RoeNets: Predicting Discontinuity of Hyperbolic Systems from Continuous Data

Shiying Xiong, Xingzhe He, Shuqi Yang, Yunjin Tong, Runze Liu, Bo Zhu

Keywords: Data-driving modeling, neural PDE, hyperbolic dynamic system

Abstract:

Predicting future discontinuous phenomena that are unobservable from the training data sets has been a challenging problem for scientific machine learning. In this paper, we introduce a novel learning paradigm to predict the emergence and evolution of various kinds of discontinuities for hyperbolic dynamic systems based on smooth observation data. At the heart of our approach is a templaterizable and data-driven Riemann solver that functions as a strong inductive prior to tackle the potential discontinuities.
The key design of our templaterized Riemann approximator is inspired by the classical Roe solver (P. L. Roe, J. Comput. Phys., vol. 43, 1981), which served as a fundamental mathematical tool for simulating various hyperbolic systems in computational physics.
By carefully designing the computing primitives, data flow, and incorporating a novel pseudoinverse processing module, we enable our data-driven predictor to inherently satisfy all the essential mathematical criteria of a Roe hyperbolic solver derived from the first principles and hence deliver accurate predictions of hyperbolic dynamics.

One-sentence Summary: We introduce a novel machine learning paradigm to predict the emergence and evolution of various kinds of discontinuities for hyperbolic dynamic systems.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): https://openreview.net/references/pdf?id=Po73yWssX