Large-Step Neural Solvers Whose Receptive Field Is a Hyperbolic Ball Sized by the Courant Number

Fast simulators for waves, shocks and fluid flow are limited by the Courant-Friedrichs-Lewy (CFL) condition: in one time step, a grid cell may only depend on cells that a physical wave could reach in that time. Explicit solvers therefore take tiny steps. Neural surrogates promise to skip ahead, and recent work shows they can take steps many times larger than the classical limit while still conserving mass. The catch is that a large step needs a large neighborhood. In the best current method, each cell's update reads a fixed local window of neighbors. That window must grow in proportion to the step size and to the grid resolution. As a result, cost grows linearly in the Courant number, and a model trained at one resolution does not transfer to another.

We propose to replace the fixed window with a multiscale one that has a natural geometry. A stack of grids from fine to coarse, indexed by position $x$ and cell size $s$, is exactly the hyperbolic upper half-plane, a negatively curved space in which horizontal distance is measured in units of the local cell size. In that geometry the Courant number is itself a hyperbolic distance. The physical domain of dependence of a step of length $\Delta t$ is a hyperbolic ball of fixed radius, centered at scale $s \approx \lambda \Delta t$, where $\lambda$ is the local wave speed. A cell's update for an arbitrarily large step can then read a constant number of multiscale tokens: the ones inside that ball. Conservation is kept exactly by feeding those tokens into face-transport heads that cannot create or destroy mass.

Core Claim

The upper half-plane carries the metric $ds_H^2 = (dx^2 + ds^2)/s^2$. Here $x$ is spatial position and $s>0$ is the cell size of a level in the grid hierarchy. The geodesic distance between two hierarchy nodes $(x_1,s_1)$ and $(x_2,s_2)$ is

$$d_H = \operatorname{arcosh}\!\left(1 + \frac{(x_1-x_2)^2 + (s_1-s_2)^2}{2 s_1 s_2}\right).$$

For two nodes on the same level $s$, this reduces to $d_H = 2\,\operatorname{arsinh}(C/2)$, where $C = |x_1-x_2|/s$ is the separation counted in cells of that level. A wave travelling for time $\Delta t$ covers $C = \lambda\Delta t/s$ cells, which is the local Courant number of level $s$.

The central observation follows directly. A step that violates CFL on the fine grid ($\lambda\Delta t \gg \Delta x$) respects CFL on every level with $s \gtrsim \lambda\Delta t$. The set of hierarchy nodes from which a fine cell may legitimately draw information is therefore well approximated by a hyperbolic ball of radius $r = O(1)$ around the node $(x, \lambda\Delta t)$. This ball contains $O(\log(\lambda\Delta t/\Delta x))$ levels with $O(1)$ nodes each, because hyperbolic balls of fixed radius have fixed volume and each level contributes a bounded number of nodes.

Our claim is empirical and falsifiable. A conservative surrogate whose face transport is computed by attention restricted to this physically centered ball will do three things. It will match a fixed-stencil conservative surrogate at moderate Courant numbers. It will keep accuracy at Courant numbers where the fixed stencil must be widened. And it will transfer across grid resolutions without retraining, while the work per cell grows only logarithmically in the step size.

Three points separate this from the closest work. First, the hyperbolic coordinates are not learned. They are the analytic coordinates of the grid's own scale hierarchy. Second, the attention region is set by the physics ($\lambda\Delta t$), not by the network. Third, the method is a time-stepping rollout with exact discrete conservation, not a steady-state map.

Relation to Prior Work

The proposal reuses its conservation machinery from FluxNet. FluxNet predicts cumulative transport across faces over a whole step using capacity-constrained heads. It keeps discrete conservation near $10^{-8}$ and stays stable at 5x the classical step on the Lighthill-Whitham-Richards (LWR) traffic model. Its authors also state the limitation that motivates us: the stencil is fixed and the model is not resolution invariant. We keep FluxNet's heads unchanged and change only how their inputs are gathered.

ShockCast supplies the other half of the physics. It predicts an allowable step from local wave speeds and conditions the solver on it. We borrow that wave-speed estimate to place the ball, but we use it to choose where to look rather than how far to step.

The geometric precedent is the Hyperbolic Neural Operator. It embeds tokens in the Lorentz hyperboloid and lets learned radii separate local specialists from global aggregators. That gives a 42.2% relative-error reduction on Elasticity, and accuracy collapses when the learned radii are fixed. Its finding supports our premise that scale-as-radius matters. However, the operator is steady-state, has no time integrator and does not conserve mass. We test whether analytic, physics-placed radii can replace learned ones in a time-dependent setting.

Two lines of work organize multiscale structure in flat space. MiSe-GNN builds hierarchical trees around high-error nodes in a flat graph. HOT uses a Euclidean multigrid hierarchy to reach CFL-rate implicit steps in the material point method. Both suggest that hierarchies, rather than wider stencils, are the right tool for large steps. Neither supplies a metric that ties the hierarchy to the domain of dependence.

The foundational embedding results, Poincaré embeddings and the Lorentz model, establish that tree hierarchies embed with low distortion only in negatively curved space. We use the Lorentz model for numerically stable distance computations. One known objection is that hyperbolic space distorts flat physical systems. It does not apply here, because we embed the scale hierarchy of the grid, which is exactly hyperbolic, and not the flat grid itself.

Several papers sound related but address a different object. Hyperbolic Entailment Cones model partial orders between concepts, not spatial causality. HGDE runs a diffusion process along network depth on a hyperbolic manifold. The Hyperbolic-PDE GNN uses "hyperbolic" only in the PDE-classification sense. None of them treats the physical domain of dependence.

Research Program

We first build the multiscale token set. For a 1D or 2D Cartesian grid, we form a dyadic hierarchy of cell averages down to the coarsest level. Each node receives fixed upper-half-space coordinates (cell center, cell size), mapped to the Lorentz model.

For each fine face at each step, we do four things. We estimate the local wave speed $\hat\lambda$, either from the solution's flux Jacobian when it is known, or with a small ShockCast-style regressor. We center a ball of radius $r$ at $(x_{\text{face}}, \hat\lambda \Delta t)$. We gather the $K$ nodes inside it, with $K$ bounded by construction. We then attend over them with a kernel $\exp(-d_H/\tau)$, where $\tau$ is a learned temperature.

The attended features feed FluxNet's capacity-constrained transport heads, which produce the cumulative mass crossing that face over $\Delta t$. The cell update is the divergence of these face transports. Conservation is therefore exact by construction, and positivity bounds are inherited from the heads. The step $\Delta t$ is an input, so one model serves a range of Courant numbers.

For parabolic problems, the dependence width scales as $\sqrt{D\Delta t}$, where $D$ is the diffusivity, rather than $\lambda\Delta t$. We expose this as a second centering rule and test whether the correct rule matters.

The analysis component is short. We will show that for a linear advection problem the ball of radius $r$ covers the exact characteristic interval at all Courant numbers. We will also bound $K$ as a function of $r$. This turns the receptive-field guarantee that ShockCast lacks into a stated geometric property, conditional on the wave-speed estimate being an upper bound.

Baselines

The naive reference point is a FluxNet model run with its own heads and stencil at the classical CFL step, unrolled for as many substeps as needed to cover the target $\Delta t$. Every large-step method must beat it in wall-clock cost at equal error. If it does not, the large-step idea itself is not paying off.

The strongest existing competitor is FluxNet trained directly at large steps. Its stencil width will be tuned per Courant number with the same hyperparameter budget, data, and GPU hours as our model. It is the only method in the file that is simultaneously conservative to machine precision and stable beyond the explicit limit. The file reports that it cuts shallow-water depth error by 81% relative to the best post-hoc correction methods. It also reports a 17.3x wall-clock speedup on Cahn–Hilliard from 1000x larger steps. No reported absolute relative-$\ell_2$ value for FluxNet was found on Lacuna, so we will rerun it ourselves.

Beating FluxNet means three things. Error must stay within noise of FluxNet at its own operating Courant number. Error must be lower at Courant numbers of 10 and above at matched FLOPs. And error must be lower when evaluated at 2x and 4x the training resolution.

We also compare against ShockCast on its circular-blast benchmark. There the file reports mass conservation within 0.2% over 100+ unrolled steps. Our method should conserve to near machine precision while matching its field error.

Three ablations isolate the key ingredient. The first, a flat multiscale variant, uses the same tokens and the same FluxNet heads but a Euclidean distance on $(x, \log s)$. This variant does not rescale horizontal distance by the local cell size, so it tests whether the hyperbolic metric itself matters. The second, a learned-radius variant in the style of the Hyperbolic Neural Operator, lets the network choose ball centers freely, with no physical placement. The third, a mis-centered variant, deliberately places the ball at $s = \Delta x$ regardless of step. It should fail at large Courant numbers if the domain-of-dependence argument is right.

Why not just…?

Why not just widen FluxNet's stencil? This works, but its cost grows linearly in the Courant number in 1D and quadratically in 2D. Its learned weights are also tied to a specific cell count, which is the resolution-invariance limitation the FluxNet authors name. Our scaling experiment measures exactly where the widened stencil's cost and cross-resolution error overtake ours. If that crossover never happens on the benchmarks below, the proposal fails.

Why not just use a Euclidean multiscale network, such as a U-Net or a multigrid hierarchy, with conservative heads? This is the real threat, and it is why the flat-multiscale ablation exists. Multigrid in the spirit of HOT and hierarchical rewiring as in MiSe-GNN both show that hierarchies help. What they lack is a rule for which coarse nodes a given fine face should read for a given $\Delta t$. A U-Net reads all of them with weights tuned to the training step.

If the flat variant with a hand-placed window matches us, the honest conclusion is that physics-placed multiscale gathering is the contribution and the hyperbolic metric is merely a convenient parameterization. We would report it that way. Our prediction is that the flat variant degrades when Courant number or resolution changes at test time, because only the hyperbolic metric measures separation in local cells.

Why not just take more small steps with a cheap surrogate, or use an implicit solver? Small steps accumulate autoregressive error, and their cost is linear in the number of steps. Implicit and optimization-based solvers such as PDHG are unconditionally stable. However, they can face non-convex landscapes on nonlinear conservation laws, and they still pay roughly 90–100 iterations per step. The naive baseline and a classical implicit reference on the transport tests quantify this directly.

Why not just use HNO with FluxNet heads? This is the obvious combination. The learned-radius ablation is exactly this model. HNO's own ablation shows radii matter, but learned radii have no reason to track $\lambda\Delta t$ when the step is changed at test time. We expect it to match us in-distribution and to fall behind on held-out Courant numbers.

Proposed Experiments

The primary suite is the FluxNet benchmark set: 1D convection–diffusion, 2D shallow water with wetting and drying, 1D LWR traffic and 2D Cahn–Hilliard. We add the linear-transport and traffic-shock tests used for PDHG as clean diagnostics with exact characteristics. We also use the circular-blast case from ShockCast, and the shallow-water and Euler-type datasets from The Well via polymathicai/the_well. Metrics are relative $\ell_2$ error over rollouts, discrete mass error, shock-position error, steps until divergence, and FLOPs and wall-clock time per unit of simulated time.

The first decisive check is a Courant sweep on linear transport and LWR from $C=1$ to $C=64$ at fixed resolution. It plots error and work per cell for each method. We predict our work grows logarithmically while widened FluxNet grows linearly. The mis-centered ablation should fail beyond $C\approx 2$.

The second check trains at one resolution and tests at 2x and 4x with the same physical $\Delta t$, so the fine-grid Courant number doubles and quadruples. Our model should hold error roughly flat. FluxNet should require retraining.

The third check runs the parabolic Cahn–Hilliard case under both centering rules. Accuracy should favor the $\sqrt{D\Delta t}$ rule. That would confirm the ball tracks the true domain of dependence rather than merely adding capacity.

Throughout, mass error must remain at FluxNet's machine-precision level. Any loss of conservation would mean the multiscale gathering broke the divergence form and would invalidate the comparison.