MS039 - Deep Learning-based Reduced Order Models in Scientific Computing
Keywords: approximation theory, neural networks, parametrized PDEs, reduced order modeling, scientific machine learning
Solving differential equations using full order models (FOMs) incurs significant computational costs, especially for real-time and many-query applications. To address this challenge, reduced order models (ROMs) have emerged as a crucial framework for generating efficient and reliable approximations, which are essential for both academic and industrial applications.
The increasing demand for efficient non-intrusive methods, coupled with the availability of large amounts of data from measurements or simulations, has spurred the development of data-driven techniques based on deep learning strategies for complexity reduction in computational science. These approaches leverage state-of-the-art machine learning algorithms capable of extracting previously unseen patterns inherent in the data. They offer novel solutions for addressing the challenges of reduced order modeling across various scientific domains.
Integrating data-driven and physics-based approaches, and investigating their approximation properties, enhance the modeling capabilities and the interpretability of these tecnhiques, enabling consistent and accurate predictions even for complex systems. This integration has given rise to several research lines that combine traditional ROMs with scientific machine learning.
This mini-symposium aims to bring together researchers actively involved in the theory, methods, and applications of data-driven methods for complexity reduction. The areas of interest include, but are not limited to, deep-learning-based reduced order modeling, approximations using neural operators, data-driven discovery of system dynamics, physics-informed deep learning, hybrid modeling, multi-fidelity methods, as well as the approximation and mathematical properties of neural networks.
The increasing demand for efficient non-intrusive methods, coupled with the availability of large amounts of data from measurements or simulations, has spurred the development of data-driven techniques based on deep learning strategies for complexity reduction in computational science. These approaches leverage state-of-the-art machine learning algorithms capable of extracting previously unseen patterns inherent in the data. They offer novel solutions for addressing the challenges of reduced order modeling across various scientific domains.
Integrating data-driven and physics-based approaches, and investigating their approximation properties, enhance the modeling capabilities and the interpretability of these tecnhiques, enabling consistent and accurate predictions even for complex systems. This integration has given rise to several research lines that combine traditional ROMs with scientific machine learning.
This mini-symposium aims to bring together researchers actively involved in the theory, methods, and applications of data-driven methods for complexity reduction. The areas of interest include, but are not limited to, deep-learning-based reduced order modeling, approximations using neural operators, data-driven discovery of system dynamics, physics-informed deep learning, hybrid modeling, multi-fidelity methods, as well as the approximation and mathematical properties of neural networks.
