Shock waves are the hardest part of simulating gases, traffic, and multiphase flow. A smooth starting state can develop a jump in finite time, and at that jump the governing equations admit many mathematically valid solutions. Only one of them is physical: the entropy solution, defined as the limit of the viscous solution as the viscosity is driven to zero. Classical high-order schemes are built to find that solution. Examples are WENO (weighted essentially non-oscillatory reconstruction, which blends several local polynomial fits and down-weights any fit that straddles a jump). Recent neural solvers instead learn the numerical flux, the rule for how much quantity crosses each cell face. Unsupervised variants learn it only from a weak-form PDE residual, so they need no reference data. A residual loss, however, is satisfied by every weak solution, physical or not, and its optimization landscape becomes rugged as discontinuities sharpen.
This note proposes training such neural flux solvers along the very limit that defines the physical answer. The flux network is first fit to a strongly viscous version of the equation, where solutions are smooth and the loss is easy to optimize. It is then warm-started through a schedule of decreasing viscosity until the inviscid target is reached. The claim is that this ordering works as an optimization aid, as earlier PDE curricula did. It also works as a selection mechanism: it should steer the learned flux toward the entropy solution without hand-coded entropy penalties. That should matter most where direct training is fragile, namely wide stencils, non-convex fluxes, and coupled systems like the Euler equations.
Core Claim
Two existing results meet here without having been combined. (U)NFV replaces the analytic flux of a conservative finite-volume update with a small convolutional network. The update is $u_i^{n+1} = u_i^n - \frac{\Delta t}{\Delta x}(F_{i+1/2} - F_{i-1/2})$, where $u_i^n$ is the cell average and $F$ the learned face flux. The network is trained either on snapshots or on a weak-form residual, directly on discontinuous Riemann data, and the paper reports empirical convergence to entropy solutions on 1D scalar laws. Characterizing possible failure modes in PINNs shows that physics-informed losses degrade from a smooth bowl into a jagged landscape as convection strengthens. Warm-starting through increasing convection speed cut error on 1D convection at $\beta=30$ from near 100% to roughly 2%. That work never touched nonlinear shocks or conservative flux architectures.
Our claim is narrower and more testable than "curricula help shock capturing." For an unsupervised conservative neural flux, a vanishing-viscosity curriculum should do two things relative to direct training at equal compute. It should reduce the rate of non-entropic outcomes, such as expansion shocks and wrong compound waves on non-convex fluxes. It should also widen the range of stencil sizes and systems that train stably at all. The ordering itself, not mere exposure to viscous data, should be responsible.
Research Program
At stage $k$ the flux network $F_\theta$ is trained so that its finite-volume rollouts $u_\theta$ satisfy the viscous conservation law in weak form:
$$ \mathcal{L}_{\varepsilon_k}(\theta) = \sum_{\phi \in \Phi} \Big| \iint \big( u_\theta \, \phi_t + f(u_\theta)\, \phi_x - \varepsilon_k \, \partial_x u_\theta \, \phi_x \big)\, dx\, dt + \int u_0 \, \phi(x,0)\, dx \Big|^2 . $$
Here $f$ is the physical flux, $u_0$ the initial state, $\Phi$ a set of smooth compactly supported test functions, and $\varepsilon_k$ the stage viscosity. The learned flux must absorb the viscous term, so early stages teach it a dissipative, smooth face-flux rule. The final stage sets $\varepsilon = 0$, recovering the unsupervised (U)NFV objective exactly. The curriculum therefore changes only the path, never the endpoint.
The schedule starts at a viscosity spanning several cells and decreases geometrically. We test two pacing rules. The first is fixed-length stages. The second is a self-paced rule in the spirit of Self-Paced Learning: advance when the stage loss plateaus, which addresses the schedule-tuning burden flagged by Krishnapriyan et al.
We build this on the (U)NFV architecture for 1D Burgers and the LWR traffic family. We then extend it in two directions where (U)NFV left gaps. The first is the non-convex Buckley–Leverett flux, whose entropy solution is a shock attached to a rarefaction and where wrong weak solutions are easy to fit. The second is the 1D compressible Euler system, three coupled conserved quantities, with the flux network outputting a vector flux. A secondary analysis tests the "architecture versus optimization" disagreement. PIANN argues that architecture matters more than loss scheduling for hyperbolic problems, and sl-PINN relies on hand-built tanh correctors because plain PINNs collapse at low viscosity. To probe this, we run the same curriculum on a plain MLP PINN for viscous Burgers. That asks whether the curriculum alone rescues a weak architecture or whether the conservative flux structure is also needed.
Baselines
The naive reference is the first-order Godunov scheme, the simplest correct shock-capturing method. Any learned flux must clearly beat it to justify itself. (U)NFV reports "up to 10x lower error than the standard Godunov method." The file gives no absolute error value, so we will reproduce Godunov and (U)NFV on identical grids and Riemann test sets ourselves.
The classical high-order reference is WENO-5, implemented following the adaptive-stencil machinery of Shu's ENO implementation. Lichtlé et al. report that (U)NFV "consistently outperformed ENO/WENO methods and matched the accuracy of Discontinuous Galerkin (DG) solvers." WENO therefore sets the bar on Euler, where no neural flux result exists.
The strongest learned competitor is direct-training unsupervised (U)NFV at the same stencil sizes, network width, optimizer, total gradient steps, and hyperparameter search budget. For it we have only the relative Godunov comparison above. Beating it means one of two outcomes: lower $L^1$ error against exact Riemann solutions on the held-out multi-wave tests, or equal error with fewer non-entropic failures and stable training at stencil widths where direct training diverges or plateaus.
Three ablations isolate the key ingredient. A shuffled-mixture model sees the same set of viscosities, sampled uniformly throughout training rather than in decreasing order. A reversed curriculum runs from sharp to viscous. A direct-training model is given an explicit discrete entropy-inequality penalty. On viscous Burgers the external reference is sl-PINN, which reports an $L^2$ error of about $5.57 \times 10^{-3}$ at $\varepsilon = 10^{-3}/\pi$, against $5.50 \times 10^{-1}$ for a standard PINN (sl-PINN). On Buckley–Leverett, PIANN reports average absolute shock-front error below $10^{-3}$ (PIANN). We will match its grid of 101 points and 51 steps over mobility ratios $M \in [2,100]$.
Why not just…?
Why not just keep training directly on Riemann problems, since (U)NFV already converges to entropy solutions? This is the real threat. On convex 1D Burgers and LWR, the curriculum may add nothing. We expect any gain to appear where the file shows direct training has not been tested: non-convex fluxes, wide stencils, and coupled Euler. The equal-compute comparison against direct (U)NFV on all three settles it. If direct training matches us everywhere, the honest conclusion is that conservative flux structure alone already solves entropy selection, which would itself be a useful negative result for the Krishnapriyan-style optimization story.
Why not just mix all viscosities into one training set? Mixing gives the network the same smooth examples without ordering, and with that data diversity it might match the curriculum. The shuffled-mixture ablation tests exactly this. Only if ordering beats mixing at equal steps does the vanishing-viscosity interpretation hold.
Why not just add an explicit entropy-inequality penalty to the loss? A penalty directly targets non-entropic solutions and needs no schedule. Its drawbacks are that it requires an entropy pair for each equation, which is awkward for systems, and that it adds a weight to tune. We include it as an ablation. If it matches the curriculum on scalar laws but fails to transfer to Euler, the curriculum's value lies in needing no equation-specific entropy design.
Why not just train supervised NFV on WENO-generated data? Where a trusted high-order solver exists, supervised training sidesteps the selection problem entirely, and it may well win on accuracy. The case for unsupervised training rests on settings without a reference solver, such as noisy real data like I-24 MOTION. We will report supervised NFV as an upper reference and state plainly when it dominates.
Proposed Experiments
Training uses isolated Riemann problems, following (U)NFV. Evaluation uses held-out multi-shock and rarefaction interactions for Burgers and the six LWR variants, Buckley–Leverett across mobility ratios, and the Sod, Lax, and Shu–Osher problems for 1D Euler. Out-of-distribution checks use 1D Burgers and 1D compressible Navier–Stokes from PDEBench, scored with nRMSE, conservation error (cRMSE), and frequency-spectrum error (fRMSE).
Shock-specific metrics are $L^1$ error against exact solutions, shock width in cells, maximum overshoot, and sign violations of discrete entropy production. The headline failure count is the fraction of rarefaction-generating initial states that the model resolves as expansion shocks, together with the fraction of Buckley–Leverett cases with the wrong shock height. For Euler we also count runs producing negative density or pressure. Every configuration is run over at least five seeds and swept across spatial stencil widths up to the $NFV^{11}_{10}$ scale. The result of interest is the seed-wise rate of training failure versus stencil width for curriculum against direct training.
Three checks are decisive. The first is whether curriculum training lowers the entropy-violation rate on Buckley–Leverett relative to both direct training and the shuffled mixture. The second is whether it trains a stable vector flux for Euler where direct unsupervised training does not. The third is whether either effect survives when total compute is held fixed. If the answer to all three is no, the proposal's central claim is refuted.
References
- (U)NFV: (Un)Supervised Neural Finite Volume Methods for Solving Hyperbolic PDEs
- Characterizing possible failure modes in physics-informed neural networks
- Self-Paced Learning for Latent Variable Models
- Physics-informed attention-based neural network for hyperbolic partial differential equations: application to the Buckley–Leverett problem
- Singular layer PINN methods for Burgers' equation
- Efficient implementation of essentially non-oscillatory shock-capturing schemes
- PDEBench: An Extensive Benchmark for Scientific Machine Learning