# Learning to Simulate Complex Physics with Graph Networks

**Authors:** Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, Peter W. Battaglia  
**arXiv:** [2002.09405](https://arxiv.org/abs/2002.09405)  
**Date:** 2020-02-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2002.09405v2/

## Key points

- Authors present a machine learning framework and model implementation for simulating challenging physical domains
- Proposed framework called "Graph Network-based Simulators" (GNS) represents the state of a physical system using particles expressed as nodes in a graph
- Dynamics of the system are computed through learned message-passing techniques
- Model can generalize from single-timestep predictions during training to different initial conditions and thousands of timesteps at test time
- Demonstrates robustness across various hyperparameter choices and evaluation metrics
- Number of message-passing steps and mitigating error accumulation through noise introduction in training data were found to be main factors influencing long-term performance
- GNS framework advances the current state-of-the-art in learned physical simulation and has potential for solving complex problems in various applications.

## Summary

In their paper titled "Learning to Simulate Complex Physics with Graph Networks," authors Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter W. Battaglia present a machine learning framework and model implementation for simulating challenging physical domains. The proposed framework called "Graph Network-based Simulators" (GNS) represents the state of a physical system using particles expressed as nodes in a graph. Through learned message-passing techniques, the dynamics of the system are computed. The results show that the model can generalize from single-timestep predictions during training to different initial conditions and thousands of timesteps at test time. It also demonstrates robustness across various hyperparameter choices and evaluation metrics. The main factors influencing long-term performance were found to be the number of message-passing steps and mitigating error accumulation through noise introduction in training data. Overall, the GNS framework advances the current state-of-the-art in learned physical simulation and has potential for solving complex problems in various applications.