Paper Number
AM8 

Session
AI and ML Based Rheological Characterization

Title
Automated classification of shear induced relaxation dynamics from in situ rheology and x-ray photon correlation spectroscopy

Presentation Date and Time
October 10, 2022 (Monday) 1:50

Track / Room
Track 6 / Mayfair

Authors (Click on name to view author profile)

  1. Horwath, James P. (Argonne National Laboratory, Advanced Photon Source)
  2. He, HongRui (Argonne National Laboratory, Materials Science Division and Center for Molecular Engineer)
  3. Zhang, Qingteng (Argonne National Laboratory, Advanced Photon Source)
  4. Chu, Miaoqi (Argonne National Laboratory, Advanced Photon Source)
  5. Dufresne, Eric (Argonne National Laboratory, Advanced Photon Source)
  6. Sankaranarayanan, Subramanian (University of Illinois, Chicago, Department of Mechanical and Industrial Engineering)
  7. Narayanan, Suresh (Argonne National Laboratory, Advanced Photon Source)
  8. Cherukara, Mathew (Argonne National Laboratory, Advanced Photon Source)

Author and Affiliation Lines (in printed abstract book)
James P. Horwath1, HongRui He1, Qingteng Zhang1, Miaoqi Chu1, Eric Dufresne1, Subramanian Sankaranarayanan2, Suresh Narayanan1 and Mathew Cherukara1
1Advanced Photon Source, Argonne National Laboratory, Lemont, IL 60439; 2Department of Mechanical and Industrial Engineering, University of Illinois, Chicago, Chicago, IL 60607

Speaker / Presenter 
Horwath, James P.

Keywords
experimental methods; computational methods; colloids; glasses; jammed systems; ML based; spectroscopy

Text of Abstract
X-ray photon correlation spectroscopy (XPCS) is a useful technique for characterizing the dynamics of evolving systems and has been used successfully in combination with rheology measurements (rheo-XPCS) to observe the relaxation of complex fluids under shear in situ. However, out-of-equilibrium dynamics can produce a variety of unique and complex two-time correlation patterns which makes quantification of dynamics, or even establishing qualitative relationships between samples, extremely difficult. Meanwhile, machine learning and computer vision provide a wide range of unsupervised techniques for processing and understanding data without requiring input from the users, which can be applied to scientific data. We have developed an unsupervised variational autoencoder (VAE) which takes raw XPCS correlation data as input, encodes this data into a feature-rich latent representation, and attempts to reproduce the input based on this compressed representation. After training the model to accurately reproduce experimental data, we are able to cluster the latent representation into distinct classes. At first pass each class corresponds to a range of relaxation rates, however the algorithm also detects small groupings of anomalous data which represent unique relaxation behavior. We then apply our method to a new set of rheo-XPCS data and demonstrate that we can correlate our classes of dynamic behavior with features in the time-dependent rheological response. In another application, we show how this automated clustering method can take user-specified data and suggest other experimental conditions and timestamps which have produced similar image features. Moving forward, we expect to utilize this type of automated data analysis to compare with physical models and enable direct interpretation of experimental XPCS data.