Recurrent Deep Kernel Learning of Dynamical Systems

Nicolò Botteghi,Paolo Motta,Andrea Manzoni,Paolo Zunino,Mengwu Guo

2024-10-08

Abstract:Digital twins require computationally-efficient reduced-order models (ROMs) that can accurately describe complex dynamics of physical assets. However, constructing ROMs from noisy high-dimensional data is challenging. In this work, we propose a data-driven, non-intrusive method that utilizes stochastic variational deep kernel learning (SVDKL) to discover low-dimensional latent spaces from data and a recurrent version of SVDKL for representing and predicting the evolution of latent dynamics. The proposed method is demonstrated with two challenging examples -- a double pendulum and a reaction-diffusion system. Results show that our framework is capable of (i) denoising and reconstructing measurements, (ii) learning compact representations of system states, (iii) predicting system evolution in low-dimensional latent spaces, and (iv) quantifying modeling uncertainties.

Machine Learning

What problem does this paper attempt to address?

The paper attempts to address the problem of constructing efficient Reduced-Order Models (ROMs) to accurately describe the complex dynamics of physical assets in the presence of high-dimensional and noisy data. Specifically, the research team proposes a data-driven, non-intrusive method that utilizes Stochastic Variational Deep Kernel Learning (SVDKL) to discover low-dimensional latent spaces from the data and represents and predicts the evolution of these latent dynamics through a recursive version of SVDKL. The main contributions of the paper include: 1. **Improved Prediction Accuracy**: By using recurrent neural networks to model latent dynamics, the long-term prediction accuracy and consistency of the SVDKL framework are improved. 2. **Handling Complex Systems**: The framework's capability is demonstrated on two challenging chaotic systems, namely the double pendulum and the reaction-diffusion system. 3. **Uncertainty Visualization**: An intuitive method is introduced to visualize and study uncertainties on latent variables by observing the standard deviation in the measurement space. In summary, the study aims to develop a non-intrusive ROM capable of handling high-dimensional and noisy data to enhance the accuracy and efficiency of digital twin technology in practical applications.

Recurrent Deep Kernel Learning of Dynamical Systems

Machine learning based digital twin for dynamical systems with multiple time-scales

Development of Self-Learning Kernel Regression Models for Virtual Sensors on Nonlinear Processes

$\textit{FastSVD-ML-ROM}$: A Reduced-Order Modeling Framework based on Machine Learning for Real-Time Applications

Machine learning based digital twin for stochastic nonlinear multi-degree of freedom dynamical system

Learning deep dynamical models from image pixels

Digital Twin Data Modelling by Randomized Orthogonal Decomposition and Deep Learning

Kernel-Based Learning of Stable Nonlinear Systems

Learning Nonautonomous Systems via Dynamic Mode Decomposition

Learning Latent Dynamics via Invariant Decomposition and (Spatio-)Temporal Transformers

Kernel Embedding Based Variational Approach for Low-dimensional Approximation of Dynamical Systems

Learning Deep Dissipative Dynamics

Deep Kernel Methods Learn Better: From Cards to Process Optimization

A parametric framework for kernel-based dynamic mode decomposition using deep learning

Linearly Recurrent Autoencoder Networks for Learning Dynamics

Learning Deep Dynamical Systems using Stable Neural ODEs

Automated Global Analysis of Experimental Dynamics through Low-Dimensional Linear Embeddings

Promises of Deep Kernel Learning for Control Synthesis

Linearly-Recurrent Autoencoder Networks for Learning Dynamics

Learning Stable Deep Dynamics Models

Probabilistic Decomposed Linear Dynamical Systems for Robust Discovery of Latent Neural Dynamics