Machine Learning for High-Dimensional Microscopy Data

We build uncertainty-aware analysis tools that turn EELS and multidimensional electron microscopy measurements into quantitative maps of nanoscale electronic properties.

This research connects electron nanoscopy, spectroscopy, and machine learning to understand how local structure controls functionality in low-dimensional and van der Waals materials.

Visit EELSfitter View publications

Why this matters

The bottleneck is no longer only acquiring microscopy data, but interpreting it reliably. Weak spectral signatures can be hidden by background, noise, and user-dependent fitting choices. We develop analysis workflows that make these measurements more quantitative, reproducible, and uncertainty-aware.

From high-dimensional microscopy data to quantitative maps

Our lab uses electron nanoscopy to probe how local structure controls functionality in quantum and low-dimensional materials. Techniques such as electron energy-loss spectroscopy, 4D-STEM, and spatially resolved imaging provide access to electronic excitations, bonding, dielectric response, strain, defects, and nanoscale heterogeneity.

The challenge is that the most informative signals are often the hardest to extract. In low-loss EELS, band-gap and excitonic features can be masked by the zero-loss peak. In core-loss spectroscopy, quantitative edge analysis depends sensitively on background subtraction. These problems become especially limiting in spectrum images, where thousands of spectra must be analysed consistently.

We address this by developing machine-learning methods that combine flexible data-driven modelling with physical constraints and uncertainty propagation. Instead of returning only a single fitted curve, our workflows estimate statistically consistent solutions and propagate their uncertainty to the final observables, enabling quantitative maps of band gaps, excitonic features, dielectric response, ionisation-edge intensities, and spatial variations in electronic structure.

EELSfitter

Open-source framework for uncertainty-aware EELS analysis

EELSfitter is our open-source Python framework for the analysis and interpretation of electron energy-loss spectroscopy measurements in transmission electron microscopy. It adapts machine-learning strategies originally developed in high-energy physics to nanoscale spectroscopy in materials science.

The framework supports model-independent background subtraction, uncertainty estimation, and quantitative mapping of electronic properties from low-loss and core-loss EELS datasets.

Visit EELSfitter documentation

Selected publications