Joint Applied Algebraic Topology Research Network/Statistical Thermodynamics & Molecular Simulations (AATRN/STMS) Series
These seminar series during October 2022 are a collaboration between the Applied Algebraic Topology Research Network (AARTN) and the Statistical Thermodynamics and Molecular Simulations (STMS) seminar series. Each event will feature two speakers, one from the STMS community and the other from the AARTN community. Both speakers, however, have interests within the purview of the other community, i.e., mathematicians who work on problems that are relevant to theoretical and computational chemistry, and statistical thermodynamicists who develop and use tools from applied mathematics. As such, the intent is to enhance discussions and collaborations between these two communities. Each seminar will comprise of two 25-minute talks (one from each community) followed by questions and discussions. These four events will all take place on the four Fridays of October 20222, from 10:45 AM-12:00 PM Eastern Standard Time (EST):
This event's talks:
Graph neural network: Representational power and limitations
Prof. Yusu Wang (University of California, San Diego)
Abstract: Graph neural networks and (high order) variants provide a flexible and powerful framework for graph analysis. However, it is also known that GNNs have limitation in terms of representation power. At the same time, it is important to understand such limitations so as to use GNNs or variants in more informed ways in practice. In this talk, I will survey some of the results in this direction.
Speaker Bio: Yusu Wang is currently Professor in the Halicioglu Data Science Institute at University of California, San Diego. She is Associate Director for Research for the NSF funded AI Institute TILOS. Prior to joining UCSD, she was Professor in the Computer Science and Engineering Department at the Ohio State University. She obtained her PhD degree from Duke University in 2004, where she received the Best PhD Dissertation Award at the Computer Science Department. From 2004-2005, she was a post-doctoral fellow at Stanford University. Yusu Wang primarily works in the fields of Computational geometry, and Computational and applied topology. She is particularly interested in developing effective and theoretically justified algorithms for data analysis using geometric and topological ideas and methods, as well as in applying them to practical domains, including computational neuroscience and material science. She received DOE Early Career Principal Investigator Award in 2006, and NSF Career Award in 2008. Her work received several best paper awards. She is on the editorial boards for SIAM Journal on Computing (SICOMP) and Journal of Computational Geometry (JoCG). She is currently a member of the Computational Geometry Steering Committee, as well as a member of AATRN Advisory Committee. She also serves in SIGACT CATCS and AWM Meetings Committee.
Combining molecular dynamics simulations and machine learning to quantify interfacial hydrophobicity
Prof. Reid Van Lehn (University of Wisconsin, Madison)
Abstract: Classical molecular dynamics (MD) simulations generate high-dimensional datasets consisting of 103-106 atomic positions at 105-107 distinct timesteps. These datasets encode spatial and temporal correlations between molecules that are physically meaningful when analyzed using statistical mechanics, but such approaches are time-consuming and currently require significant human intervention. In this talk, I will describe machine learning (ML) approaches to efficiently analyze the output of MD simulations. As a representative case study, I will focus on understanding the hydrophobicity of functionalized interfaces. Hydrophobicity is a key property that influences behavior in aqueous environments but is challenging to predict for chemically heterogeneous interfaces – i.e., interfaces that have nonpolar and polar groups in close (~nm) proximity – which are abundant in biological and industrial systems. I will show how ML can extract dominant features from MD-derived solvent configurations to quantify hydrophobicity using significantly less MD data than conventional simulations. I will also demonstrate how the transformation of MD output into various data representations influences the selection of ML method and corresponding prediction accuracy. Finally, I will show that the topological analysis of MD output can outperform more complex ML models, highlighting future opportunities for the intersection of topological data analysis with MD.
Speaker Bio: Reid Van Lehn is the Hunt-Hougen Associate Professor in the Department of Chemical and Biological Engineering at the University of Wisconsin-Madison. He received his Ph.D. in Materials Science and Engineering from MIT under the supervision of Prof. Alfredo Alexander-Katz, then performed research as a NIH Ruth-Kirschstein postdoctoral fellow with Prof. Tom F. Miller III at Caltech. He joined UW-Madison in May 2016, where his group develops and applies molecular simulation methods to characterize, predict, and engineer the physicochemical properties of synthetic and biological soft materials.