RECAST: A Machine-Learning Framework for Correction and Super-Resolution of Coarse-Grid PDE Solvers

RECAST: A Machine-Learning Framework for Correction and Super-Resolution of Coarse-Grid PDE Solvers

Summary

This paper introduces RECAST (Recurrent Error Correction And Super-resolution of coarse-grid Trajectories), a machine-learning framework designed to restore accuracy lost when solving time-dependent partial differential equations (PDEs) on coarse computational grids. The framework addresses two distinct sources of error that arise from coarse-grid evolution: representation error (the coarse grid cannot explicitly capture fine-scale spatial structures) and evolution error (the coarse solver advances resolved scales based on under-resolved dynamics, producing numerical diffusion, phase and dispersion errors, incorrect wave speeds, and omission of sub-grid interactions).

RECAST couples two learned components with complementary roles. A super-resolution model (SHRED-SR) learns a nonlinear map from coarse-grid information to fine-grid states, reducing representation error by recovering sub-grid spatial content. A separate correction model (SHRED-delta) is inserted within the coarse-solver loop, reducing evolution error by modifying the provisional coarse update before it is fed back into the next time step. Both components are based on the Shallow Recurrent Decoder (SHRED) architecture, which uses a recurrent LSTM encoder to extract dynamical information from temporal histories and a shallow nonlinear decoder to map the latent representation to the required output space.

The training workflow proceeds in three stages. First, SHRED-SR is trained through supervised reconstruction of fine-grid states from lagged coarse-grid histories. Second, SHRED-delta is pretrained using one-step supervised learning, where the target is the difference between the next fine-grid state projected onto the coarse grid and the provisional state predicted by the coarse solver. Third, SHRED-delta is refined through solver-in-the-loop (SOL) rollout training, in which its corrections are recursively inserted into the coarse-grid solver over multiple time steps, training it against accumulated trajectory error rather than only instantaneous residual error. The SOL training uses a staged rollout curriculum with horizons of 2, 4, and 8 steps, with model selection based on a fixed-length rollout validation error over 100 time steps.

During post-training deployment, RECAST advances the coarse state using the numerical solver, applies the learned SHRED-delta correction, feeds the corrected state into the next time step, and uses SHRED-SR to reconstruct the corresponding fine-grid solution. This produces both a dynamically corrected coarse trajectory and a fine-grid reconstruction.

The framework is evaluated on six one-dimensional PDE systems: variable-coefficient advection-diffusion, Korteweg-de Vries-Burgers, electric-field propagation in an inhomogeneous dielectric, FitzHugh-Nagumo reaction-diffusion, linear Schrödinger equation, and shallow-water/Saint-Venant system. These cases span linear and nonlinear dynamics, heterogeneous coefficients, dispersion, diffusion, reaction kinetics, wave propagation, and balance-law effects. Reference data consists of 500 trajectories with 1000 recorded time instances per trajectory, with fine grids of 256 spatial points and coarse grids of 32 or 16 points (8× and 16× spatial coarsening). The temporal input to both networks is a lagged coarse-state sequence of length 50.

Across the six PDE cases, RECAST reduces the time-averaged relative error by approximately 50-92% compared with the uncorrected coarse-grid solver over 1000-step closed-loop rollouts from unseen initial conditions. The pure coarse-grid solver has approximately 2× to 13× the time-averaged relative error of RECAST. The error plots show that applying SHRED-SR only as an offline reconstruction of the uncorrected coarse trajectory gives errors of comparable magnitude to the pure coarse solver in most cases, confirming that super-resolution alone cannot recover an accurate fine-grid solution once the underlying coarse dynamics have drifted. In contrast, RECAST maintains consistently lower errors because SHRED-delta corrects the coarse state inside the time-stepping loop before SHRED-SR reconstructs the fine-grid field.

When evaluated offline using projected coarse-grid histories obtained by block-averaging the fine-grid test trajectories, SHRED-SR reduces the error of the projected coarse representation by factors of approximately 2× to 9× across the six PDE cases, confirming its ability to recover fine-grid spatial information when supplied with dynamically consistent coarse histories.

The framework also demonstrates parameter-dependent generalization. For advection-diffusion, with the advection speed varied over [0.3, 3.0], RECAST reduces error by approximately 4.5× to 34× (78% to 97% lower error) across tested parameter bins. For KdV-Burgers, with viscosity varied over [0.01, 1.0], the error reduction is approximately 9× to 30× (89% to 97% lower error) across tested viscosity bins. These results suggest the learned correction acts as a parameter-dependent closure that adapts to changes in the influence of unresolved scales on the coarse trajectory.

RECAST is compared with P2C2Net, a contemporary PDE-preserved coarse-correction network. For advection-diffusion over a 1000-step rollout, the time-averaged mean relative error is 4.69×10−1 for the pure coarse solver, 2.09×10−1 for P2C2Net, 1.54×10−1 for P2C2Net followed by offline SHRED-SR, 1.35×10−1 for the SHRED-delta corrected coarse solver, and 6.83×10−2 for RECAST. RECAST reduces the error by about 67% relative to P2C2Net and about 56% relative to P2C2Net+SHRED-SR. For KdV-Burgers, the corresponding errors are 6.36×10−1, 5.44×10−1, 4.27×10−1, 3.58×10−1, and 1.59×10−1, with RECAST reducing error by about 71% relative to P2C2Net and about 63% relative to P2C2Net+SHRED-SR. In an extended 5000-step rollout, P2C2Net's prediction degrades substantially for KdV-Burgers and becomes unstable for advection-diffusion, whereas RECAST continues to track the fine-grid reference more closely. P2C2Net's error reaches and then exceeds the pure coarse-grid-solver error around or shortly after time step 1000.

The paper notes that although the demonstrations are on one-dimensional PDEs, the observed 8× to 16× spatial coarsening has important implications for higher-dimensional simulations: in two dimensions, the same per-direction coarsening would reduce grid nodes by 64× to 256×; in three dimensions, by 512× to 4096×. Future work includes testing on two- and three-dimensional systems, evaluating on larger-scale problems and broader ranges of PDEs and solvers, and directly assessing computational scaling including wall-clock cost and inference overhead. The code is available at github.com/MariRe1992/recast.

Improvements to AI systems:

1. Hybrid Physics-ML Solver with Dual Error Correction – Integrate RECAST’s architecture into any PDE solver (e.g., CFD, weather, plasma) to automatically separate and correct representation error (via super-resolution) and evolution error (via in-loop delta correction). The improved system can run simulations on 8–16× coarser grids while maintaining fine-grid accuracy, reducing computational cost by orders of magnitude in higher dimensions.

2. Solver-in-the-Loop Training for Long-Horizon Stability – Adopt the staged rollout curriculum (2, 4, 8 steps) with fixed-length rollout validation (100 steps) to train correction networks against accumulated trajectory error, not just instantaneous residuals. The improved system will remain stable and accurate over thousands of time steps, unlike models trained only on one-step losses that drift or blow up.

3. Parameter-Adaptive Closure Modeling – Use the learned delta-correction as a parameter-dependent closure that generalizes across varying physical coefficients (e.g., advection speed, viscosity). The improved system can automatically adjust its sub-grid corrections when system parameters change, without retraining, enabling robust deployment in real-world scenarios with uncertain or time-varying conditions.

4. Two-Stage Super-Resolution + Correction Pipeline – Combine a recurrent LSTM encoder (SHRED) with a shallow decoder for both super-resolution and error correction, sharing temporal dynamics extraction. The improved system can reconstruct fine-grid fields from coarse histories while simultaneously keeping the coarse trajectory dynamically consistent, preventing error accumulation that pure super-resolution suffers from.

5. Offline Super-Resolution with Dynamically Consistent Inputs – Use the framework to train super-resolution models on projected fine-grid histories (block-averaged) to achieve 2–9× error reduction in spatial reconstruction. The improved system can be deployed as a standalone post-processing tool for existing coarse simulations, recovering sub-grid features when the coarse dynamics are trustworthy.

6. Stability-Aware Model Selection – Implement the fixed-length rollout validation (100 steps) for model selection during training. The improved system will automatically reject models that perform well on short horizons but fail over longer rollouts, ensuring deployment-ready reliability for time-critical applications.

7. Generalizable Coarse-Grid Acceleration for High-Dimensional PDEs – Leverage the demonstrated 8–16× per-direction coarsening to enable 64–256× node reduction in 2D and 512–4096× in 3D. The improved system can make previously intractable high-fidelity simulations feasible on standard hardware, enabling real-time control, uncertainty quantification, or digital twin applications.

8. Hybrid Deployment for Unseen Initial Conditions – Use the framework’s ability to generalize to unseen initial conditions (tested on 500 trajectories) to build AI-enhanced solvers that can be safely applied to new scenarios without retraining, provided the underlying physics class remains the same. The improved system can serve as a drop-in accelerator for legacy solvers.

9. Error-Aware Uncertainty Estimation – Combine the super-resolution and correction outputs to estimate where the coarse solver is most unreliable (e.g., where SHRED-delta corrections are large). The improved system can flag regions of high model uncertainty, guiding adaptive mesh refinement or triggering full-resolution recomputation only where needed.

10. Transferable Correction for Multi-Physics Coupling – Apply the same framework to coupled PDE systems (e.g., reaction-diffusion with advection) by training separate SHRED-delta modules per equation while sharing the LSTM encoder. The improved system can handle multi-scale, multi-physics interactions on coarse grids without sacrificing accuracy, enabling whole-system simulation at reduced cost.