NeurIPS 2026 · Poster

PhySPRING: Structure-Preserving Reduction of Physics-Informed Twins via GNN

A differentiable GNN framework that learns reduced spring–mass digital twins while preserving explicit physical structure for efficient simulation and downstream robot evaluation.

Yixiong Jing^* Xingyuan Chen^* Guangming Wang^† Olaf Wysocki Haibing Wu Brian Sheil

University of Cambridge Shanghai Jiao Tong University CV4DT Lab

* Equal contribution † Corresponding author: Guangming Wang

Paper PDF Code · Coming soon

01 · OVERVIEW

Reduce the model, preserve the physics.

Physics-informed digital twins are often reconstructed at the resolution needed for visual geometry, even when downstream dynamics can be represented with far fewer degrees of freedom. PhySPRING learns a hierarchy of progressively coarser graph topologies together with their mechanical parameters, keeping each level as an explicit spring–mass system rather than replacing simulation with a black-box dynamics model.

02 · METHOD

End-to-end structure-preserving reduction

From dense reconstruction to simulation-ready reduced-order models.

PhySPRING architecture showing hierarchical Neural-CLASP graph coarsening, Galerkin reduction, decoder reconstruction, and forward dynamics. — **PhySPRING architecture.** The framework progressively coarsens the input spring-mass graph with Neural-CLASP, preserves physical structure through Galerkin reduction, reconstructs multi-resolution states with hierarchical decoders, and evaluates reduced models through forward dynamics.

A · TOPOLOGY

Learned graph coarsening

A hierarchical GNN learns discrete node assignments from geometric, topological, and dynamic features.

B · PHYSICS

Explicit physical form

Galerkin projection carries the system matrices across scales so each reduced model remains a spring–mass system.

C · OPTIMIZATION

Observation-supervised

Topology and mechanical parameters are trained through differentiable forward rollouts against observed motion.

03 · RESULTS

Faster rollouts with stable physical and visual fidelity

On the PhysTwin benchmark, the reduced hierarchy maintains useful dynamics while lowering simulation cost.

2.30×up to simulation speed-up

3learned reduction levels in the paper

Explicitspring–mass structure at every level

PhysTwin optimization

Reconstruction, re-simulation, and future prediction

Each object is shown with four synchronized videos so the observation, PhysTwin baseline, SPRING-Dense, and SPRING-Sparse can be compared directly over time.

Object 01

Rope

ObservationReference

PhysTwinBaseline

SPRING-DenseOurs

SPRING-SparseOurs

Object 02

Sloth

ObservationReference

PhysTwinBaseline

SPRING-DenseOurs

SPRING-SparseOurs

04 · REAL2SIM

Robot manipulation across four reduction levels

Visualize the same grasping task with Level-0 through Level-3 reduced twins to compare task-level behavior as the physical model becomes progressively coarser.

Level-0Full-resolution twin

Level-1Reduced twin

Level-2Further-reduced twin

Level-3Most compact twin

Level-0Full-resolution twin

Level-1Reduced twin

Level-2Further-reduced twin

Level-3Most compact twin

05 · ABSTRACT

Physics-informed model order reduction for deformable digital twins

PhySPRING addresses the mismatch between visually dense reconstruction and the lower physical complexity needed for task-relevant dynamics. The method learns a hierarchy of coarsened graph topologies and mechanical parameters with a differentiable GNN, while preserving each level as an explicit spring–mass system. Experiments on PhysTwin show improved dense reconstruction and prediction, stable fidelity across reduced models, and faster simulation. The reduced models are further evaluated in a Real2Sim robot policy pipeline, where they preserve manipulation utility while improving action-sampling efficiency.

06 · CITATION

How to cite

@article{jing2026physpring,
  title   = {PhySPRING: Structure-Preserving Reduction of Physics-Informed Twins via GNN},
  author  = {Yixiong Jing and Xingyuan Chen and Guangming Wang and Olaf Wysocki and Haibing Wu and Brian Sheil},
  journal = {arXiv preprint arXiv:2605.07687},
  year    = {2026},
  note    = {NeurIPS 2026 Poster}
}