The two most prominent, unsolved challenges in computer-assisted modeling of physical and engineering systems relate to complex/multiscale modeling and uncertainty quantification (UQ). Each subject poses significant challenges on its own, and few efforts have been undertaken to tackle them simultaneously (mainly in the area of weather prediction, and only very recently in, say, UQ of DFT computations). We argue that a concerted approach is not only of paramount importance in making progress in many disciplinary applications (as multiscale problems are inherently characterized by uncertainty) but also necessary in enabling a new era in simulation- based analysis and design - a new approach to modeling, which functionally integrates our cur- rent and developing modeling tools with machine learning/data mining. PROMISe combines the complementary expertise of the participating focus groups in these areas, in order to address fundamental, long-standing modeling issues, such as:
- What are the appropriate variables for the coarse-grained description? How many are needed, and how do these relate with the fine-scale degrees-of-freedom (what in our language are “restriction” and “lifting” maps)?
- What is the appropriate form of coarse-grained models (e.g (non)local, stochastic, what order spatial derivatives, how much memory etc)?
- What is the predictive ability of the coarse-grained models i.e. how much information is lost and how does this manifest itself in uncertainty with regards to the predictions?
- What are the “missing physics” in the coarse-grained models currently employed and when/how do these affect our predictive ability?
- How predictive can coarse-grained descriptions be in problems without apparent scale separation or even with cascades of different scales?
- How can quantifying predictive uncertainty in coarse-grained models can lead to adaptive schemes?
Organization of the minisymposium on "Towards Data-driven, Predictive Multiscale Simulations" during the SIAM Uncertainty Quantification Conference in Lausanne, April 5-8 2016
Despite the continuous increase in our computational capabilities, the ultimate goal of predictive simulations remains elusive. The key challenges are: High-dimensionality of uncertainties; Information fusion, e.g., multi-fidelity, multi-scale/physics models, and experiments; Model-form uncertainties induced by limited data and incomplete physics; Cost of information acquisition, i.e., the cost of doing simulations/experiments. The purpose of this minisymposium is to address these roadblocks and achieve groundbreaking advances by promoting synergies between applied mathematics, computational physics, and data sciences. Specific topics include but are not limited to: Data-driven model identification; Learning from high-dimensional data; Task-specific information acquisition policies; Non-linear dimension- reduction for coarse graining. The minisymposium included 3 sessions and 12 talks in total.
The TUMQCD collaboration involves members from TUM, physics department T30f, BNL, Michigan State University, and Fermilab. Mission of the TUMQCD collaboration is to complement effective field theory methods and lattice techniques to calculate the properties of strongly coupled systems at zero and finite temperature.