples impact mechanics to explicit primitives, but only for discrete fracture of monolithic objects, not accumulated particulate wear.
Finally, Monocular Dynamic View Synthesis: A Reality Check shows that good aggregate image scores can hide broken internal dynamics. We take that warning seriously and evaluate the impact statistics $p$ and $q$ themselves, not only the final erosion map.
Research Program
We will first build the data that does not yet exist. We will generate a corpus of one-way-coupled particle-tracking simulations using a standard RANS turbulence closure with Saffman lift. The geometries come from parameterized families: straight risers, 90° elbows across $R/d$ from 1 to 6, tees, contractions, and a nozzle-and-body geometry approximating the fracturing tool as described by Sheng et al. Particle diameters span 1–1000 µm and gas speeds 3–20 m/s. Each run logs every wall impact with its position, speed, angle and particle size, rather than only the integrated erosion. Because the logs are kept raw, any wear law can be applied afterwards, and $p$ and $q$ have direct targets. We will check four-way coupling on a subset to measure where the one-way assumption breaks for dense loading.
The model represents the wall as a point cloud carrying normals and a geometric context feature. It combines local continuous convolutions with global attention over a downsampled set of wall points, following FluidFormer's local-plus-global design but applied to surfaces. Each block is conditioned on $d$, $U$, the particle-to-gas density ratio and the Stokes number. The model's outputs are a nonnegative impact density $p$ and a small mixture model for $q$ at each wall point. Training uses a Poisson likelihood for impact counts and a mixture log-likelihood for the logged speeds and angles. Erosion is never a training target; it is computed from the formula above.
The refinement stage runs $T$ recurrent steps, with $T$ between 2 and 4. Each step traces 200–500 parcels through a flow field solved on a mesh roughly 50 times coarser than the reference. It computes the per-point residual between observed and predicted impact counts and speed-angle moments, and feeds that residual back as an input feature. We will study how accuracy trades against the number of parcels and the number of steps.
Baselines
The naive reference point is the empirical erosion-versus-size curve fitted on training geometries, applied unchanged to a new geometry. This is the "do nothing beyond lookup" baseline that a practitioner already has. Beating it on held-out geometries is the minimum bar.
The strongest existing practice is to simply run the cheap simulation: coarse-mesh particle tracking given the same wall-clock budget as our full pipeline, including the network's inference time. This is the fairest competitor because it uses no learning and no training corpus, and it improves monotonically with compute. We will give it the same coarse flow solution our refinement uses, plus every remaining second of budget as extra parcels. Beating it means lower relative error in the erosion map, and lower error in the location of peak wear, at equal wall-clock. We expect the gap to be largest at the extremes of particle size, where wall impacts are rare and sampling noise dominates.
The strongest learned competitor is the same architecture, data and tuning budget, trained to regress the erosion map directly under each wear law. This isolates the factorization, which is our key ingredient. It wins if the factored model's advantage disappears on held-out diameters and held-out wear laws.
A second ablation removes the recurrent refinement. Comparing single-pass prediction with refined prediction isolates the ReSplat-style feedback. A third ablation feeds the coarse parcels as input once, with no recurrence, to test whether iteration matters beyond simply having the extra data.
No reported result was found on Lacuna for any learned surrogate of wall-erosion maps. The quantitative anchors from the file are physical. Ogunlela et al. place the riser erosion peak at 300–350 µm, and Sheng et al. predict erosion depths of 8–10 mm matching recovered field tools. Our high-fidelity reference simulations must themselves reproduce the first anchor before any comparison is meaningful.
All learned models share one training set, one hyperparameter search budget of 20 trials, and one A100-class GPU.
Why not just…?
Why not just fit response surfaces or a small neural network on operating parameters, as in the walnut separation study? Those models map a handful of scalars to a scalar for one fixed geometry. They cannot take a new pipe shape as input, and they do not produce a map of where wear occurs, which is what the nozzle fingerprints in Sheng et al. show matters. We will still include a per-geometry scalar surrogate in the peak-diameter test. If it matches our model on peak location within a known geometry, we will report that honestly. The claimed advantage is generalization across geometries, not accuracy within a single one.
Why not just run more parcels on a coarse mesh? This is the real threat, and it is why it is our primary baseline at matched compute. Ogunlela et al. find that wall reach for small particles is governed by near-wall dynamics, and Zhou finds that lift forces matter below 15 µm. Both suggest that a coarse mesh biases impact statistics in exactly the regime where the peak forms. Extra parcels reduce noise but not this bias, whereas our learned prior was trained on fine-mesh references. If the coarse simulation matches us at equal time, the learned component adds nothing, and we will say so.
Why not just regress erosion end-to-end and let the network discover the physics? The ablation tests this directly. Our prediction is that direct regression must relearn each wear law's magnitude scale, which the four laws in Ogunlela et al. disagree on substantially. It must also extrapolate a sharp peak in particle size, whereas the factored model only needs to interpolate smooth functions of the Stokes number. FluidFormer's finding that unconstrained neural predictors fail on unseen boundaries supports the expectation, but the experiment decides.
Proposed Experiments
The decisive checks are four held-out splits.
The first holds out particle sizes. We train on 1–250 µm and 500–1000 µm and test on 250–500 µm, measuring error in the peak diameter on the reconstructed coal-riser configuration (1.2 m diameter, 12.5 kg/s, steel density 7700 kg/m³). Success means placing the peak inside 300–350 µm.
The second holds out geometry. We train on risers and elbows and test on tees and the nozzle-tool geometry. We measure relative L2 error of the erosion map, distance error of the peak-wear location, and rank correlation of predicted wear across candidate designs. We also check whether the two-lobe and bottom-loop patterns appear.
The third holds out the wear law. We train impact statistics using logs only, then evaluate erosion under the Oka and McLaury laws, which were never used in model selection.
The fourth tests whether the statistics themselves are right, following the Reality Check lesson. We report error in the impact density and divergence between predicted and logged speed-angle distributions separately from erosion error, so a model cannot score well through compensating errors.
All results are reported against wall-clock time, where the reference fine-mesh simulation defines the hours-scale cost being amortized.
References
- ReSplat: Learning Recurrent Gaussian Splats
- CFD Approach to the Influence of Particle Size on Erosive Wear in Coal Riser Pipes
- Investigating the Effect of Particle Size on Erosive Wear in Industrial Coal Pneumatic Transport Using Computational Fluid Dynamics
- CFD analysis and field observation of tool erosion caused by abrasive waterjet fracturing
- Numerical simulation of pneumatic conveying characteristics of micron particles in horizontal pipe
- Numerical simulation of flow characteristics in vertical elbows of pneumatic conveying
- The Influence of Air Curtains on Fire Smoke in Tunnels
- GS-LRM: Large Reconstruction Model for 3D Gaussian Splatting
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos
- FluidFormer: Transformer with Continuous Convolution for Particle-based Fluid Simulation
- PIVNO: Particle Image Velocimetry Neural Operator
- Fracture-GS: Dynamic Fracture Simulation with Physics-Integrated Gaussian Splatting
- Monocular Dynamic View Synthesis: A Reality Check
- Parameters optimization of a walnut shell-kernel separation device using response surface methodology and ANN coupled genetic algorithm