MaNGO — Adaptable Graph Network Simulators via Meta-Learning

Philipp Dahlinger, Tai Hoang, Denis Blessing, Niklas Freymuth, Gerhard Neumann

Autonomous Learning Robots, Karlsruhe Institute of Technology (KIT), Karlsruhe
Conference: NeurIPS 2025

Paper Code arXiv
MaNGO Schematic Overview

Learning to simulate physics while inferring simulation parameters from similar trials leads to our proposed Meta Neural Graph Operator (MaNGO) approach. The context set is aggregated to form a latent representation of material properties. Given an unseen initial state, the Graph Network Simulator (GNS) uses this latent representation to generate trials that follow the material laws of the context set, enabling accurate predictions for new conditions. The example prediction aligns perfectly with the ground truth data.

Abstract

Accurately simulating physics is crucial across scientific domains, with applications spanning from robotics to materials science. While traditional mesh-based simulations are precise, they are often computationally expensive and require knowledge of physical parameters, such as material properties. In contrast, data-driven approaches like Graph Network Simulators (GNSs) offer faster inference but suffer from two key limitations: Firstly, they must be retrained from scratch for even minor variations in physical parameters, and secondly they require labor-intensive data collection for each new parameter setting. This is inefficient, as simulations with varying parameters often share a common underlying latent structure. In this work, we address these challenges by learning this shared structure through meta-learning, enabling fast adaptation to new physical parameters without retraining. To this end, we propose a novel architecture that generates a latent representation by encoding graph trajectories using conditional neural processes (CNPs). To mitigate error accumulation over time, we combine CNPs with a novel neural operator architecture. We validate our approach, Meta Neural Graph Operator (MaNGO), on several dynamics prediction tasks with varying material properties, demonstrating superior performance over existing GNS methods. Notably, MaNGO achieves accuracy on unseen material properties close to that of an oracle model.

Method

MaNGO (Meta Neural Graph Operator) uses meta-learning to make graph-based simulators adaptable to new physical systems. It learns from many related simulations, each with different material properties or conditions, and builds a shared understanding that helps it generalize to unseen scenarios.

Meta-Learning and Physical Parameters: During training, MaNGO learns to infer a latent description of the underlying simulation parameters directly from context data. Using a Conditional Neural Process (CNP) framework, it encodes this information into a compact representation that guides the simulation. This enables MaNGO to generate trials consistent with the inferred material behavior in an end-to-end, data-driven manner.

Spatiotemporal Encoder: The encoder extracts meaningful information from sequences of graphs that evolve over time. It processes time-dependent changes with a 1D convolution and summarizes spatial relationships using a Deep Set aggregation. This way, the encoder builds a compact, translation-invariant representation of the system’s motion and structure.

MaNGO Decoder: The MaNGO decoder follows a neural operator paradigm, predicting the full simulation trajectory from start to end in a single forward pass. It alternates between message passing across nodes (spatial reasoning) and convolution across time (temporal reasoning). This design enables faster simulation times and reduces the error propagation that commonly occurs in autoregressive learned simulators.

MaNGO Decoder Architecture

Overview of the MaNGO decoder architecture. It takes the latent representation from the encoder and an initial state as input, and outputs a sequence of future graphs.

Summary: Together, these components allow MaNGO to simulate new materials and environments without retraining from scratch. Its meta-learning setup and graph-based design make it a powerful tool for modeling diverse physical phenomena in a data-efficient and generalizable way.

Qualitative Results

Example simulation demonstrating MaNGO’s ability to adapt to unseen physical conditions.

Quantitative Results

Quantitative Results 1

Performance comparison of MaNGO and baseline methods across four datasets. We report the mean and 95% confidence interval of the Full Rollout MSE over five runs. The x-axis shows the number of context samples used for meta-learning. MaNGO consistently outperforms non-meta-learning and alternative meta-learning baselines, achieving performance close to the oracle model with access to ground-truth parameters.

Quantitative Results 2

Left: Comparison of different GNS decoders with oracle information. MaNGO outperforms both MGN and EGNO, with the gap most pronounced in the Sphere Cloth Coupling task due to its complex dynamics.
Right: Visualization of the 2D latent space for Deformable Plate (Easy), color-coded by Poisson’s ratio. The sharp bend at a Poisson ratio of 0 separates materials with positive and negative Poisson behavior, showing that MaNGO’s latent representation effectively distinguishes between fundamentally different material regimes.

BibTeX


        @article{dahlinger2025mango,
          title={MaNGO — Adaptable Graph Network Simulators via Meta-Learning},
          author={Dahlinger, Philipp and Hoang, Tai and Blessing, Denis and Freymuth, Niklas and Neumann, Gerhard},
          journal={arXiv preprint arXiv:2510.05874},
          year={2025}
        }