NeurIPS 2022

Timezone: »

Adversarial Noise Injection for Learned Turbulence Simulations

Jingtong Su · Julia Kempe · Drummond Fielding · Nikolaos Tsilivis · Miles Cranmer · Shirley Ho

in Machine Learning and the Physical Sciences »

Machine learning is a powerful way to learn effective dynamics of physical simulations, and has seen great interest from the community in recent years. Recent work has shown that deep neural networks trained in an end-to-end manner seem capable to learn to predict turbulent dynamics on coarse grids more accurately than classical solvers. All these works point out that adding Gaussian noise to the input during training is indispensable to improve the stability and roll-out performance of learned simulators, as an alternative to training through multiple steps. In this work we bring insights from robust machine learning and propose to inject adversarial noise to bring machine learning systems a step further towards improving generalization in ML-assisted physical simulations. We advocate that training our models on these worst case perturbation instead of model-agnostic Gaussian noise might lead to better rollout and hope that adversarial noise injection becomes a standard tool for ML-based simulations. We show experimentally in the 2D-setting that for certain classes of turbulence adversarial noise can help stabilize model rollouts, maintain a lower loss and preserve other physical properties such as energy. In addition, we identify a potentially more challenging task, driven 2D-turbulence and show that while none of the noise-based attempts significantly improve rollout, adversarial noise helps.

[ Poster]

Author Information

Jingtong Su (New York University)

Julia Kempe (New York University)

Drummond Fielding (Center for Computational Astrophysics, Flatiron Institute)

Nikolaos Tsilivis (New York University)

Miles Cranmer (Princeton University)

Shirley Ho (Flatiron Institute)

Shirley Ho is a group leader and acting director at Flatiron Institute at Simons foundation, a research professor of physics and an affiliated faculty at Center for Data Science at NYU. Ho also holds associate (adjunct) professorship at Carnegie Mellon University and visiting appointment at Princeton University. She was a senior scientist at Berkeley National Lab from 2016-2018 and a Cooper-Siegel Development chair professor at Carnegie Mellon University before that. Ho was a Seaborg and Chamberlain Fellow from 2008-2011 at Berkeley Lab, after receiving her PhD in Astrophysics from Princeton University in 2008 under supervision of David Spergel. Ho graduated summa cum laude with a B.A. in Physics and a B.A. in Computer Science from UC Berkeley. A cited expert in cosmology, machine learning applications in astrophysics and data science,her interests are using deep learning accelerated simulations to understand the Universe, and other astrophysical phenomena. She tries her best to balance her love for the Universe, the machine and life especially during these crazy times.