|
Technical Thrust Area
Nanotechnology and Lower Scale Phenomena Committee
Chair: Anter El-Azab, Purdue University
Vice-Chair: Jaroslaw Knap, U.S. Army Research Laboratory
Amartya Banerjee, University of California Los Angeles Members Webinar SeriesOctober 26, 2022 Speaker: Stefanie Reese, RWTH Aachen University Title: Data-Driven Mechanics for Elastic and Inelastic Problems Including Uncertainty Quantification The data-driven computing paradigm for mechanical systems as proposed in [1] is further extended. Its main principle, to define the solution of boundary value problems explicitly in the material data has several advantages. While generally accepted physical principles such as conservation laws and thermodynamic constraints are enforced, any functional modelling of the material data is avoided. The open-ended process of modelling and calibration as the material data set grows is thus circumvented. Model uncertainties regarding the choice of functions, proper usage within its range of validity and loss of information are thus minimized. One main challenge is the extension to inelastic material behaviours, also included in the talk. The approach of a previous work [2] is extended by material-independent thermodynamic constraints, involving the Helmholtz free energy as first principal data, which may be obtained by increasingly accurate low-scale simulations. Further important aspects are the data search [3] as well as the uncertainty quantification [4]. The talk will be concluded by structural examples. [1] Kirchdoerfer, T., & Ortiz, M. (2016). Data-driven computational mechanics. Computer Methods in Applied Mechanics and Engineering, 304, 81-101. Join via Zoom: https://us06web.zoom.us/j/84541274013?pwd=ZDhuM09BaUlVVlBuREhEV0sweGFlZz09 October 12, 2022 Speaker: Steve WaiChing Sun, Columbia University Title: Geometric Learning for Discovering Complex Material Behaviors of Microstructures Plasticity models often include a scalar-valued yield function to implicitly represent the boundary between elastic and plastic material states. However, a surface can also be represented explicitly by a manifold of which the tangential space is Euclidean. In this work, we introduce the concept of yielding manifold. This yielding manifold is reconstructed by stitching local multi-resolution patches together. This treatment enables us to construct a highly complex and precise yield envelope by breaking it down into multiple coordinate charts. The global atlas is then built to enforce the consistency of the machine learning generated yield surface. In contrast to data from smooth manifolds, data collected from sensors and numerical simulations are often discrete. As such, we propose to use graph embedding to create latent space inferred from 3D microstructural data represented via weighted graphs. Vectors of this latent space are used as both geometrical descriptors and internal variables of which a decoder can be used to convert predictions in the latent space back as a weight graph. This approach establishes a direct connection between nonlinear dimensional reduced simulations with machine learning constitutive models. Potential applications of geometric reinforcement learning for inverse problems and the design of experiments in the limited data regime will also be discussed. August 31, 2022 Speaker: Michael Shields, Johns Hopkins University Machine Learning (ML) and Uncertainty Quantification (UQ) have gained widespread popularity in the scientific community, to such an extent that ML/UQ seems to appear in some capacity in nearly all modern scientific investigations. This is particularly true in physics-based modeling where machine learning algorithms are being specially designed to adhere to physical principles, such as the popular physics-informed neural networks (PINNs). In this talk, we will discuss the relationship between these two important research areas in the context of physics-based modeling, with an emphasis on simulating materials systems. We will specifically discuss how the two areas complement one another to enhance modeling capability by making the critical distinction between UQ for ML and ML for UQ. In the former case, we will argue that modern ML methods require UQ as an integral component and show recent advances in the learning of uncertainty-aware Bayesian Neural Networks. In the latter case, we will argue that UQ can be viewed as an exercise in ML and will show how modern ML methods (from Hamiltonian Neural Networks to manifold learning) enable UQ in physical systems while, in many cases, remaining constrained by the underlying physics. Applications to materials modeling will be shown ranging from equation of state modeling for warm dense matter to hierarchical multi-scale models for structural mechanics. July 27, 2022 Speaker: Michael Falk, Johns Hopkins University Title: Questing for Structural Predictors of Plastic and Failure Response in Glasses For decades materials scientists, mechanicians and physicists have searched for structural predictors for plastic flow and failure in glasses and other amorphous materials. Due to the lack of any crystalline order in these materials, disorder rules the day on many scales. This makes quantifying the microstructures of amorphous materials difficult. Such quantification is necessary for building predictive theories that can guide materials design and the development of processing to improve mechanical properties. Here I will consider some particular case studies from my own work with collaborators: the use of machine learning to fit a constitutive model to molecular dynamics data, the use of computer modeling to quantify of the local yield surface in two- and three-dimensions, and an attempt to use an equation-free method to harvest simulation data for the quantification of plastic constitutive response in a 3D binary glass. Reflections on these efforts will be discussed in order to consider the prospects for harnessing machine learning for the development of physically interpretable structural characterization and materials response theories. June 29, 2022 Speaker: Markus Hütter, Eindhoven University of Technology Title: Molecular Approach to Plasticity in Polymer Glasses: A Journey A major benefit of multiscale modeling is that it helps to shed light on constitutive assumptions in macroscopic approaches to mechanics [1]; the stress tensor and the rate of plastic deformation are of particular interest. Nonequilibrium statistical mechanics is a powerful technique in this field, and it has been applied to study the plastic deformation of solids [2,3]. In this presentation, the focus is on multiscale modeling of solids in the glassy state. Structural glasses are particularly interesting (and challenging) for two reasons: they can age in the course of time, and the microscopic carriers of plastic deformation have not been identified to date (in contrast to the well-known dislocations for crystalline materials). Molecular simulations will be used to study glassy materials, in particular polymers. The goal is to establish a fine level of description that is suitable for a subsequent coarse-graining step to the macroscopic continuum level. May 25, 2022 Speaker: Alejandro Strachan, Purdue University;
Discussant: Arun Mannodi Kanakkithod, Purdue University
Data Using Machine Learning
The synergy between principles-based modeling and data science is playing an increasingly important role in materials science and engineering. In addition, there is significant interest in using machine learning (ML) tools to extract physical laws as well as symmetries and associated invariants from data. I will discuss recent progress in my group on machine learning applied to multiscale modeling and in the development of interpretable models that balance accuracy with parsimony. 1)Yoo P, Sakano M, Desai S, Islam MM, Liao P, Strachan A. Neural network reactive force field for C, H, N, and O systems. npj Computational Materials. 2021 Jan 22;7(1):1-0. Discussion Forum
|