Many one-dimensional (1D) flow models used for rivers, pipe networks and blood vessels contain terms that cannot be written as the change in a conserved quantity. These are called non-conservative products. When the solution jumps, for example where a channel suddenly narrows, where pipes meet, or where a barge moves along a canal, such a term multiplies a jump by a spike. That product has no defined value, so the model cannot say which jump is correct. The standard fix, a "path" that specifies how states are joined inside the jump, is chosen by convention. Straight-line paths are known to converge to wrong answers (Exact solution for Riemann problems of the shear shallow water model). Junctions in networks have a parallel problem: they need an extra rule, usually "equal pressure in every branch," chosen because it is physically plausible (Coupling conditions for gas networks governed by the isothermal Euler equations). Moving obstacles need a third rule, a capacity limit that produces jumps classical theory forbids.
This note argues that for a large class of these problems all three rules are one object, and that it can be measured. The ambiguity exists because a 1D model has collapsed a real two-dimensional (2D) geometry into a point. The physics that selects the right jump is therefore not an unknown vanishing viscosity. It is the flow in the collapsed dimension, and a well-posed, conservative 2D model can compute it. We propose to resolve that geometry in 2D, extract the effective 1D jump relation, and check whether it fits a path family that is cheap to evaluate and consistent across static width steps, branching junctions and moving obstacles. We will also test when equal pressure is the correct limit of that relation.
Core Claim
Take the width-averaged shallow-water equations, the standard 1D river model. The pressure term picks up a factor proportional to the derivative of the channel width. Where the width jumps, this term is a non-conservative product. A moving partial blockage, such as a barge, is then a moving width jump. In 1D it behaves exactly like the moving bottleneck of traffic models: it caps the discharge relative to the obstacle and can force a downward jump that classical entropy rules call inadmissible. Such a jump is called a non-classical shock (A conservative scheme for non-classical solutions to a strongly coupled PDE-ODE problem). At a channel junction, the common closure of equal water level is the shallow-water version of the equal-pressure condition.
Our claim has three parts. First, a single low-dimensional path family, fitted from 2D reference flows, reproduces the 2D-averaged jump states at static steps, junctions and moving obstacles better than the straight-line path, equal level, or energy conservation. Second, it does so without being refitted for each configuration type. Third, the error of equal pressure grows predictably with flow speed, roughly as
$$\frac{|h_i - h_j|}{h} \sim C\,\mathrm{Fr}^2,$$
where $h_i, h_j$ are depths at the junction faces of branches $i,j$ and $\mathrm{Fr}$ is the Froude number, the ratio of flow speed to wave speed. The $\mathrm{Fr}^2$ form comes from comparing dynamic pressure, which scales like $u^2/2$, with hydrostatic head $gh$. The gas analog replaces $\mathrm{Fr}$ by the Mach number $M$. This would explain why equal pressure works so well in the low-Mach regime studied by An Asymptotic-Preserving Scheme for Isentropic Flow in Pipe Networks, and predict where it breaks.
Relation to Prior Work
Two papers motivate the claim. The exact shear-shallow-water Riemann solver of Exact solution for Riemann problems of the shear shallow water model shows that straight paths give unphysical jumps. It also shows that schemes miss the exact answer even at 2,000 cells, because numerical errors create dispersive wave trains. The paper names vanishing small-scale physics as the true selector but treats it as intractable. We replace that selector with resolved geometry in the cases where geometry is the cause. Numerical schemes for coupled systems of nonconservative hyperbolic equations defines path-conservative Kirchhoff conditions at a single static interface using a first-order relaxation scheme, still with straight paths. We keep its idea that coupling is a path statement and extend it to multi-branch and moving interfaces, with paths taken from data rather than chosen.
Several papers constrain the design. A central scheme for coupled hyperbolic systems shows that naive linear coupling of relaxation variables creates artifacts, so any fitted closure must be applied to physical states. It must not be applied to auxiliary variables. Interacting moving bottlenecks in traffic flow reports spurious oscillations when constraints switch on and off. Our moving-obstacle tests therefore include activation and deactivation. Banda et al. show equal pressure beats a subsonic bound at a 1-to-1 joint. That is the closure we test against, not one we assume.
Other papers supply methods. The in-cell reconstruction of Chalons et al. keeps the moving jump sharp on a fixed grid. Well-balanced path-conservative central-upwind schemes from Novel Local Characteristic Decomposition Based Path-Conservative Central-Upwind Schemes are both the 1D solver and the 2D reference solver on Cartesian grids with blocked cells. The entropy-stable treatment of Entropy stable schemes for the shear shallow water model Equations shows a non-conservative term can be entropy-neutral. We will test whether fitted paths dissipate or produce energy.
Research Program
The work has four stages.
Reference data. We build a 2D shallow-water library on Cartesian grids with blocked cells. It covers abrupt contractions and expansions over a range of width ratios, symmetric and side-entry T- and Y-junctions, and a rectangular obstacle of blockage fraction $\beta$ moving at speed $V$. Each case runs Riemann-type initial data plus slow, quasi-steady sweeps. We average the 2D solutions across the width at stations a few channel widths from the feature, so the 1D states are measured outside the local 2D region. The quantity extracted is the jump map: the incoming 1D characteristic data in, and the face states and discharges out.
Closure families. We fit nested families in order of complexity. The first is equal level or equal pressure. The second is Bernoulli energy conservation. The third is momentum balance with a constant loss coefficient. The fourth is a DLM path, that is, a path in the Dal Maso–LeFloch–Murat sense, deformed from the straight line by a two-to-four-parameter function of the local Froude number and geometry ratio. The last is a small monotone neural map. Every family enforces exact mass conservation and requires the outgoing characteristic count to match the incoming data. A closure that fails these checks is reported as inadmissible rather than kept.
1D solver. We embed the selected path in a well-balanced path-conservative central-upwind scheme on a network graph. At moving obstacles we add in-cell reconstruction, which keeps the width jump inside one cell.
Gas analog. We repeat the junction study for 2D isentropic Euler flow in a T-junction across a Mach sweep, to test the $M^2$ prediction for equal pressure.
Baselines
The first baseline is the naive reference: a straight-line path for the width term with no special junction rule. It is the default in path-conservative codes. Beating it is necessary but not impressive, because Exact solution for Riemann problems of the shear shallow water model already shows straight paths can be wrong.
The strongest existing competitor at junctions is equal pressure (equal level), the closure validated by Coupling conditions for gas networks governed by the isothermal Euler equations. It is also used with mass conservation at 1-to-2 and 2-to-1 junctions by An Asymptotic-Preserving Scheme for Isentropic Flow in Pipe Networks. Banda et al. report qualitatively that only equal pressure "prevented a non-physical pressure jump," and Redle and Herty report first-order accuracy in time and second-order in space. No reported result was found on Lacuna for junction-state error against a resolved 2D reference. Our 2D library therefore supplies that number for every baseline under identical conditions.
The strongest competitor for moving obstacles is the in-cell reconstruction scheme of Chalons et al. with a constant capacity cap. It reports convergence between $O(\Delta x^{0.8})$ and $O(\Delta x^{1.5})$ and exact mass conservation in the 1D traffic setting.
All 1D methods share the same grid, the same central-upwind core, and the same time step and CFL number, where CFL is the time-step limit set by wave speed. They also get the same fitting budget: each closure's parameters are fitted on one training split of the 2D library and scored on held-out geometries and Froude ranges. The proposal wins if the fitted path cuts width-averaged face-state error relative to equal pressure by a clear margin, for example half, at $\mathrm{Fr} \gtrsim 0.3$. It must stay no worse at low $\mathrm{Fr}$, and the same parameters must work across all three configuration types.
Two ablations isolate the key ingredient, which is geometry-derived selection. In the first, the same flexible path family is fitted to 1D data regularized by artificial viscosity instead of 2D data. If it does as well, the collapsed dimension is not the selector. In the second, closures are fitted on steady data only and tested on transients.
Why not just…?
Why not just use a hydraulic loss coefficient? Engineering practice adds a constant minor-loss term to Bernoulli or momentum balance. This is a real threat, and it sits inside our nested families. If a constant coefficient matches the 2D data across geometries, Froude numbers and moving cases, we will say so. The contribution then shrinks to two things: deriving the coefficient from first principles as a path, and showing it carries over to moving non-classical jumps. We expect a constant coefficient to fail in transients and at moving obstacles, where the effective loss depends on obstacle speed. The held-out transient and moving tests settle this.
Why not just simulate everything in 2D? On a large network this costs far more, and our 2D reference uses Cartesian blocked-cell grids. Entropy stable schemes for the shear shallow water model Equations notes that unstructured grids for complex geometry remain open. We will report wall-clock time per simulated hour for both models on the same hardware. The 1D closure only earns its place if it is much cheaper at comparable accuracy.
Why not just derive paths from vanishing viscosity? For geometric non-conservative terms, the unresolved structure is lateral flow separation and recirculation, not a thin viscous shock layer. The first ablation tests this directly. Nkonga and Chandrashekar's dispersive wave trains also show that numerical regularization does not converge to a single physical answer.
Why not just learn the whole flux with a neural finite-volume scheme? (U)NFV reports up to 10x lower error than Godunov on I-24 traffic data, but only for 1D scalar laws without junctions or moving constraints. An end-to-end learned flux would not guarantee mass conservation at junctions, and it would not expose the closure for inspection. We include the small neural jump map as the most flexible family and an upper bound. If it is much better than the low-parameter path, that says the physics has more structure than our path family captures.
Proposed Experiments
The first decisive check is the Froude sweep at T- and Y-junctions. It tests whether equal-pressure error scales as $\mathrm{Fr}^2$, using a log-log slope fitted over $\mathrm{Fr} \in [0.05, 0.8]$. The gas T-junction Mach sweep is the parallel test, and we check it against the low-Mach success reported by Redle and Herty. If the slope is far from 2, our explanation of why equal pressure works is wrong.
The second check is transfer. A path fitted only on static width steps is applied unchanged to junctions and to the moving-obstacle cases. It is scored on width-averaged depth and discharge error ($L^1$ over a window of several channel widths), on upstream surge height, and on the size of the downstream drop behind the obstacle. Failure to transfer would show that the three ambiguities are not one object, which is the central claim.
The third check repeats Banda et al.'s variable-sound-speed joint ($a_1=6, a_2=8$) as a sanity test of the gas pipeline. It also repeats the Chalons et al. Riemann suite (Cases 0 to IV) with the moving-width model, confirming that sharp jumps and their convergence rate survive the change of path.
The fourth check covers obstacle activation and deactivation, as a barge enters and leaves a narrow reach. We measure oscillation amplitude against the in-cell baseline, in light of the oscillations reported by Goatin et al.
The fifth check computes the energy produced across each fitted jump. Following the entropy-neutrality analysis of Yadav et al., a physically selected path should never create energy. A fitted closure that does is evidence of overfitting, and we discard it.
References
- Exact solution for Riemann problems of the shear shallow water model
- Coupling conditions for gas networks governed by the isothermal Euler equations
- A conservative scheme for non-classical solutions to a strongly coupled PDE-ODE problem
- An Asymptotic-Preserving Scheme for Isentropic Flow in Pipe Networks
- Numerical Schemes for Coupled Systems of Nonconservative Hyperbolic Equations
- A Central Scheme for Two Coupled Hyperbolic Systems
- Interacting moving bottlenecks in traffic flow
- Novel local characteristic decomposition based path-conservative central-upwind schemes
- Entropy Stable Schemes for the Shear Shallow Water Model Equations
- (U)NFV: (Un)Supervised Neural Finite Volume Methods for Solving Hyperbolic PDEs