Simulating fluids or waves on a fine grid is expensive, and machine-learned surrogates (fast neural networks trained to imitate a simulator) need many fine simulations to become accurate. Two independent observations suggest this cost can be cut sharply. First, a cheap coarse-grid simulator can be run densely and still carries most of the physics. Pre-training a surrogate on coarse data and fine-tuning on a little fine data already works well. Second, consecutive time steps of a simulation are highly redundant. Paying the fine solver for every step of every trajectory wastes budget on steps that teach the model almost nothing. This note joins the two. We run the coarse solver at every time step of every trajectory, and call the fine solver only at the few steps where the coarse answer is likely to be badly and unpredictably wrong.
The bet is that the gap between coarse and fine solutions is far more redundant in time than the solution itself. A vortex street or a travelling soliton changes every step. The coarse grid's error, though, tends to stay steady and smooth between short bursts: vortex detachment, shock formation, soliton collisions. If that holds, a model that learns only the coarse-to-fine correction can be trained from a small set of fine-solver steps placed at those bursts, and the coarse solver anchors everything in between. This also repairs a known failure of current time-step selection methods. Today, the inputs handed to the expensive solver come from a neural surrogate and can be unphysical. Here they would come from a real (coarse) physical solver.
Core Claim
The closest existing work is Active Learning with Selective Time-Step Acquisition for PDEs (STAP). STAP builds "hybrid rollouts": a learned surrogate fills most time steps, a single numerical solver is queried only at a selected mask of steps, and the mask is chosen to minimize disagreement across a committee of surrogates. STAP motivates this proposal because it shows that time steps carry very unequal value; on Kuramoto–Sivashinsky it was the only method to significantly beat random sampling. It also constrains the design in three ways. It assumes one solver and one fidelity. It builds its surrogate from scratch. And its surrogate-generated states can crash the solver: on the KdV equation this forced an ad-hoc filter.
Our claim is that a three-tier rollout outperforms STAP's two-tier rollout at equal fine-solver cost. The three tiers are a dense coarse solver, a learned correction, and sparse fine-solver steps. The gain comes because acquisition targets the learned correction rather than the full state. Concretely, let $u^L_t$ be the coarse-solver state at step $t$ and $P$ an upsampling to the fine grid. Let $F_H$ be one step of the fine solver. The model $G_\theta$ predicts the next fine state from the current fine estimate and the coarse trajectory, written as a correction:
$$\hat u^H_{t+1} = P(u^L_{t+1}) + G_\theta\big(\hat u^H_t,\, u^L_{t:t+1}\big).$$
A fine-solver query at step $t$ returns the training pair $(\hat u^H_t, F_H(\hat u^H_t))$. Such a pair is a valid label for the one-step fine map, whatever the input state. Each query costs $c_H$, each coarse step costs $c_L \ll c_H$, and acquisition chooses trajectories and masks to keep total cost under a budget $B$.
What is new is not "multi-fidelity" or "selective time steps" separately; both exist. It is the combination of three things. Acquisition scores the predicted uncertainty of the correction $G_\theta$, not of the state. The coarse solver supplies a physical skeleton that keeps inputs to the fine solver near realizable states. And the surrogate starts from coarse pre-training rather than from scratch.
Research Program
The work has three parts.
The first part is an empirical test of the premise, done before any method is built. On benchmarks where dense fine and coarse trajectories can be generated offline, we measure how the step-to-step change of the correction $\delta_t = u^H_t - P(u^L_t)$ compares with that of $u^H_t$. We also measure how concentrated $\delta_t$'s change is in a few steps; the fraction of total variation in the top 10% of steps is the summary statistic. If the correction is not more temporally concentrated than the state, the central mechanism fails, and we report that.
The second part is the method. We take the cylinder-wake setup of Multi-fidelity prediction of fluid flow and temperature field based on transfer learning using Fourier Neural Operator, which pre-trains a Fourier Neural Operator (a network that learns maps between functions in frequency space and can be evaluated at any resolution) on coarse grids and fine-tunes on few fine samples. We use that approach to initialize $G_\theta$ on dense coarse trajectories. We then run STAP-style greedy mask optimization, with the variance target swapped from state to correction. We also add a cheap screening score so that the full committee is evaluated on only a fraction of candidate steps; STAP's acquisition cost grows with the square of the committee size times trajectory length.
To avoid overwriting shared physics during fine-tuning, we borrow the annealed two-fidelity loss weighting of Adaptive Multitask Neural Network for High-Fidelity Wake Flow Modeling of Wind Farms rather than plain fine-tuning. This choice follows Multi-fidelity learning for interatomic potentials: low-level forces and high-level energies are all you need, where a shared trunk with per-fidelity heads gave 45–70% improvements against 15–20% for sequential fine-tuning. That result came from interatomic potentials, so we treat it as a design prior, not as evidence for fluids.
The third part is the stiffness question raised by Transfer learning for chemically accurate interatomic neural network potentials, which found that too much low-fidelity pre-training relative to high-fidelity data can make representations rigid. We vary the amount of coarse pre-training and check whether pre-training on the correction-target parameterization is less prone to this than pre-training on raw states. The correction parameterization asks the network to learn only what the coarse solver gets wrong.
Two cited works define the limits of this design. The FNO multi-fidelity paper found that its coarsest wake grid ($32\times64$) degraded transfer through spatial artifacts. So we sweep coarse resolution, and we expect the method to fail when the coarse solver is too poor to localize events. KeyWorld supplies the architectural intuition: a heavy model on key frames and a light interpolator between them. It chooses key frames by a geometric heuristic on a single-fidelity video. We instead choose steps by correction uncertainty, and the "interpolator" is a real coarse physical solver rather than a learned frame filler.
Baselines
The naive reference point is training from scratch on fine-solver trajectories sampled uniformly in time, with uniformly chosen initial conditions, at the same total budget $B$. Every method must beat this to justify itself.
The strongest existing method on the time-selection question is STAP, which we re-implement on its own five equations. We give it a coarse-pretrained surrogate so that our pre-training is not confounded with our acquisition, and it gets the same committee size, the same greedy-plus-bit-flip optimizer, the same budget and the same tuning trials. STAP's numerical error values per PDE are not available: no reported result found on Lacuna for STAP's per-PDE errors. The only reported finding usable as a target is that STAP alone significantly beat random sampling on Kuramoto–Sivashinsky and reduced the 50th, 95th and 99th error quantiles. Beating it means lower mean and 95th-percentile rollout error at matched fine-solver cost on at least three of the five equations, including KS.
The strongest multi-fidelity baseline on the cylinder wake is the FNO multi-fidelity model: coarse pre-training plus fine-tuning on uniformly spaced fine snapshots. Its reported MAE and MRE ratios against a high-fidelity model are as low as 0.16 and 0.11 with a $128\times256$ low-fidelity grid, in Multi-fidelity prediction of fluid flow and temperature field based on transfer learning using Fourier Neural Operator. Beating it means matching those ratios with a clearly smaller fine-snapshot budget; our target is half.
Two ablations isolate the key ingredient. One keeps the full three-tier rollout but scores acquisition on state variance, as in STAP. The other keeps correction-targeted acquisition but feeds the fine solver surrogate-only states instead of coarse-anchored ones. The first isolates the choice of acquisition target; the second isolates the physical-anchoring benefit, measured by solver failure rate on KdV.
Why not just…?
Why not just pre-train on coarse data and then run STAP unchanged? This is the most direct competitor, and it is a baseline above rather than something argued away. Pre-training alone does not tell STAP which steps matter for the fidelity gap. A state-uncertainty score will still spend queries on high-motion steps where the coarse solver is already right. The state-variance ablation measures exactly this. If it matches the full method, our contribution shrinks to coarse anchoring and solver stability, and we will say so.
Why not just pick fine-solver steps by a cheap heuristic on the coarse trajectory, such as large step-to-step change or RDP turning points as in KeyWorld? This is a real threat. It needs no committee, and the coarse solver is free to inspect. We include it as a baseline at identical budget. It will show where our premise fails or holds: "fast coarse change" and "large coarse error" coincide on smooth advection but should diverge where under-resolution creates error without dramatic motion, such as KS chaos and turbulent CNS. If the heuristic ties our method everywhere, the honest conclusion is that the cheap rule suffices, and that the measurement study in the first part explains why.
Why not just spend the budget on more coarse trajectories, or uniformly strided fine snapshots? VideoCompressa (VideoCompressa) and STAP both show that uniform temporal sampling keeps redundancy. More coarse data alone cannot remove the coarse solver's systematic bias, and the stiffness result warns it may even hurt. We run a cost-matched sweep that reallocates the budget between coarse and fine, and report the best allocation for each baseline, so that no method wins merely by a better split of the budget.
Proposed Experiments
We use STAP's five PDEs (Burgers, KdV, Kuramoto–Sivashinsky, 2D incompressible and compressible Navier–Stokes). For each, we generate a coarse solver at 2×, 4× and 8× spatial coarsening, plus the FNO-MFM single-cylinder wake at its four resolutions. Costs are measured in solver wall-clock time, not counted samples. Metrics are mean and 50/95/99th-percentile autoregressive rollout error against held-out fine trajectories, MAE and MRE ratios on the wake, and the fraction of fine-solver calls that fail or blow up.
The decisive checks follow from the claim. The temporal-concentration measurement must show correction variation concentrated in a small fraction of steps. Correction-targeted acquisition must beat both state-targeted acquisition and the coarse-change heuristic at matched cost on KS and CNS. Coarse anchoring must cut KdV solver failures without any filtering heuristic. The coarse-resolution sweep should show the advantage disappearing at the coarsest grids, consistent with the $32\times64$ degradation. That last check maps where the coarse solver is too poor to tell us where it is wrong.
References
- Active Learning with Selective Time-Step Acquisition for PDEs
- Multi-fidelity prediction of fluid flow based on transfer learning using Fourier neural operator
- Adaptive Multitask Neural Network for High-Fidelity Wake Flow Modeling of Wind Farms
- Multi-fidelity learning for interatomic potentials: low-level forces and high-level energies are all you need*
- Transfer learning for chemically accurate interatomic neural network potentials
- KeyWorld: Key Frame Reasoning Enables Effective and Efficient World Models
- VideoCompressa: Data-Efficient Video Understanding via Joint Temporal Compression and Spatial Reconstruction