1804 Machine Learning and Reduced-order Models for Complex Systems

Maruti Mudunuru, Los Alamos National Laboratory
Satish Karra, Los Alamos National Laboratory
Velimir Vesselinov, Los Alamos National Laboratory
Junuthula Reddy, Texas A&M University
Kalyana Nakshatrala, University of Houston
Large and complex computational mechanics applications demand efficient methods for reducing computational costs by data- and model-reduction techniques extracting important features and exploiting underlying mathematical/physical structure. These are needed for many practical applications including, energy and global security, where dynamics of the analyzed system is composed of many interacting parts. This is particularly relevant for addressing important problems like real-time sensing, structural health monitoring, nuclear waste disposal, biomechanics, carbon capture and storage, hydraulic fracturing, adaptive control, enhanced oil recovery, and environmental remediation. To reduce the computational costs, strategies such as physics-informed dimensionality reduction, proper orthogonal decomposition, graph/network-based methods, supervised and unsupervised machine learning techniques such as support vector machines, neural networks, non-negative matrix factorization, blind source separation, and clustering analysis, to name a few, can be explored within the theme of this mini-symposium.

The mini-symposium will serve as a platform to exchange ideas by gathering scientists working in different fields with the common idea of using machine learning and/or model/solution/data reduction techniques for engineering applications. Topics of interest may involve but not limited to:
• Data-driven proxy models to replace detailed computational models for managing computational costs. 
• Designing efficient model-order reduction methods for systems that require repeated evaluation of complex numerical simulations with massive computational requirements.
• Machine learning techniques to identify changes in system response (under uncertainities) from either experimental/field data or high-fidelity numerical simulations. Examples include, machine learning techniques to detect damage, failure, flow structures, mixing patterns, crack propagation, new subsurface signals etc.
• Inferential procedures for linking models to data for evaluating reliability and performance of complex systems. 
• Solution reduction methods and surrogate models for inverse problems like parameter estimation, structural optimization, signal detection, and signal discrimination.
• Network analysis and graph-based models for geosciences, biomechanics, and structural health monitoring applications. Example include, new trends in modeling flow in networks of pipes, fractures, and blood vessels.