Research map / Probabilistic and scientific modeling / Dynamical systems
MaNGO — Adaptable Graph Network Simulators via Meta-Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MaNGO, a meta-learning graph network simulator that encodes context trajectories with a CNP (deep-set + 1D CNN encoder) and decodes full trajectories with MPN + 1D temporal convolution blocks, adapting to unseen material parameters without retraining.
Problem
Graph network simulators must be retrained from scratch for each new physical parameter setting and require new labeled data per setting, while inverse estimation of material parameters is ill-posed and needs knowledge of the governing PDE.
Approach
Combine conditional neural processes for context-based latent encoding of simulation parameters with a neural-operator-style trajectory decoder that replaces EGNO's equivariant layers with message passing plus 1D temporal convolutions, predicting whole trajectories in one pass; an auxiliary head predicts the true simulation parameter during meta-training.
Key results
- MaNGO with few context samples (2-3) achieves low full-rollout MSE on all four tasks, close to the MaNGO-Oracle upper bound, and outperforms non-meta-learning baselines (EGNO, MGN).
- EGNO fails to learn the Planar Bending task; the paper proves EGNN maps planar inputs to planar outputs, so it cannot represent the required deformation.
- MaNGO predicts a 100-step SCC trajectory in about 13 ms versus about 500 ms for autoregressive MGN; latent representation also robust up to 10% noise and 10% node dropout on Deformable Plate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MaNGO meta-learning with a context set instead of non-meta-learning simulation | MaNGO · Deformable Plate (Easy), Deformable Plate (Hard), Planar Bending, and Sphere Cloth Coupling | Full Rollout MSE | Helped |
| Increase the context set size | MaNGO meta-learning · Four simulation datasets | Full Rollout MSE | Helped |
| Use MaNGO instead of MGN or EGNO as the decoder | Oracle simulation with known material properties · Sphere Cloth Coupling | MSE | Helped |
| Use EGNO for prediction | EGNO · Planar Bending | deformation prediction | Hurt |
| Add Gaussian noise to context data during training | MaNGO encoder · Deformable Plate robustness evaluation | normalized MSE under noisy and sparse observations: 0.05 | Helped |
| Reduce batch shuffling to mimic the CNP setup (ablation) | MGN decoder · CNP training setup | performance | Hurt |
Surprising findings
- EGNO, due to its E(3)-equivariance constraint, is provably unable to produce non-planar deformations from a flat initial plane in the Planar Bending task.
- Batched inference with the MaNGO decoder scales GPU memory linearly with the number of predicted time steps.
Limitations
- Assumes a fixed graph topology and fixed number of nodes across trials.
- Predicting full trajectories in one pass requires substantial GPU memory that grows linearly with the number of time steps.
- Only supports a single observation modality at test time (e.g., no point clouds from cameras).
Experimental setup
Datasets: Deformable Plate (SOFA-generated, extended) Planar Bending (new) Sphere Cloth Coupling (NVIDIA Isaac Sim, inspired by Physion++)
Benchmarks: Deformable Plate (easy) Deformable Plate (hard) Planar Bending Sphere Cloth Coupling
Metrics: mean squared error (full rollout) normalized MSE under noise and node dropout inference time
Baselines: EGNO MeshGraphNet (MGN) MaNGO (Two-Stage) MaNGO-Oracle MaNGO without context
Builds on: conditional neural processes equivariant graph neural operator (EGNO) MeshGraphNet message passing neural networks deep sets
Closest papers by meaning
- Point Cloud Sequence Encoding for Material-conditioned Graph Network Simulators (NeurIPS 2026 · similarity 0.89)
- Diffusion-Based Hierarchical Graph Neural Networks for Simulating Nonlinear Solid Mechanics (NeurIPS 2025 · similarity 0.87)
- EvoMesh: Adaptive Physical Simulation with Hierarchical Graph Evolutions (ICML 2025 · similarity 0.86)
- Physics-Informed DeepONets for drift-diffusion on metric graphs: simulation and parameter identification (ICML 2025 · similarity 0.85)
- Graph Mamba Operator: A Latent Simulator for Interacting Particle Systems (NeurIPS 2026 · similarity 0.85)
- DEL: Discrete Element Learner for Learning 3D Particle Dynamics with Neural Rendering (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.