Mitigating Compounding Errors in Autoregressive Physical Simulators

Version 1 of 11,788 papers2,173 conceptsfrozen 2026-10-04

Traditional numerical solvers for fluid dynamics or structural mechanics are computationally expensive, often requiring supercomputers to resolve complex interactions. To bridge this gap, researchers have turned to neural surrogate models that learn to predict the next state of a system based on its current one. However, these models face a fundamental hurdle: compounding error.

When a model is used autoregressively—feeding its own previous output back in as the next input—minor inaccuracies at each step accumulate. Over long horizons, this "drift" causes the simulation to diverge from physical reality, leading to "explosions" or nonsensical trajectories.

The Feedback Loop of Failure

The core challenge in neural simulation is not just achieving high single-step accuracy, but maintaining rollout stability. As Alvaro Sanchez-Gonzalez and colleagues demonstrated in their work on Graph Networks as Learnable Physics Engines, a model that is 99% accurate on a single step can still fail completely after 100 steps because it enters "out-of-distribution" states it never saw during training.

This phenomenon, often called autoregressive drift, is a recurring theme across domains. Whether simulating urban air pollution or customer transactions, the feedback loop amplifies noise into structural degradation. Kaiwen Zhang and Zhenyu Tang describe this in the context of autonomous driving as a "short-horizon barrier," where world models eventually degrade into "nonsense" without specific interventions.

Architectural Inductive Biases

One way to combat drift is to bake physical constraints directly into the network's "DNA." By restricting what the model can learn, we prevent it from wandering into unphysical states.

Congyue Deng and the team at Tsinghua University addressed this with Alt-ConvLSTM, which mimics the mathematical structure of the forward Euler method. By alternating state updates, the model gains a physical inductive bias that helps it simulate soft-tissue dynamics more reliably than standard black-box LSTMs.

Similarly, Koen Minartz and Vlado Menkovski introduced Equivariant Probabilistic Neural Simulation (EPNS). Their insight was that simulators should respect symmetries (like rotation or translation). By ensuring the model outputs equivariant distributions, they significantly improved rollout stability in celestial and biological systems.

Multi-Scale Information Flow

Physical phenomena often happen at multiple scales simultaneously—think of large ocean currents and small surface ripples. Standard Graph Neural Networks (GNNs) often struggle with this, as they only pass information between immediate neighbors, leading to slow global communication and error buildup.

To solve this, Mario Lino and Stathi Fotiadis developed MultiScaleGNN, which processes information across different spatial resolutions. This hierarchical approach is echoed in MultiScale MeshGraphNets by Tobias Pfaff and Alexander Pritzel. By passing messages across both fine and coarse representations, these models capture global pressure changes and local turbulence more accurately, which naturally dampens the accumulation of local errors.

Yadi Cao and Menglei Chai further refined this with BSMS-GNN, using a "bi-stride" pooling strategy that respects the graph's topology, preventing the "over-smoothing" that often precedes simulation divergence.

Preserving the Underlying Physics

Perhaps the most robust way to ensure long-term stability is to ensure the model never "forgets" the laws of physics. Hao Sun and Lu Lu proposed the PDE-Preserved Neural Network (PPNN), which embeds discretized physical laws as fixed layers within the network. Unlike models that only use physics in the loss function, PPNN "bakes" the equations into the architecture, forcing the output to remain physically consistent over time.

This "physics-informed" philosophy is also seen in the work of Aldo Lipani and colleagues, who used Physics-Informed Neural Networks (PINNs) to simulate industrial furnaces. By embedding energy-balance equations, they created a surrogate that is both faster than traditional solvers and more stable than pure data-driven models.

Training for the Long Haul

Finally, researchers are changing how these models are trained. Instead of training on perfect "ground truth" pairs, models are increasingly trained on their own noisy outputs.

Fernando González and Simon Bernard 0001 highlight the importance of temporal unrolling and input denoising. By exposing the model to its own small errors during training, it learns to "self-correct" rather than letting those errors snowball. This "chain-of-forward" training, as Kaiwen Zhang calls it, is essential for creating world models that can generate consistent trajectories for minutes rather than seconds.

Papers in this version

Showing 60 of 1,788 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_mitigating_compounding_errors_in_autoreg_290c1015,
  title = {Mitigating Compounding Errors in Autoregressive Physical Simulators},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mitigating-compounding-errors-in-autoregressive-physical-simulators/rdr_290c10153079447096a36164b9252574},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_mitigating_compounding_errors_in_autoreg_290c1015_v1,
  title = {Mitigating Compounding Errors in Autoregressive Physical Simulators},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mitigating-compounding-errors-in-autoregressive-physical-simulators/rdr_290c10153079447096a36164b9252574/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}