Learned simulators for particle-based fluids can match reference positions over short rollouts while getting something more global wrong. Particles may clump, open spurious holes, or mix at the wrong rate. Today's scores, mean squared position error and a distribution-matching distance, compare snapshots, so they see little of how the flow has rearranged particles over time. We propose a cheap, history-carrying record of that rearrangement. Connect the particles into a triangulation: triangles in 2D, tetrahedra in 3D, using the standard Delaunay rule that no particle lies inside any element's circumscribed circle or sphere. As particles move, the triangulation changes only through local swaps called flips. In 2D a flip replaces the shared edge of two triangles with the other diagonal. In 3D the analogous 2-3 move turns two tetrahedra into three, which is a Pachner move from 3-manifold topology. Logging these flips along the trajectory gives each particle and each region a "flip ledger": a running count of neighbor exchanges, which works as a Lagrangian memory of the flow.
The opportunity is that this ledger has two kinds of mathematical backing that ordinary neighbor lists lack. First, a result from 3-manifold topology says that on a surface, the number of flips needed to transform one triangulation into another tracks how tangled the motion that relates them is. Second, a functorial nerve theorem guarantees that the triangulated particle complex faithfully represents the shape of the fluid region, and does so consistently as the particles move. Put together, the ledger should detect surrogate failures, such as wrong mixing, clustering, and voids, that snapshot metrics miss. A topology penalty built on the same complex should also be able to suppress those failures, without the extra inference-time solver steps that current fixes rely on.
Core Claim
For a learned surrogate rolled out on standard Lagrangian benchmarks, two properties should be visibly wrong whenever the rollout is physically wrong. The first is the flip-rate trajectory, meaning flips per particle per unit time accumulated along the rollout. The second is the low-dimensional holes of the particle complex at the smoothing-kernel scale. These errors should appear even when position error and Sinkhorn divergence look acceptable. Of the two, the flip ledger is the history-dependent signal, and the hole count is the snapshot signal it complements.
The claim has two testable parts. The first concerns diagnosis. Flip-rate error should correlate with long-horizon rollout failure better than position error, Sinkhorn divergence, or neighbor-count histograms do. The second concerns correction. A training-time penalty on spurious kernel-scale holes should approach the particle-distribution quality of inference-time SPH relaxation, at lower inference cost.
The novelty lies in turning triangulation moves into a tracked physical history variable. That history is treated as the quantity to measure and constrain. Prior work does not use the triangulation this way. Learned simulators use neighbor graphs only as numerical scaffolding. Work on Pachner graphs explores them statically, with no notion of a physical path through them.
Grounding in Prior Work
The fluid side supplies the problem and the benchmark. LagrangeBench standardizes learned particle simulators, including graph network simulators (GNS) and equivariant variants. It covers 2D and 3D Taylor-Green vortex, reverse Poiseuille flow, lid-driven cavity, and 2D dam break, and scores models with position error, kinetic-energy error and Sinkhorn divergence. It also reports that EGNN fails to converge. Neural SPH diagnoses particle clustering and tensile instability (artificial voids forming under tension) in these surrogates. It fixes them by running one to five classical SPH relaxation steps during inference. Symmetric Basis Convolutions takes a different route and bakes parity symmetry into the convolution itself. These papers motivate our work, since they establish that the failure exists and that physical priors help. They also constrain it: any new prior must beat an inference-time fix that is already very strong.
The history-tracking side supplies the idea of a particle as a ledger. In integral viscoelastic SPH (Santelli et al.; multi-mode and fractional SPH), each particle carries its own deformation history so that memory integrals can be evaluated without back-tracking through a grid. Angerman et al. carry a scalar microstructure variable, the fraction of frictional contacts, along particles in shear-thickening suspensions. These works carry continuum strain history. We carry a combinatorial history of who neighbors whom. That history is more coarse-grained than strain, but it is cheap to compute and has topological meaning.
The topology side supplies the method and its guarantees. Lackenby and Purcell prove that the spine graph of a surface is quasi-isometric to its mapping class group. A spine graph is roughly a graph of triangulation-like structures joined by local moves. The mapping class group is the group of ways to deform a surface, up to continuous wiggling. The theorem therefore says that flip distance between triangulations measures, up to bounded factors, how tangled the deformation relating them is. In a 2D periodic box, the particles are punctures on a torus, and the flow over a time window defines such a deformation. The accumulated flip count is an upper bound on the flip distance, and so it is a computable proxy for the topological complexity of the flow.
We use this result only as motivation for comparisons, for two reasons. The bounding constants depend on the surface, here the number of punctures, and are not known to be calculable. The paper's triangulation-complexity theorem also covers only closed manifolds with pseudo-Anosov monodromy. We therefore compare surrogates only at fixed particle number, where the same constants apply to every model.
Learning 3-Manifold Triangulations studies Pachner graphs statically and hits a memory ceiling at depth 5 even with 200 GB of RAM. We never enumerate that graph. We only walk the single path the physics dictates, so the explosion does not arise. Bauer et al.'s unified functorial nerve theorem gives conditions, which include Delaunay and Voronoi covers, under which the discrete complex is homotopy equivalent to the union of particle balls. The equivalence is natural across a filtration, meaning it holds consistently across the growing ball radii. This licenses reading the holes of the Delaunay-based complex as holes of the fluid region at kernel radius $h$. It also makes interleaving distance a stable way to compare predicted and reference complexes.
Research Program
The first component is the ledger itself. We will maintain a kinetic Delaunay triangulation over each rollout, frame by frame, recording every flip and attributing it to the particles involved. In 2D these are edge flips. In 3D they are 2-3 and 3-2 moves, with occasional repair when frame subsampling is large. LagrangeBench keeps only every 100th solver step, so flips between frames must be recovered by retriangulating and computing a flip path between consecutive frames. That path gives an upper bound on the true count, and we will record the bound's looseness against full-resolution reference runs.
From the ledger we derive the flip-rate curve $\phi(t)$, defined as the mean number of flips per particle per unit time. We also derive its spatial map, and the per-particle accumulated count, which is the history variable. Alongside the ledger we compute the homology of the alpha complex at radius $h$ for each frame. The alpha complex is the sub-complex of the Delaunay triangulation made of elements small enough to fit inside balls of radius $h$. Its homology counts connected components and holes, and the count is justified by the nerve theorem.
The second component is analysis. On the 2D periodic flows, we test whether accumulated flip counts on reference SPH data grow with the flow's stretching, as the mapping-class motivation predicts. If they do not, the theoretical motivation fails and we report that. We then test whether the gap between surrogate and reference in $\phi(t)$, and in the number of spurious kernel-scale holes, predicts eventual rollout divergence earlier and more reliably than snapshot metrics.
The third component is correction. We will train GNS with an added loss that penalizes kernel-scale holes and short-edge clusters in the predicted particle complex, using a differentiable surrogate for each. For holes, the surrogate penalizes alpha-complex elements whose circumradius exceeds the radius $h$ at which they should fill in. For clusters, it penalizes Delaunay edges far shorter than the reference spacing. We then compare this model with Neural SPH relaxation.
Baselines
The naive reference point is plain GNS trained as in LagrangeBench, evaluated with the existing metrics only. It sets the floor that both the diagnostic and the correction must improve on.
The strongest existing method is Neural SPH, meaning GNS plus inference-time SPH relaxation and external-force smoothing. It is the strongest competitor because it targets exactly the clustering and void failures we target, and it reports a very large improvement. On 2D lid-driven cavity, Sinkhorn divergence falls from $3.1\times10^{-4}$ for baseline GNS to $2.8\times10^{-7}$, at roughly 5–15% added inference time per step. We will rerun Neural SPH with its published relaxation settings, on the same datasets, splits, and rollout lengths, and with the same backbone, training budget, and hyperparameter search budget as our penalized model. Beating it means one of two outcomes. The first is matching its Sinkhorn divergence within a factor of a few, with zero inference overhead. The second is beating it on flip-rate error and spurious-hole count at equal or better Sinkhorn divergence.
Symmetric Basis Convolutions is a secondary architectural baseline on its own benchmark suite. It reports a 30% divergence-error reduction over LinCConv with about 30,000 parameters. We include it to check whether a better backbone already removes the defects our ledger detects.
For the diagnostic claim, the competitors are snapshot metrics: position MSE, Sinkhorn divergence, and the neighbor-count histograms used in Neural SPH. Beating them means achieving a higher rank correlation with a held-out measure of long-horizon failure, which we define as the time until position error exceeds a fixed threshold.
The key ablation replaces Delaunay flips with turnover in the fixed-cutoff neighbor list, that is, counting particles that enter or leave each radius-$h$ neighborhood. This isolates whether the triangulation structure matters, or whether any neighbor-change count would do just as well. No reported result was found on Lacuna for flip-based or topology-based evaluation of Lagrangian surrogates.
Why not just…?
Why not just use neighbor-count histograms or density error, as Neural SPH already does? These are snapshot statistics. They can flag clustering at a given frame, but they say nothing about whether particles have exchanged neighbors at the right rate over time. Mixing errors can occur with perfectly plausible density. Our diagnostic experiment ranks surrogates on long-horizon failure using each metric. If neighbor histograms predict failure as well as flip ledgers, the diagnostic claim fails and we will say so.
Why not just run Neural SPH relaxation and stop? It is a real threat on particle-distribution quality, and its lid-driven-cavity number is hard to beat. However, it repairs positions after the network errs, so the learned dynamics stay uncorrected. It also adds solver cost at every step. Our correction experiment asks whether a training-time topology penalty reaches comparable Sinkhorn divergence with no inference overhead. If it does not, we fall back to the ledger as an evaluation tool, applied to Neural SPH outputs as well, to check whether relaxation restores correct mixing or only correct density.
Why not just count changes in the cutoff neighbor list instead of Delaunay flips? This is the most serious alternative, because it is cheaper and needs no triangulation. Its weakness is that the count depends on the arbitrary cutoff radius and carries no guarantee. The Delaunay flip count, by contrast, is tied to flip distance and hence to the complexity of the flow's mapping class, and the alpha complex is tied to the fluid region's shape by the nerve theorem. The ablation above settles the question empirically. If cutoff turnover matches flips on every test, the topological machinery adds nothing in practice, and we would recommend the cheaper count.
Why not just compute persistent homology of the raw point cloud with a generic distance-based complex? Generic complexes built from pairwise distances lack the clean Delaunay/Voronoi nerve guarantee in Bauer et al., and they grow quickly with particle count. More importantly, they are snapshot measures and do not supply the flip history. We include them as a hole-counting comparison on the 2D datasets to measure what the alpha-complex choice buys in cost and stability.
Proposed Experiments
The first experiment validates the ledger on reference data. On 2D Taylor-Green vortex and reverse Poiseuille flow, both periodic, we compare accumulated flip counts from LagrangeBench reference trajectories with an independent stretching measure computed from the same trajectories. We expect a monotone relationship at fixed particle number. We also quantify how much frame subsampling inflates or undercounts flips relative to full-resolution reruns.
The second experiment is the decisive diagnostic check. We train GNS, SEGNN, and Neural SPH on all LagrangeBench datasets with matched budgets. We then compute rank correlations between each early-time metric, measured over the first 10% of a rollout, and long-horizon failure time, comparing flip-rate error, spurious-hole count, Sinkhorn divergence, position MSE, neighbor histograms, and cutoff turnover. The diagnostic claim stands only if flip-rate error beats all snapshot metrics and cutoff turnover on at least the periodic datasets.
The third experiment is the decisive correction check. GNS with the topology penalty is compared against plain GNS and Neural SPH on 2D lid-driven cavity and dam break, the bounded domains where clustering was worst. We report Sinkhorn divergence against the published $2.8\times10^{-7}$, kinetic-energy error, flip-rate error, and wall-clock inference cost. The 3D datasets test whether the 2-3-move ledger remains affordable.
As a stretch experiment, we will replace the accumulated flip count with a fading-memory version: past flips are summed with weights that decay over time, in the same way integral viscoelastic laws weight past strain. This version would be a candidate microstructure variable for the viscoelastic SPH settings of Santelli et al., but we will pursue it only if the first three experiments succeed.
References
- LagrangeBench: A Lagrangian Fluid Mechanics Benchmarking Suite
- Neural SPH: Improved Neural Modeling of Lagrangian Fluid Dynamics
- Symmetric Basis Convolutions for Learning Lagrangian Fluid Mechanics
- Integral fractional viscoelastic models in SPH: LAOS simulations versus experimental data
- Smoothed Particle Hydrodynamics simulations of integral multi-mode and fractional viscoelastic models
- Microstructural smoothed particle hydrodynamics model and simulations of discontinuous shear-thickening fluids
- The triangulation complexity of fibred3–manifolds
- Learning 3-manifold triangulations
- A unified view on the functorial nerve theorem and its variations