Function space data representation of temporal signals for machine learning
Temporal signals emerge in material science from both experiments and computer simulations to understand a physical phenomenon. For example, it is often desirable to collect responses of materials for a particular loading that changes over time to elicit steady-state or equilibrium behavior. A characteristic signal is collected for each material of interest and analyzed to extract scientific knowledge about the material. In this talk, we will be discussing a particular characterization technique called cyclic voltammetry (CV) that characterizes a material in an electrochemical environment such as a battery cell. CV experiment involves sweeping over a time-varying voltage load applied to the electro-chemical cell and collecting current output as a temporal response. CV is a critical tool in characterizing and understanding behavior of potential battery materials for clean and efficient energy storage.
We will be using CV data collected from hundreds and thousands of experiments that needs to be analyzed in a multivariate fashion. Our goal is to compare different materials based on their respective CV signals and develop a method to classify materials based on the shape of the signal. Thus we are interested in the structure of the signal as represented by moment representation i.e. interactions between output and inputs jointly over time. In particular, we propose use of the Gaussian process (GP) as a means to identify function space representations of CV signals. GPs allows users to obtain mathematical representations of the data by selecting appropriate basis in the form of kernels (covariance functions) that can be leveraged to encode physics based constraints on the signals. This talk would cover underlying mathematics of GPs, function spaces and their applications to signal processing. Applying these methods on CV data of materials, we showcase how machine learning and signal processing can be combined to realize data driven material discovery and design.
Date and Time
If you are not a robot, please complete the ReCAPTCHA to display virtual attendance info.
- Starts 03 December 2020 03:26 PM
- Ends 18 January 2021 04:59 PM
- All times are America/New_York
- No Admission Charge
Kiran Vaddi is a PhD candidate at the department of Materials Design and Innovation at the University at Buffalo. He obtained his bachelors and master from Indian Institute of Technology Madras in mechanical engineering. His research interest is in using machine learning and applied mathematics for scientific discovery. His current focus is on building data representation for materials discovery exploiting the inherent topological and geometrical structure.