Data-driven and AI-assisted Microscopy Analysis

Turning high-dimensional microscopy data into interpretable spectral, structural, and compositional insight.

Modern transmission electron microscopes (TEM) produce increasingly large and high-dimensional datasets, including three-dimensional spectrum imaging and four-dimensional scanning transmission electron microscopy (4D-STEM) measurements. At each probe position, the recorded signal may consist of a spectrum containing thousands of energy channels or a diffraction pattern containing tens of thousands of detector pixels. Extracting interpretable information from these datasets remains challenging because of limited signal-to-noise ratios, particularly under low-dose conditions, as well as the computational demands and subjectivity associated with conventional analysis. I develop data-driven approaches that make those datasets more interpretable while preserving a clear connection to the underlying physics.

Visual highlights

DRCA workflow from high-dimensional microscopy data to clustered features
From multidimensional microscopy data to interpretable feature maps through dimensionality reduction and clustering.
Non-negative matrix factorization analysis of 4D-STEM data
A data-driven route from 4D-STEM measurements to interpretable components.
Architecture of a one-dimensional convolutional autoencoder for spectrum data
A one-dimensional convolutional autoencoder learns a compact representation of microscopy spectra.

Selected outputs