Here we present a general framework for learning simulation, and provide a
single model implementation that yields state-of-the-art performance across a
variety of challenging physical domains, involving fluids, rigid solids, and
deformable materials interacting with one another. Our framework---which we
term "Graph Network-based Simulators" (GNS)---represents the state of a
physical system with particles, expressed as nodes in a graph, and computes
dynamics via learned message-passing. Our results show that our model can
generalize from single-timestep predictions with thousands of particles during
training, to different initial conditions, thousands of timesteps, and at least
an order of magnitude more particles at test time. Our model was robust to
hyperparameter choices across various evaluation metrics: the main determinants
of long-term performance were the number of message-passing steps, and
mitigating the accumulation of error by corrupting the training data with
noise. Our GNS framework is the most accurate general-purpose learned physics
simulator to date, and holds promise for solving a wide range of complex
forward and inverse problems.
Description
[2002.09405] Learning to Simulate Complex Physics with Graph Networks
%0 Journal Article
%1 sanchezgonzalez2020learning
%A Sanchez-Gonzalez, Alvaro
%A Godwin, Jonathan
%A Pfaff, Tobias
%A Ying, Rex
%A Leskovec, Jure
%A Battaglia, Peter W.
%D 2020
%K deep-learning graphs physics simulations
%T Learning to Simulate Complex Physics with Graph Networks
%U http://arxiv.org/abs/2002.09405
%X Here we present a general framework for learning simulation, and provide a
single model implementation that yields state-of-the-art performance across a
variety of challenging physical domains, involving fluids, rigid solids, and
deformable materials interacting with one another. Our framework---which we
term "Graph Network-based Simulators" (GNS)---represents the state of a
physical system with particles, expressed as nodes in a graph, and computes
dynamics via learned message-passing. Our results show that our model can
generalize from single-timestep predictions with thousands of particles during
training, to different initial conditions, thousands of timesteps, and at least
an order of magnitude more particles at test time. Our model was robust to
hyperparameter choices across various evaluation metrics: the main determinants
of long-term performance were the number of message-passing steps, and
mitigating the accumulation of error by corrupting the training data with
noise. Our GNS framework is the most accurate general-purpose learned physics
simulator to date, and holds promise for solving a wide range of complex
forward and inverse problems.
@article{sanchezgonzalez2020learning,
abstract = {Here we present a general framework for learning simulation, and provide a
single model implementation that yields state-of-the-art performance across a
variety of challenging physical domains, involving fluids, rigid solids, and
deformable materials interacting with one another. Our framework---which we
term "Graph Network-based Simulators" (GNS)---represents the state of a
physical system with particles, expressed as nodes in a graph, and computes
dynamics via learned message-passing. Our results show that our model can
generalize from single-timestep predictions with thousands of particles during
training, to different initial conditions, thousands of timesteps, and at least
an order of magnitude more particles at test time. Our model was robust to
hyperparameter choices across various evaluation metrics: the main determinants
of long-term performance were the number of message-passing steps, and
mitigating the accumulation of error by corrupting the training data with
noise. Our GNS framework is the most accurate general-purpose learned physics
simulator to date, and holds promise for solving a wide range of complex
forward and inverse problems.},
added-at = {2020-03-15T03:55:25.000+0100},
author = {Sanchez-Gonzalez, Alvaro and Godwin, Jonathan and Pfaff, Tobias and Ying, Rex and Leskovec, Jure and Battaglia, Peter W.},
biburl = {https://www.bibsonomy.org/bibtex/2ed3ecb75d98c4dbc0de6cb2a8e39523a/kirk86},
description = {[2002.09405] Learning to Simulate Complex Physics with Graph Networks},
interhash = {7b2f45fdcabe38ad5ff4cf058034c089},
intrahash = {ed3ecb75d98c4dbc0de6cb2a8e39523a},
keywords = {deep-learning graphs physics simulations},
note = {cite arxiv:2002.09405Comment: Submitted to ICML 2020},
timestamp = {2020-03-15T03:55:25.000+0100},
title = {Learning to Simulate Complex Physics with Graph Networks},
url = {http://arxiv.org/abs/2002.09405},
year = 2020
}