Paper Number
ML18                         My Program 

Session
AI and ML in Rheology

Title
Rheometric signal denoising using latent space modelling

Presentation Date and Time
October 23, 2025 (Thursday) 9:25

Track / Room
Track 6 / Sweeney Ballroom C

Authors (Click on name to view author profile)

  1. Das, Mohua (Massachusetts Institute of Technology)
  2. Vadillo, Damien C. (Corporate Research Analytical Laboratory, 3M)
  3. Perego, Alessandro (Corporate Research Analytical Laboratory, 3M)
  4. McKinley, Gareth H. (Massachusetts Institute of Technology, Mechanical Engineering)

Author and Affiliation Lines (in printed abstract book)
Mohua Das1, Damien C. Vadillo2, Alessandro Perego2 and Gareth H. McKinley1
1Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139; 2Corporate Research Analytical Laboratory, 3M, St Paul, MN, MN 55144

Speaker / Presenter 
Das, Mohua

Keywords
experimental methods; artificial intelligence; machine learning; rheometry

Text of Abstract
In rheometry, precise evaluation of the viscoelastic properties of materials depends on accurately capturing undistorted response signals, such as strain (or angular displacement) and stress (or torque). However, these signals are frequently affected by substantial noise originating from the rheometer hardware itself or from external sources. Unlike many traditional filtering problems, in rheometry it is essential to accurately determine both the phase information and the amplitude of the measured signal. Conventional signal denoising methods, including moving average filters, Savitzky-Golay filters, Wiener filters, and wavelet-based techniques, often require manual parameter tuning for each signal. This not only makes the process labor-intensive and time-consuming but can also result in the loss of important signal features, especially in the frequency domain. To address these limitations, we introduce a novel data-driven denoising framework based on latent space modeling techniques. By learning a compact representation of the underlying noise-free or ‘undistorted’ signal, the model effectively separates noise from meaningful content while preserving critical features across both time and frequency domains. The architecture leverages contemporary deep learning techniques—such as encoder-decoder structures and generative models—to learn mappings from noisy inputs to clean outputs using paired training data. Preliminary results demonstrate that the proposed approach outperforms traditional denoising techniques, effectively reducing noise while preserving signal integrity in both time and frequency domains. Moreover, the method offers a significant advantage over classical techniques by automatically adapting to the specific characteristics of the measured signal and requires minimal manual tuning. This scalable, automated approach offers a robust solution for processing noisy rheometric datasets, facilitating more precise determination of complex material behavior.