Predicting Fields on Moving Domains Before the Data Arrives

Many systems we want to control are spread out in space, and their boundaries move. Examples include temperature in a deforming part, flow in a channel with a flexing wall, and electrical waves in a beating heart. Measurements of such systems reach the controller late because of sensing, network and compute delays. The standard fix for delay is a forward model, a simulator run in parallel with the real system that predicts where the system will be when the next action lands. Existing forward models do this for small state vectors such as joint angles or a drone's position. They do not handle a whole field whose domain changes shape during the delay. Fast field simulators for irregular shapes have the opposite gap. They work by mapping each shape onto a fixed reference square or disk, but they treat the shape as known and static, and they are run offline.

We propose a delay-compensating estimator that splits the state into two parts: a few numbers describing the shape, and a field stored on the fixed reference grid. The estimator predicts both forward over the measured delay. The key observation is a change of variables. When a moving domain is pulled back to a fixed reference grid, the boundary motion becomes an extra transport term whose speed is set by how fast the shape is changing. Field prediction on a moving domain therefore splits into two problems: a cheap low-dimensional forecast of the shape, and a field forecast on a grid that never moves. We expect the error of today's methods to concentrate in a band near the moving boundary and to grow with shape speed times delay, and we expect this split to remove most of that error. The claim is precise enough that it could turn out false.

Core Claim

Let $u(x,t)$ be a scalar field on a physical domain $\Omega(t)$ governed by $\partial_t u = F(u,\nabla u)$. Let $\Phi_{z(t)}:\Omega(t)\to\Omega_r$ be a map onto a fixed reference domain $\Omega_r$, determined by low-dimensional shape coefficients $z(t)$. Define the pulled-back field $\tilde u(\xi,t) = u(\Phi_{z(t)}^{-1}(\xi),t)$. The chain rule then gives

$$\partial_t \tilde u = \tilde F(\tilde u;\, J_z) + \nabla_x u \cdot \frac{d}{dt}\Phi_{z(t)}^{-1}(\xi),$$

where $\tilde F$ is the original operator rewritten with the map's Jacobian $J_z$. The last term is a "mesh velocity" transport that is nonzero only when the shape moves. Two consequences follow.

The first concerns geometry. Any estimator that propagates the field over a delay $\tau$ with stale geometry, meaning it holds $z$ fixed at its last measured value, makes an error that to first order scales like $\tau\,\|\dot z\|\,\|\nabla u\|$. That error is concentrated where shape changes move points the most, typically near the boundary.

The second concerns structure. If $z$ is forecast explicitly, the field forecaster only has to learn dynamics on a fixed grid, with the transport term supplied in closed form from $\dot z$.

Our claim is that this split predictor beats both masked Cartesian forecasters and canonical-grid surrogates with stale geometry. The advantage should appear specifically near boundaries and should grow with shape speed and delay. On fixed domains, where $\dot z = 0$, the advantage should vanish.

Relation to Prior Work

The geometric side supplies the method. PhyGeoNet maps vascular-like channels to a rectangle with elliptic coordinates and solves steady PDEs there with convolutions. DIMON learns solution operators on a reference geometry by pulling each shape back through a diffeomorphism (a smooth, invertible warp), with shapes parameterized by PCA of deformation fields. DIMON reports 3.2–7.5% relative $L_2$ error on time-dependent reaction-diffusion over ellipses with a hole and an ~18,000x speedup over finite-element cardiac simulation.

We reuse DIMON's PCA shape coding because it makes evaluating the map cheap, and cheap evaluation matters when the whole computation must fit inside the delay. We depart from DIMON in three ways. Our shapes move during the prediction horizon. The shape trajectory itself is forecast rather than registered after the fact. The model also runs online and ingests delayed sparse measurements.

NUNO instead splits irregular domains into local uniform patches with a static K-D tree. It is evidence that grid-based models beat global interpolation, with a 61% error reduction on (2+1)D channel flow. Its static partition is exactly what breaks when the boundary moves, so we treat it as a reference point rather than the base. Gauge Flow Matching maps convex sets to a unit ball for constrained generation. It shows that reference-space mappings pay off for speed and feasibility, but its sets are static and it does no estimation. We do not build on it.

The estimation side supplies the reset-and-rollout structure. Event-triggered predictor feedback resets an internal predictor whenever a delayed packet arrives and proves stability for low-dimensional nonlinear plants under time-varying delay. Asynchronous MPC on the Minitaur rolls a learned model forward by the planning latency. ACDA handles stochastic, unobserved delays with a recurrent belief model and provides real WiFi delay traces (DLib, DOff). All three motivate predictor-based compensation, and all three assume a lumped state vector with no spatial field and no moving boundary. We transplant the rewind–assimilate–reroll pattern to distributed fields.

One paper constrains the design: One Operator to Rule Them All?. It shows that neural PDE solvers conditioned on boundary data learn a family of boundary-indexed operators. Their relative $L_2$ error rises from about 0.07–0.1 in distribution to 0.4–0.6 under boundary shift. A canonical-grid model conditioned on $z$ could fall into the same trap and become a shape-indexed lookup table. For this reason we test out-of-distribution shape trajectories explicitly. We also supply the transport term analytically rather than asking the network to infer it from $z$ alone.

Research Program

The estimator has three components.

Shape forecaster. A small recurrent network forecasts $z(t)$ over the delay horizon from past shape observations. Shapes are encoded as PCA coefficients of the deformation from a template, as in DIMON. The forecaster returns $z(t+\tau)$ and $\dot z$ along the way.

Canonical field forecaster. A Fourier neural operator, a network that learns PDE solution maps by working in the frequency domain on a regular grid, runs on $\Omega_r$. It takes the current canonical field, the Jacobian fields $J_z$, and the analytic transport term computed from $\dot z$, and steps forward. The Fourier neural operator is chosen because PDEBench finds it generally strongest on grid forecasting.

Delayed assimilation. When a measurement packet stamped at time $t-\psi$ arrives, the system rewinds to $t-\psi$. It pulls the sparse sensor readings back to $\Omega_r$ using the shape at the stamp time, not the current shape. It then corrects the field with a learned encoder and rerolls both forecasters forward to $t+\tau_{\text{act}}$. This mirrors the predictor reset of the event-triggered literature.

We restrict to scalar fields because DIMON's open problem of transporting vectors through the map is real and orthogonal to our question. We handle flow through 2D vorticity, which is a scalar.

The analysis has three parts. We check the first-order error prediction for stale geometry against measured error as $\tau$ and shape speed vary. We test whether a closed-form transport term helps over letting the network learn it. We also measure wall-clock time per reset, since an estimator slower than the delay it is meant to cancel is useless.

Data and Benchmarks

No existing benchmark couples moving domains with delayed observations, so we build two.

Moving-ellipse reaction-diffusion. This extends DIMON's ellipse-with-a-hole family. Shape coefficients follow smooth random trajectories with controllable speed, and a moving-mesh finite-element solver provides ground truth.

Flexible-wall channel. This is 2D incompressible flow in a channel whose wall shape follows a random low-mode trajectory, tracked through its vorticity.

Sensors are sparse points on and near the boundary. Delays come from three sources: constant values, a synthetic random process, and ACDA's DLib and DOff WiFi traces rescaled to the simulation timestep.

A third, fixed-domain control set is drawn from PDEBench 2D diffusion-reaction. On it the shape never moves, so the method should tie its baselines. A gain there would mean our gains come from something other than the mechanism we claim.

The primary metric is relative $L_2$ error of the delay-compensated field at actuation time. We report it over the whole domain and, separately, in a boundary band of fixed width in reference coordinates, binned by delay and by shape speed. Secondary metrics are PDEBench's boundary error (bRMSE), shape-forecast error, and wall-clock per reset against the delay distribution.

A modest closed-loop test asks whether better estimates matter downstream. In that test, boundary heating drives the reaction-diffusion field toward a target under DLib delays, using the same controller on top of each estimator, and we report tracking error.

Baselines

Zero-order hold. The naive reference passes the last assimilated field, warped to the current geometry, straight to the controller with no forward prediction. It fixes the scale of how much delay hurts at all. A method that does not clearly beat it at the median DLib delay has nothing to offer.

Masked Cartesian FNO. The strongest generic competitor rasterizes the moving domain onto a fine Cartesian grid with a binary mask. A Fourier neural operator conditioned on $\tau$ is trained to output the future masked field from the delayed measurements and recent masks. It is the strongest generic option because PDEBench reports FNO dominance on grid forecasting and because it needs no mapping machinery. PhyGeoNet's 85% versus 34% pressure-error comparison is against a fully connected PINN, not this model, so it does not settle the question. We give this baseline a grid fine enough to match our parameter count and wall-clock within 10%.

Stale-geometry canonical surrogate. The strongest existing method is a DIMON-style canonical operator. It uses identical architecture, data and shape coding, rolls the field forward on $\Omega_r$, and holds the shape at its last measured value. DIMON reports 3.2–7.5% relative $L_2$ on its static reaction-diffusion family. We found no reported result on Lacuna for any method doing delay-compensated field estimation on moving domains. We will therefore first reproduce DIMON-like accuracy on our generator with $\tau=0$ to confirm the comparison is sound. This baseline isolates our key ingredient, the forecasting of geometry over the delay.

Learned-transport ablation. This removes the analytic transport term and lets the network infer motion effects from the forecast $z$ trajectory. It tells us whether the change of variables matters, or only having the future shape.

Mesh GNN. A message-passing network on the moving mesh is included because it represents the mesh-free school of thought.

Tuning and fairness. All methods get the same simulations, the same delay traces and the same 30-trial hyperparameter budget, with five seeds each.

What counts as winning. We count it as beating the strongest baseline if our boundary-band error at the 90th-percentile DLib delay is lower than the stale-geometry surrogate's by more than the spread across seeds. The gap must also widen with shape speed, and per-reset wall-clock must stay under the median delay.

Why not just…?

Why not just rasterize the domain and let a masked FNO learn the motion? This is the most serious threat. A large enough Cartesian model might learn boundary motion implicitly. The evidence against it is indirect. PhyGeoNet and NUNO show that staircase boundaries and global gridding cost accuracy near boundaries, but only on static domains. Our boundary-band metric binned by shape speed settles the question directly. If the masked FNO matches us in that band at high speeds and equal compute, the reference grid is not earning its keep. We would report that as a negative result.

Why not just run an existing canonical surrogate like DIMON with the latest available shape? It is the cheapest upgrade, and on slowly moving shapes it may be enough. The first-order error estimate predicts its error grows with $\tau\|\dot z\|$. Our experiment plots error against that product. If the stale surrogate's curve is flat, geometry forecasting is unnecessary and the proposal collapses to "use DIMON online." The fixed-domain control shows what tying looks like.

Why not just use a graph network on the moving mesh and skip mappings entirely? Graph models handle moving meshes natively. NUNO reports graph neural operators training 2x to 30x slower, and PointNet reaching about 56% error versus about 5% on 3D heatsinks. Training speed is not latency, though, so we measure per-reset inference time against the delay distribution rather than assume it. If the graph model is both accurate and fast enough, our advantage reduces to engineering convenience, and we will say so.

Decisive Checks

Four results would establish or refute the idea. First, the boundary-band error gap between our method and the stale-geometry surrogate should grow roughly linearly in $\tau\|\dot z\|$, as the change of variables predicts. Second, that gap should vanish on the fixed PDEBench control. Third, under shape trajectories faster than any seen in training, removing the analytic transport term should hurt more than it does in distribution. That would be the boundary-indexed failure of One Operator to Rule Them All? reappearing in learned form, and the closed-form term guarding against it. Fourth, the closed-loop heating task should show lower tracking error with our estimator under the same controller and DLib delays. If the first check fails, the mechanism is wrong regardless of headline accuracy.