Augmenting Physical Models with Neural Residuals

This direction is merged into another direction: residual dynamics learning, physics-informed error correction. Its last version stays readable and citable here.

Version 1 of 1546 papers788 conceptsfrozen 2026-10-04

When we simulate a robot arm or a weather system, we face a fundamental trade-off. Classical physics models are interpretable and generalize well, but they are often "too simple," failing to account for messy real-world effects like friction, turbulence, or material aging. Conversely, pure deep learning models can capture these complexities but are notoriously data-hungry and prone to "hallucinating" physically impossible behavior.

A growing body of research suggests a "best of both worlds" solution: residual physics modeling. Instead of asking a neural network to learn the entire system from scratch, we use a known physical model to provide a baseline and task the neural network with learning only the residual—the gap between theory and reality.

The Intuition of the Residual

The core idea is a simple decomposition of the system's dynamics: $$\text{Actual State} = \text{Physics Model}(\text{State}) + \text{Neural Network}(\text{State})$$

In this framework, the physics model handles the "global" structure—things we are certain about, like gravity or conservation of mass. The neural network acts as a "local" corrector, learning the specific, unmodeled nuances of the environment.

As Andreas René Geist and Sebastian Trimpe explain in Structured learning of rigid-body dynamics: A survey and unified view, this approach simplifies the learning task. Because the network doesn't have to "re-discover" Newton's laws, its optimization landscape is much flatter and easier to navigate. This leads to what Hongbin Lin and Xiangyu Chu call "Physical Knowledge Distillation," where models can reach target accuracy with up to 99% less training data than pure machine learning approaches.

Bridging the Simulation-to-Reality Gap

This hybrid approach is particularly effective in robotics, where the "sim-to-real" gap often causes controllers to fail when moved from a perfect simulator to a dusty lab floor.

In Simulator-Augmented Interaction Networks (SAIN), Anurag Ajay and Joshua B. Tenenbaum demonstrate how a system can use a standard physics engine for baseline predictions while a neural network learns to compensate for complex contact dynamics and friction. Similarly, Philipp Foehn introduced NeuroBEM, which combines aerodynamic first principles with neural residuals to reduce trajectory errors by 50% during high-speed quadrotor flight, capturing effects like blade flapping that standard equations omit.

For soft robots, which are notoriously difficult to model analytically due to their infinite degrees of freedom, Curtis C. Johnson and Marc D. Killpack showed that a residual "error DNN" can eliminate steady-state errors caused by material plasticity, effectively "teaching" the robot about its own mechanical imperfections.

Stability and Physical Grounding

One of the primary risks of adding a neural network to a physical system is the loss of guarantees. If the network makes a wild prediction, the system could become unstable.

To solve this, researchers like Thomas Beckers have integrated Gaussian Process Regression with Lyapunov stability theory to ensure that even as the model learns unmodeled dynamics, the robot's tracking error remains bounded. In the realm of climate modeling, Ali J. Ramadhan and John C. Marshall use Neural Differential Equations (NDEs) to ensure that the learned residuals still respect fundamental conservation laws, preventing the "physical hallucinations" common in purely data-driven weather forecasts.

Beyond Robotics: Materials and Media

The residual philosophy is spreading into other domains where "perfect" models are computationally expensive:

  • Materials Science: Bin Wang and Uri M. Ascher developed "Neural Material," which uses a neural network to learn the residual stress and damping of complex deformable objects, allowing for realistic simulations of heterogeneous materials from sparse data.
  • Video Generation: Tharindu Wickremasinghe and Yu Yuan introduced NewtonGen, which uses a Neural Newtonian Dynamics module to ensure that objects in AI-generated videos follow realistic trajectories, using residuals to handle non-linear deformations.
  • Reinforcement Learning: The Neural Motion Simulator (MoSim) by Chenjie Hao decomposes state transitions into rigid-body dynamics and a residual corrector, enabling "zero-shot" reinforcement learning where agents learn complex skills entirely within a neural world model.

By treating physical laws as a foundation rather than a constraint, this research direction allows us to build models that are as robust as classical physics and as flexible as modern deep learning. As Roshan Sharma and Bernt Lie demonstrated with their hybrid helicopter model, this approach doesn't just improve accuracy—it can eventually help researchers "discover" the missing symbolic equations that were absent from their original theories.

Papers in this version

Showing 60 of 546 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_augmenting_physical_models_with_neural_r_6f689fc1,
  title = {Augmenting Physical Models with Neural Residuals},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/augmenting-physical-models-with-neural-residuals/rdr_6f689fc141a14218915aabd19e1f7e06},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_augmenting_physical_models_with_neural_r_6f689fc1_v1,
  title = {Augmenting Physical Models with Neural Residuals},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/augmenting-physical-models-with-neural-residuals/rdr_6f689fc141a14218915aabd19e1f7e06/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}