Discovering Governing Equations from Partial Measurements with Deep Delay Autoencoders

Abstract

A central challenge in data-driven model discovery is the presence of hidden, or latent, variables that are not directly measured but are dynamically important. Takens' theorem provides conditions for when it is possible to augment these partial measurements with time delayed information, resulting in an attractor that is diffeomorphic to that of the original full-state system. However, the coordinate transformation back to the original attractor is typically unknown, and learning the dynamics in the embedding space has remained an open challenge for decades. Here, we design a custom deep autoencoder network to learn a coordinate transformation from the delay embedded space into a new space where it is possible to represent the dynamics in a sparse, closed form. We demonstrate this approach on the Lorenz, R össler, and Lotka-Volterra systems, learning dynamics from a single measurement variable. As a challenging example, we learn a Lorenz analogue from a single scalar variable extracted from a video of a chaotic waterwheel experiment. The resulting modeling framework combines deep learning to uncover effective coordinates and the sparse identification of nonlinear dynamics (SINDy) for interpretable modeling. Thus, we show that it is possible to simultaneously learn a closedform model and the associated coordinate system for partially observed dynamics.

Decoding depression during subcallosal cingulate deep brain stimulation

“…The resulting correlation suggests the SCC accounts for only part of the, consistent with MDD being a disorder arising across distributed networks. Improving on this performance for future readouts may require coverge of brain regions beyond SCC, or computational models that can infer network-wide states from limited SCC observations [41].…”

Characterization of Acoustic Emissions From Analogue Rocks Using Sparse Regression‐DMDc

“…Several methods are available in open source python notebooks (Demo et al, 2018;Kaptanoglu et al, 2021), making their use and implementation in a discovery pipeline efficient and advantageous. All these methods, which are aimed at providing interpretable models, can also be used in combination with deep learning (Bakarji et al, 2022;Champion et al, 2019;Chen et al, 2019;Gin et al, 2021;Lusch et al, 2018). Thus data-driven modeling (Brunton et al, 2016a), like DMDc, offer flexible and robust mathematical framework for advancing scientific efforts and unraveling scientific questions.…”

“…Can we discriminate events related to individual or coupled process, such as the emergence of damage as a result of cracking induced by fluid movement and changes in the water content during interactions with subsurface clay-bearing porous mediums (e.g., at the interface of natural rock and engineered barriers in nuclear waste disposal repositories)? The value of extending DMDc can be captured by the fact that even when the correct variables are unknown, such data-driven methods can be used to discover advantageous latent representations that are often directly related to unknown and not directly measured physics (Bakarji et al, 2022;Champion et al, 2019;Chen et al, 2019). Several methods are available in open source python notebooks (Demo et al, 2018;Kaptanoglu et al, 2021), making their use and implementation in a discovery pipeline efficient and advantageous.…”

Dimensionally Consistent Learning with Buckingham Pi

“…More recently, machine learning algorithms have shown promise in discovering scientific laws [18], differential equations [19][20][21], and deep network inputoutput function approximations [22][23][24] from simulation and experimental data alone, with considerable progress in the field of fluid mechanics [25][26][27][28]. Particularly, unsupervised learning techniques have recently made significant progress in distilling physical data into interpretable dynamical models that automate the scientific process [29][30][31][32][33]. However, these methods do not take into account the units of their training data, which can significantly constrain the hypothesis class to physically meaningful solutions.…”