Atomic electron tomography (AET) enables the determination of 3D atomic structures by acquiring a sequence of 2D TEM projection measurements of a particle and then computationally solving for its underlying 3D representation. AET is a challenging and labor-intensive experiment! In this talk, we offer two computational methods to alleviate these challenges and make the reconstruction procedure more robust.
First, we describe a method that solves directly for the locations and properties of individual atoms from projection measurements. This is in contrast to classical tomography algorithms that first solve for a volume and then extract the atomic structure. We parameterize a particle as a collection of atoms each represented by a Gaussian. We show that this parameterization imparts a strong prior on the reconstruction that avoids physically implausible artifacts often present in volumetric reconstructions due to noise and missing wedge effects. These reconstruction improvements further translate to higher fidelity atomic structure identification.
Second, we tackle the problem of carbon contamination: over the time it takes to capture the tomographic projection series, amorphous carbon often accumulates on the sample surface, making it difficult to reconstruct the underlying static sample and causing laboriously collected datasets to be discarded. We use implicit neural representations, which compactly represent large 3D+time data cubes and impose flexible space-time priors, to directly model and solve for the 3D temporal dynamics of the sample This allows us to computationally remove the contamination and recover an uncorrupted reconstruction of the static sample of interest.
