摘要: Motivation: Constructing models from observed data to be able to describe sufficiently well a given system is the key goal of system identification. Properly identified models are invaluable for many down-stream tasks, including system design and control. While linear system identification has reached full maturity with regards to theoretical guarantees and reliable tools, the nonlinear counter part is far less understood. When building non-linear models, one could rely on classical engineering or physics principles, but one can also try to learn them directly from a data set containing many past observations. A promising direction of the second category is centered around deep neural ODEs [Chen et al., 2018], where the underlying differential equations of the system are modeled explicitly using a neural network.Challenge: Thermoacoustics studies systems dominated by dynamic interaction between pressure waves …