Machine learning-based prediction of material properties from position averaged convergent beam electron diffraction

NAISS 2023/5-44


NAISS Medium Compute

Principal Investigator:

Magnus Röding


Chalmers tekniska högskola

Start Date:


End Date:


Primary Classification:

10106: Probability Theory and Statistics

Secondary Classification:

10304: Condensed Matter Physics

Tertiary Classification:

10105: Computational Mathematics




This proposal is made in connection to the projects "Microstructures and mass transport - a machine learning approach" (Formas, grant no 2019-01295) and "Visualizing the dynamics of strong light-matter interactions using NEX-GEN-STEM" (VR, grant no 2020-04986). Although the PI already has a current small compute project on Alvis (SNIC 2022/22-961), this is a new research direction that was not in the pipeline during the writing of earlier proposals, and initial tests have made it clear that much larger computational resources are needed for this proposal. Position averaged convergent beam electron diffraction (PACBED) is a relatively new technique for characterizing materials. Detecting the diffraction pattern from an electron probe/beam through a thin material, averaged over many positions, yields an image that is highly informative with respect to sample thickness (number of atomic layers), tilt angle relative to the beam, and phase (atomic lattice structure). This is highly useful, in particular for characterizing novel 2D materials e.g. graphene structures, including spatially variable thickness and tilt angles which are crucial for the material performance in applications. The conventional method for PACBED analysis is visual inspection and/or least squares fitting to match the data (a 2D image) to a precomputed library of simulated PACBED images. However, this is highly challenging due to variations in e.g. translation, rotation, and scaling of the pattern, and computationally heavy during "inference" due to the large number of comparisons that have to be made. In two previous papers by other authors, convolutional neural networks have been used for PACBED analysis, showing promising results. In our ongoing work, we significantly advance that previous work with more realistic noise models, better CNN architectures, and more appropriate loss functions that give demonstrably better results. Also, we demonstrate that classification of the phase (atomic lattice structure) can be performed. Further, we perform an optimization over a set of experimental parameters that has not been done before, to provide a guide to others regarding how to optimize the experimental design. The project involves both experimental specialists from the Eva Olsson group at Physics at Chalmers, and AI expertise from the Mathematical Sciences and Electrical Engineering departments. The code and the data will be published open access on Zenodo, ensuring optimal spread of the methods and new insights on materials design and properties.