Peter Battaglia: Excited to present “Learning to Simulate Complex Physics with Graph Networks”.
Our model can generate realistic simulations, and generalizes to much larger systems and longer trajectories than its training.
w/ @spectralhippo @RexYing0923 @jure https://t.co/fEimZ0LBOJ
27 replies, 2616 likes
Sam Schoenholz: I am shocked / amazed by the fidelity of these learned simulations.
2 replies, 58 likes
Kyle Cranmer: 🤯👇🔥
2 replies, 40 likes
Andrew Davison: Graph networks sounds spot-on to me as the right way to simulate physics with a lot of interaction and contacts. @PeterWBattaglia you surely have to get this up and running on a graph processor; it would fly I assume. #SpatialAI
1 replies, 39 likes
Shane Gu 顾世翔: Another example of extreme generalization from @PeterWBattaglia @jure et al on amortizing particle simulation using graph NNs. Train 2.5k -> test 28k. Great impactful results for computational fluid dynamics, CG and physics simulation for robotics control. https://arxiv.org/abs/2002.09405 https://t.co/Fkc0TlyS5F
0 replies, 33 likes
Xavier Bresson: Graph neural networks for computational fluid dynamics (CFD) !
CFD is applied to many industrial problems : aircraft design, aerospace, weather prediction, cardiovascular system, movie special effects, etc
0 replies, 27 likes
Shane Gu 顾世翔: An incredible feature of neural nets is that by choosing certain architectures they may exhibit extreme generalization. Similarly in sequence prediction, by choosing the right RNNs, the model may generalize extremely. https://twitter.com/TingwuWang/status/1229587903971516416
0 replies, 22 likes
Brant Robertson: A fascinating must-read for computational physicists interested in the convergence of machine learning and predictive simulation.
0 replies, 16 likes
Shirley Ho: Approximating simulations with #graphnetworks rock!
0 replies, 6 likes
ML and Data Projects To Know: 📙 Learning to Simulate Complex Physics with Graph Networks
Authors: Alvaro Sanchez-Gonzalez, Jonathan Godwin, @spectralhippo, @RexYing0923, @jure, @PeterWBattaglia
Paper: https://arxiv.org/abs/2002.09405 https://t.co/XY9QGZEOww
0 replies, 5 likes
They propose a framework for high-precision physics simulations of liquids, etc.,using Graph Neural Network(GNN). We apply GNN to nearby particles to predict the motion of a particle at the next time step. It is also robust to the choice of hyperparameters https://t.co/rqTxSdkjrS
1 replies, 5 likes
Daisuke Okanohara: A wide range of physical simulations (fluid, rigid, solids, and deformable materials) can be emulated precisely by particle-based graph NN where each particle corresponds to a node, and neighbors are connected. https://arxiv.org/abs/2002.09405 https://drive.google.com/file/d/1Ri3RkuqyZ1xhg0QkXqzwfkxUCvn_0YKn/view?usp=sharing
0 replies, 5 likes
Rodolfo Rosini ☕️✨: Wow this is terribly interesting. Never imagined about using graphs to simulate physics, only higher level structures.
0 replies, 2 likes
akira: Very impressive results
0 replies, 1 likes
OGAWA, Tadashi: =>
"Learning to Simulate Complex Physics with Graph Networks", DeepMind and Stanford, arXiv, Feb 21, 2020 https://arxiv.org/abs/2002.09405
Particles, expressed as nodes in a graph, and computes dynamics via learned message-passing
Videos https://sites.google.com/view/learning-to-simulate/home#h.p_hjnaJ6k8y0wo https://t.co/F7xNM89ojq
1 replies, 0 likes
Found on Mar 10 2020 at https://arxiv.org/pdf/2002.09405.pdf