🥭 MaNGO: Adaptable Graph Network Simulators via Meta-Learning

This repository will host the official code release for our NeurIPS 2025 paper:

MaNGO: Adaptable Graph Network Simulators via Meta-Learning
Philipp Dahlinger, Tai Hoang, Denis Blessing, Niklas Freymuth, Gerhard Neumann Karlsruhe Institute of Technology (KIT)

Check out our project page for an overview and visualizations of all tasks!

---

Overview

MaNGO introduces a Meta Neural Graph Operator that enables Graph Network Simulators to adapt across different physical systems.
The method combines meta-learning with a neural operator based architecture, allowing the simulator to generalize to unseen material properties and predict full trajectories efficiently and stably.

---

Installation

We use uv as our environment manager. To set up the environment, run:

# Create a virtual environment (recommended)
uv venv
uv sync

For developing access, run:

uv pip install -e .

---

Usage

If you just want to have a look at the decoder code as a baseline for your own experiments, check it out in src/mango/simulator/ml_decoder/mango_decoder.py.

For training the full MaNGO model on the datasets, you can use the training script train.py. You need to provide a hydra config file. As an example, you can run

uv python train.py +experiment/final_exp/cnn_deepset_mango=dp_easy_v5 +platform=local_multirun

For that to work you need to download the dataset here: Dataset Download Link. Put the hdf5 files into a folder ../datasets/mango/ relative to the root of this repository (or update the path in the dataset configs in configs/dataset/)

Contact

For questions, please contact:
Philipp Dahlinger – philipp.dahlinger@kit.edu