Spatially Adaptive Computation for Physical Systems

This direction is split into other directions: uniform grid inefficiency. Its last version stays readable and citable here.

Version 1 of 162 papers62 conceptsfrozen 2026-10-04

In most physical simulations, the "interesting" parts of the world are small. Whether it is the sharp edge of a shockwave, the swirling vortices behind a wing, or the steep chemical gradients at a combustion front, the critical physics is usually concentrated in a tiny fraction of the total space. The rest of the domain—the vast stretches of empty air or smooth, laminar flow—is relatively "boring."

Traditional numerical methods and modern neural operators often ignore this imbalance. By applying a uniform grid or a constant computational budget across the entire domain, they waste massive amounts of energy processing simple regions with the same intensity as complex ones. Spatially adaptive computation is the effort to fix this by "picking battles": concentrating resolution, bits, and layers only where the physics demands it.

The Problem with Uniformity

The inefficiency of uniform grids is a classic problem in numerical analysis. As William Gropp and David Keyes noted in their foundational work on domain decomposition, applying high-resolution resources to an entire domain is computationally prohibitive. If you need a fine mesh to capture a boundary layer, a uniform grid forces you to use that same fine mesh in the far-field where nothing is happening.

This isn't just a matter of grid density; it’s a matter of mathematical convergence. Frank de Hoog and D. R. Jackett demonstrated that nonuniform grids can maintain high-order global accuracy even when local truncation errors look messy, provided the points are clustered where the solution changes rapidly.

Learning Where to Look

In the era of deep learning, the challenge has shifted from manually designing meshes to teaching models to identify "hotspots" automatically. Recent research has moved toward "physics-aware" architectures that dynamically allocate their internal resources.

For example, Haiyang Xin and colleagues developed Skip-Block Routing (SBR), which allows a neural operator to skip deep layers of computation in stable regions of a physical field. By focusing "brain power" only on turbulent vortices, the model doubles its speed without losing accuracy. Similarly, Winfried van den Dool and Max Welling introduced Adaptive Mesh Quantization (AMQ), which uses a tiny auxiliary model to decide how many bits of precision to assign to different regions. High-complexity areas get more bits; smooth areas get fewer.

Goal-Oriented and Error-Driven Refinement

A key intuition in this field is that we shouldn't just refine where the physics is complex, but where the error is likely to affect the final answer. This is known as Goal-Oriented Adaptivity.

Joseph Gregory Wallwork and Nicolas Barral have shown that by focusing mesh resolution on specific "Quantities of Interest"—like the power output of a tidal turbine—simulations can achieve high precision with significantly fewer resources than global refinement.

In the neural surrogate space, Xuan Minh Vuong Nguyen and Elie Hachem take this further with MiSe-GNN. Their model learns to predict its own error maps. It then uses these maps to dynamically rewire its internal graph structure, creating "highways" for information flow in regions where the physics is most likely to be miscalculated.

Hierarchical and Meshless Representations

To escape the "pixelation" of uniform grids, researchers are increasingly turning to hierarchical structures and point-based methods:

  • Octrees and Quadtrees: Jia-Heng Tang developed OctField, which uses a differentiable octree to focus computation on intricate surface details in 3D modeling, leaving empty space unprocessed.
  • Adaptive Tokenization: In fluid simulation, Zeyi Xu and Bingbing Ni introduced the AMR-Transformer, which uses Adaptive Mesh Refinement to create "tokens" only where shockwaves or vortices exist, reducing the computational load of Transformers by up to 60x.
  • Point Clouds: Ali Kashefi and Leonidas Guibas demonstrated that treating fluid flow as a point cloud allows models to bypass grid-based waste entirely, achieving massive speedups on irregular geometries.

Closing the Loop with Reinforcement Learning

The ultimate goal is a simulation that manages its own resolution in real-time. Tailin Wu and Qingqing Zhao achieved this with LAMP, a framework where a Graph Neural Network (GNN) and a reinforcement learning agent work together. The agent learns a policy to coarsen or refine the mesh as the simulation progresses, ensuring that computational power always follows the most dynamic features of the system.

By moving away from the "one-size-fits-all" approach of uniform grids, these methods allow us to simulate larger, more complex systems—from 2-billion-degree-of-freedom oil reservoirs Shizhe Li to high-dimensional quantum structures E Weinan—without needing a supercomputer for every step.

Papers in this version

Showing 60 of 62 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_spatially_adaptive_computation_for_physi_f4f20148,
  title = {Spatially Adaptive Computation for Physical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatially-adaptive-computation-for-physical-systems/rdr_f4f20148fdfa45d4b3e8c8c470fd9e84},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_spatially_adaptive_computation_for_physi_f4f20148_v1,
  title = {Spatially Adaptive Computation for Physical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatially-adaptive-computation-for-physical-systems/rdr_f4f20148fdfa45d4b3e8c8c470fd9e84/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}