Learning Sub-Grid Physics for Efficient High-Fidelity Simulation

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 1219 papers222 conceptsfrozen 2026-10-04

Simulating the physical world—from the swirling currents of the ocean to the microscopic propagation of cracks in a turbine blade—presents a fundamental computational bottleneck. To achieve high fidelity, traditional numerical solvers must resolve every relevant spatial and temporal scale. For turbulent flows or phase transitions, this requires grids so fine that the computational cost becomes prohibitive for real-time engineering or long-term climate forecasting.

A growing body of research suggests a way out of this "resolution trap": instead of brute-forcing the grid, we can use machine learning to discover the sub-grid physics that allow coarse simulations to behave like fine ones.

The Resolution Bottleneck and the Closure Problem

In classical fluid dynamics, Direct Numerical Simulation (DNS) is the gold standard because it resolves all scales of motion. However, as Xiangyu Hu and others note, the cost of DNS scales aggressively with the Reynolds number, making it impractical for most industrial applications. When we move to a coarser grid to save time, we lose the "fine-grained" details—like tiny vortices or sharp interfaces—that actually drive the large-scale behavior.

This creates the closure problem: how do we account for the influence of unresolved small-scale features on the resolved large-scale flow? Traditional approaches use hand-tuned mathematical models (like LES or RANS), but these often fail in complex, chaotic regimes.

Learning Optimal Discretization

One powerful solution is to let the data define the discretization itself. Suryanarayana Maddu and colleagues introduced STENCIL-NET, which uses neural networks to learn "solution-adaptive" discretization rules. Instead of using a fixed mathematical formula to calculate derivatives, the model learns the optimal weights for a local stencil based on the current state of the simulation. This allows the solver to maintain high-order accuracy even on grids that would normally be too coarse to capture sharp gradients or shock waves.

Similarly, Chuanbo Hua and Michael Poli have explored Graph Spline Networks to create efficient surrogate models. By treating the simulation as a dynamic system on a graph, these models can capture complex spatial dependencies without the rigid memory constraints of traditional grid-based deep learning.

Super-Resolution and Generative Synthesis

If a coarse simulation is fast but "blurry," can we simply "enhance" it? This is the intuition behind Super-Resolution (SR) for scientific data. Amir Barati Farimani and Pranshu Pant developed SR-DNS Net, which uses a U-Net architecture to reconstruct high-fidelity turbulent data from low-resolution inputs.

Beyond simple upscaling, generative models like Generative Adversarial Networks (GANs) can synthesize the statistical "texture" of turbulence. Claudia Drygala and Hanno Gottschalk demonstrated that GANs can generate high-resolution turbulence snapshots in seconds, capturing the underlying chaotic distributions that would otherwise take weeks of supercomputing time to simulate.

Hybrid Solvers: Blending Physics and ML

The most robust results often come from "blending" traditional solvers with neural operators. Saaketh Desai and Rémi Dingreville proposed a framework where a neural operator "leaps" forward in time, and a numerical solver periodically corrects the trajectory. This hybrid approach maintains physical consistency while achieving speedups of over 16x.

In the realm of fluid-structure interaction, Kiwon Um and Nils Thuerey developed MLFLIP, which uses neural networks to "inject" splash and droplet details into coarse liquid simulations. By learning the relationship between the coarse flow and the fine-scale splashes, the model provides the visual and physical richness of a high-resolution simulation at a fraction of the cost.

Bridging the Micro-Macro Gap

The ultimate goal of this research direction is to automate the discovery of macroscopic laws from microscopic observations. Seungjoon Lee and Yannis G. Kevrekidis have shown that machine learning can "discover" hidden PDEs directly from fine-scale data. This effectively solves the "closure problem" by learning a reduced-order model that captures the essential physics of a complex system—like shale gas transport or polymer melts—without needing to simulate every molecule.

By treating neural networks not as black-box replacements for physics, but as learned components of a multiscale solver, researchers are creating a new generation of simulation tools that are both fast enough for design optimization and accurate enough for scientific discovery.

Papers in this version

Showing 60 of 219 papers.

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_learning_sub_grid_physics_for_efficient__d9125f3b,
  title = {Learning Sub-Grid Physics for Efficient High-Fidelity Simulation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-sub-grid-physics-for-efficient-high-fidelity-simulation/rdr_d9125f3b786e45b8a227d00dcd2323e7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_learning_sub_grid_physics_for_efficient__d9125f3b_v1,
  title = {Learning Sub-Grid Physics for Efficient High-Fidelity Simulation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-sub-grid-physics-for-efficient-high-fidelity-simulation/rdr_d9125f3b786e45b8a227d00dcd2323e7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}