TriCAMS

DUKE NCCU NC State UNC

Conference Schedule

Saturday, 9/17
Sunday, 9/18

Saturday, September 17, 2022

9:30 - 10:00 am EDT	Welcome Coffee
10:00 - 10:15 am EDT	Welcoming Remarks SAS 2203 Alina Chertock, North Carolina State University
10:15 - 11:05 am EDT	On Inverse Problems, Digital Twins and Model Based Design SAS 2203 Speaker John Burns, Virginia Tech University Session Chair Hien Tran, North Carolina State University Abstract We discuss modeling and numerical issues involved with constructing finite dimensional mathematical physics based models that can be used for control, optimization and design of infinite dimensional systems. We focus on the approximate-then-design paradigm that is key to many new model based design and digital twins technologies. These methods are rapidly becoming the standard approach to model based system engineering and is enabled by a growth of software environments for model creation, simulation and post-processing. Roughly speaking, these software tools facilitate the construction of finite dimensional system level computational models by connecting (in software) finite dimensional component models. These component models are saved in a ``library'' of domain models which are linked by software to generate system level models. We consider the case where one or more of the physical components are described by partial or delay differential equations and the process of building a finite dimensional component models involves some type of discretizations. Using this process to develop system level computational models that are suitable for simulation, optimization and control leads to modeling and approximation requirements that are more stringent than models to be used only for simulation. Examples are used to demonstrate how problems can arise if these issues are ignored. In particular, we focus on inverse, optimization and optimal control problems for some thermal fluids systems. The goal of this talk is not to present a detailed discussion of the latest research accomplishments in this area, but rather to raise the awareness of new trends in model based system engineering and to point out some numerical issues, challenges and opportunities for research. Applications to thermal fluid systems and delay equations are used to motivate and illustrate the importance of consistent approximations, dual convergence, parametric smoothness and numerical conditioning.
11:10 am - 12:00 pm EDT	Energy Landscapes, Metastability, and Transition Paths SAS 2203 Speaker Katie Newhall, University of North Carolina, Chapel Hill Session Chair Greg Forest, University of North Carolina, Chapel Hill Abstract The classic example of metastability (infrequent jumps between deterministically-stable states) arises in noisy systems when the thermal energy is small relative to the energy barrier separating two energy-minimizing states. My work seeks to extend this idea to infinite dimensional systems and systems with non-gradient forces, extending the usefulness of the underlying energy landscape in the classic metastability analysis. Such example systems are a spatially-extended magnetic system with spatially-correlated noise designed to sample the Gibbs distribution relative to a defined energy functional, and a polymer bead-spring model of chromosome dynamics with additional stochastically-binding proteins that push the system out of equilibrium.
12:00 - 1:30 pm EDT	Lunch/Free Time
1:30 - 2:20 pm EDT	Decomposed Convolutional Deep Networks: on Graphs, Groups, and Across Domains SAS 2203 Speaker Xiuyuan Cheng, Duke University Session Chair Hongkai Zhao, Duke University Abstract Deep convolutional neural networks (CNN) have been applied to data on Euclidean as well as non-Euclidean domains. In this talk, we introduce a framework of decomposing convolutional filters over a truncated set of basis, which applies to the standard CNN, graph neural networks, as well as group-equivariant networks. The basis decomposition reduces the model and computational complexities of deep CNNs with an automatically imposed filter regularity. On graphs, the decomposed convolution provides a unified framework for several graph convolution models, including both spectral and spatial constructions. Theoretically, the graph convolution with low-rank local filters provably enlarges expressiveness to represent graph signals more than spectral graph convolutions. Empirical advantages are demonstrated on facial expression and action recognition datasets. For group equivariant CNNs, a joint basis decomposition over space and group geometries achieves group equivariance with provable representation stability with respect to geometric deformations of input data. We introduce the methods to obtain rotation equivariant and scaling-group equivariant representations of image data. At last, when allowing the light-weighted basis layer to be adapted to varying modals in data, the decomposition therein provides an efficient way of invariant feature learning across domains and conditional image generation. Joint work with Qiang Qiu, Ze Wang, Zichen Miao, Wei Zhu, Robert Calderbank, and Guillermo Sapiro.
2:25 - 2:55 pm EDT	Coffee Break
3:00 - 4:30 pm EDT	Lightning Session SAS 2203 Session Chair Boyce Griffith, University of North Carolina, Chapel Hill
4:30 - 5:30 pm EDT	Panel Discussion on Opportunities for Faculty, Postdocs, and Graduate Students SAS 1102 Panelists James Beaudoin, Simulations Plus Fariba Fahroo, Air Force Office of Scientific Research (AFOSR) Mark Lovern, CERTARA Joe Myers, Army Research Office (ARO) Virginia Pasour, Army Research Office (ARO) Inma Sorribes, CERTARA Haiying Zhou, Simulations Plus Moderators Greg Forest, University of North Carolina, Chapel Hill Semyon Tsynkov, North Carolina State University
5:30 - 7:30 pm EDT	Social and Poster Session SAS 2nd Floor Atrium Poster Presenters Laryssa Abdala, University of North Carolina, Chapel Hill Anna Coletti, University of North Carolina, Chapel Hill Katherine Daftari, University of North Carolina, Chapel Hill Kristen Giesbrecht, University of North Carolina, Chapel Hill Shiying Li, University of North Carolina, Chapel Hill K. Medlin, University of North Carolina, Chapel Hill Michael Redle, North Carolina State University Edward Terrell, University of North Carolina, Chapel Hill David Wells, University of North Carolina, Chapel Hill Kristen Windoloski, North Carolina State University Lewei Zhao, Beaumont Health System, Michigan See all Abstracts Here!

Sunday, September 18, 2022

9:00 - 9:30 am EDT	Welcome Coffee
9:30 - 10:20 am EDT	Forecasting Weapon Interceptions for Columbia SAS 2203 Speaker Gaolin Milledge, North Carolina Central University Session Chair Kimberly Weems, North Carolina Central University Abstract We explored various methods to make predictions for weapon interceptions for Colombia. The methods that we have experimented include ETS, ARIMA, STL, NNAR, TBATS and Ensemble methods. STL (Seasonal and Trend decomposition using Loess) has proved to be most effective in both training and testing for this type of data. The forecasting is reasonable for year 2019. However, no method works well to make predictions for year 2020 data. The pandemic might be one of the reasons that 2020 weapons interceptions is hard to predict.
10:25 - 11:15 am EDT	Synergies between Uncertainty Quantification, Sensitivity Analysis, and Control Design for Physical and Biological Models SAS 2203 Speaker Ralph Smith, North Carolina State University Session Chair Alina Chertock, North Carolina State University Abstract The fields of control theory, sensitivity analysis, and uncertainty quantification have many synergistic objectives, yet they are often investigated in isolation. For example, robust control design requires quantification of plant errors, and feedback to accommodate unmodeled dynamics. Addressing model discrepancy also constitutes one of the primary challenges of uncertainty quantification and is an area where deep learning can serve a fundamental role for applications with sufficient data. Furthermore, various sensitivity and identifiability techniques, employed in uncertainty quantification to determine influential parameters, have their origin in the control literature. This presentation will focus on topics which jointly arise in uncertainty quantification and control theory, with examples from nuclear engineering, smart material systems, and quantitative systems pharmacology (QSP). Topics will include Bayesian techniques for parameter inference, aspects of sensitivity analysis, physics-informed quantification of model discrepancy terms, and the construction of reduced-order or surrogate models for complex systems. The presentation will conclude with discussion regarding the role of digital twins and virtual populations for engineering and biological systems.
11:15 - 11:30 am EDT	Coffee Break
11:30 am - 1:00 pm EDT	Memorial Session for Tom Banks SAS 1102
1:00 - 2:30 pm EDT	Lunch Reception in SAS Hall SAS 2nd Floor Atrium

Website Templates

created with

Website Builder Software .