Enhancing Localized Deformation Analysis in Materials Science Using AI/ML: A Comparative Study of Traditional DIC and AI/ML Approaches

Materials typically fail at their weakest points. Understanding the formation, evolution, and micromechanical behavior of these weak points under mechanical load is critical in both engineering applications and the development of advanced materials. Resolving the localized deformation and strain processes provides insights into these micromechanisms, enabling the design of stronger, more resilient materials and the development of predictive models for material behavior.

In-situ Scanning Electron Microscopy (SEM) allows real-time observation of material deformation at submicron resolution. This high level of detail provides crucial insights into microstructural features, such as cracks and voids, and how they evolve under load. SEM is particularly important for understanding the behavior of advanced materials, enabling the study of localized strain that precedes failure, thus informing the design of more reliable materials. For this project, we will be able to provide database of collected in-situ SEM data. The students may even have opportunities to design their own experiments and perform on that on our SEM.

Objectives

1. Develop Traditional DIC Using Conventional Image Recognition Techniques: 

Implement traditional DIC methods to track surface features between sequential SEM images. This will serve as the baseline method for measuring deformation and strain.

2. Compare Traditional DIC with AI/ML-Aided DIC:

Enhance traditional DIC with AI/ML techniques to improve feature recognition and tracking. This objective focuses on addressing the limitations of traditional DIC, such as feature deformation and noise, using machine learning models.

3. AI/ML-Based Strain Map Generation Directly from SEM Data:

Explore the use of AI/ML models to generate strain maps directly from raw SEM images, bypassing traditional pixel-tracking methods. This advanced approach will allow faster and more accurate analysis, particularly in challenging conditions where traditional DIC may fail.

Deliverable

A software with UI will be expected to generate strain map from in-situ SEM data using traditional and AI/ML based techniques. 

Motivations

Traditional DIC and Its Limitations: Digital Image Correlation (DIC) is a common method for analyzing deformation by tracking surface features between sequential images. While effective in many applications, DIC has limitations when applied to in-situ SEM. It relies on distinct surface features, which may be faint or unstable in SEM images. Additionally, noise and the small field of view at high magnification can hinder the accuracy of traditional DIC methods, particularly at submicron scales. These challenges necessitate more robust approaches for fine-scale deformation analysis.

AI/ML Solutions for DIC: Artificial Intelligence (AI) and Machine Learning (ML) can overcome the limitations of traditional DIC by learning to recognize complex, evolving features in noisy SEM data. AI/ML approaches, such as Convolutional Neural Networks (CNNs), can adaptively track microstructural changes, even when the features deform or disappear, providing more robust and accurate displacement field generation. Unlike DIC, which depends on pixel-based correlations, AI/ML models can predict how features behave over time, improving accuracy and processing speed for strain measurements.

Qualifications

Minimum Qualifications:
  • Be familiar with a coding language
  • Self-motivation

Preferred Qualifications:
  • Be familiar with GPU calculation and/or numerical methods for PDEs  


Details

Project Partner:

Tianyi Chen

NDA/IPA:

No Agreement Required

Number Groups:

2

Project Status:

Accepting Applicants

Keywords:
Machine Learning (ML)GPUMolecular DynamicsConsultancyMachine Learning (ML)
Card Image Capstone