|Time:||July 15, 2022, 2:00 p.m. (CEST)|
|Download as iCal:||
When: 15 July 2022
Where: V 47.04, Pfaffenwaldring 47, Campus Vaihingen
Time: 2 - 5 pm
We cordially invite you to join us for the three inaugural lectures of
- Mathias Niepert on "Machine Learning with Discrete Structures and Algorithms" from 2-3 pm
- Felix Fritzen on "Physics-infusion in Data-Driven Mechanics" from 3-4 pm
- Marc-André Keip on "Smart, Soft, Solid -- Structures, Scales, Stability" from 4-5 pm.
All lectures will be held at V 47.04 at Pfaffenwaldring 47.
Machine Learning with Discrete Structures and Algorithms
Machine learning at scale has led to impressive results ranging from text-based image generation, reasoning with natural language, and code synthesis to name just a few of the most recent applications. The machine learning approaches driving these methods are also successfully applied to a broad range of problems in the simulation sciences. Due to their impressive results, these developments make some of us question the need for incorporating prior knowledge in form of symbolic (discrete) structures and algorithms. Is compute and data at scale all we need? We will make an argument that incorporating discrete (symbolic) structures and algorithms in machine learning models is indeed advantageous in numerous application domains such as Biology, the Material Sciences, and Physics, where data is often more suitably represented with graphs and on irregular grids. Biomedical networks modeling genes, drugs, and their side effects, for example, can be represented as a multi-relational and heterogeneous graph. My group’s research is concerned with the development of machine learning methods that use discrete structures such as graphs. We also address the problem of learning and leveraging structure from data where it is missing, combining discrete algorithms and probabilistic models with continuous gradient-based learning. We will show that discrete structures and algorithms appear in numerous unexpected places such as ML-based PDE solvers and that modeling them explicitly is indeed beneficial and sometimes even required. Especially machine learning models with the aim to exhibit some form of explanatory properties have to rely on symbolic representations. The talk will also cover some biomedical and SimTech-related applications and future research directions.
Physics-infusion in Data-Driven Mechanics
Data-driven models, machine-learned surrogates and artificial intelligence are omnipresent today (for a reason): The illusion of a black-box model that can discover relations of arbitrary complexity from automatic data processing is appealing. Throwing more data to the model is expected to improve the performance, of course! And adding more parameters will help in solving today's problems with ease while shaping the path to realizing tomorrow's brave new world. After a few hard lessons, some set-backs and a steep learning curve, after facing fighting technical issues, discovering a lack of quality data and facing unforeseen complications, large parts of the scientific community admit that, presumingly, a balance point between a data-centered and a model-centered approach might be a more reasonable compromise. Recently, people have started to infuse well-established mathematical and physical properties into the machine learning process. By looking at the relevant class of hyper-elastic material models we will revisit some of the physical principles, their implications and data-driving modeling strategies that can lead to more reliable Data-Integrated Simulation Science.
Smart, Soft, Solid -- Structures, Scales, Stability
We discuss some of the group's results in the field of chemo-magneto-electro-mechanical coupling of smart and soft solids based on continuum theories and corresponding numerical simulations. As a model problem we consider the behavior of magnetorheological elastomers (MREs) across several length scales. MREs are soft solids that are composed of an elastomer matrix and magnetizable inclusions. We investigate the micro-, meso- and macroscopic response of MREs with a particular focus on the evolution of magnetic microstructures, the homogenization of micro-magnetically informed mesostructures, and the analysis of material as well as structural instabilities across scales. The view is then extended towards the computational multiscale simulation of porous media, where we analyze in particular the creation of wrinkling patterns of hydrogel thin films.