The proposed symposium will include plenary talks from approximately 15 internationally-renowned researchers in the areas of Uncertainty Quantification (UQ) and Computational Statistics/Machine Learning. Emphasis of the meeting is in identifying synergies and common themes for these communities and proposing innovative research directions that can accelerate the impact of uncertainty modeling in engineering and the sciences as well as demonstrate the capabilities of computational uncertainty quantification methods and tools in various problems.
The proposed symposium will bring together perspectives from scientists in computational physical modeling, computational mathematics, and computational statistics/machine learning communities, in view of problems related to data-driven uncertainty quantification. This objective is motivated by the common, fundamental challenges that are encountered both in the modeling highly complex, multiscale, multiphysics processes as well as data-driven sciences:
Big data: current computational models are capable of generating petabyte-sized databases providing insight into processes taking place at extremely fine spatio-temporal scales. Apart from the obvious difficulties in visualizing, storing and processing such data, the fundamental challenge lies in extracting knowledge about the salient features and understanding of the driving mechanisms. This is more pronounced in the context of multiscale simulations which can only be enabled by the ability to identify accurate but computationally-tractable reduced descriptions even in cases where physical insight is unavailable. Learning from data and identifying structural characteristics is a typical task in several computational statistics and machine learning applications. Coupling such tasks with physical models and enabling physical insight by leveraging computational statistics/machine-learning tools, will be one of the challenges addressed.
High-dimension: Physical models are characterized by high-dimensional inputs and outputs. Dimensionality reduction is generally a necessary step in enabling the simulation of such systems. Overcoming the curse of dimensionality is a common objective within the Computational Mathematics UQ community and Computational Statistics machine learning community. This is tightly connected with the ability to identify salient input-output features that enable the construction of predictive models even when limited data is available.
Uncertainty: Predictive modeling and uncertainty quantification are more than just another research direction relevant to science and engineering. They constitute a different way of thinking that impacts practically all aspects of scientific and engineering analysis and design. Rather than deriving deterministic answers to complex problems, distributions (error-bars) are obtained that account for our incomplete and often inaccurate information about the problems of interest. An important objective centers around the development of probabilistic frameworks for systems identification, model validation, analysis, design, optimization and control under uncertainty.
The proposed symposium will establish a bidirectional relationship between the aforementioned scientific communities. The development of a common language as well as the different perspectives by which common problems are addressed will enable significant advances in the methods and computational tools. Such progress will impact a broad range of engineering disciplines, e.g. multiscale materials engineering, fluid/structure interaction, engineering design, energy and the environment but also physical and biological sciences such as catalysis, condensed matter physics, systems biology, and bio-chemistry.
The lectures and discussions will be video-recorded and will be made publicly available through a dedicated webpage under the webportal of the TUM Institute for Advanced Studies.
Furthermore, a special issue in the Journal of Computational Physics is planned that will include review articles and specific works demonstrating the integration of the UQ/Computational Statistics areas and the potential synergies that can be developed. Elsevier will attach the video presentations to each online paper thus further increasing the impact of the special issue. Prof. Zabaras is an Associate Editor of JCP and will coordinate the publication and review process.
This event has been supported by the Air Force Office of Scientific Research, Air Force Material Command, USAF under Award No. FA9550-15-1-0185.