Skip to main content
Back to timeline
arXivSource publication:

Atoms, not pixels: 4D-TEMPO fits each atom directly and attaches an error bar

Synopsis

Electron ptychography previously reconstructed a pixel image and read atoms off it, so it could not report per-atom uncertainty. 4D-TEMPO puts atoms, microscope and detector into one Poisson likelihood fitted directly to recorded electron counts, returning each atom's 3D position, thermal width and a Fisher-information error bar. On a simulated WSe2 bilayer, starting from the published reconstruction, in-plane error falls from 4.8 to 3.2 pm and depth error drops 40–61% against its 3D peaks.

Source-provided article image: 4D-TEMPO: Differentiable 4D-STEM ptychography via fast Physics Optimizers
arXiv

Interpretation

The output of atomic imaging becomes a list of atoms with error bars rather than a pixel image; each atom is a Gaussian primitive with a 3D position, a thermal width and a model-conditional error bar. Earlier pipelines reconstructed a pixel image and read atoms off it, or fitted atoms under a loss on atoms; neither gave per-atom uncertainty. On a simulated WSe2 bilayer, starting from the published reconstruction with no structural priors, fitted atoms reach 3.2 pm in-plane error against 4.8 pm for the published 3D peaks, and cut depth error by 40–61% against those peaks.

The method writes atoms, microscope and detector into one differentiable physical renderer, using a multislice model to map Gaussian atoms to expected diffraction patterns and comparing them with measured counts. The renderer uses a Poisson likelihood fitted directly to recorded electron counts rather than least squares on intensities. On experimental WSe2 data, the atoms-only fit reaches 0.450 held-out Poisson deviance per pixel against 0.455 for fold_slice; with free probe modes the fit leads on all three benchmarks: the WSe2 bilayer, a 21 nm PrScO3 crystal and a PZT film.

The same differentiable renderer supplies its Jacobian and the Fisher information, which yields an error bar for every atom, identifies which parameters the data determine, and computes in advance the dose each parameter needs. Previously there was no per-atom uncertainty, and no way to state before a scan how many electrons a given parameter requires. When the model is well specified, 69% of position coordinates fall within one standard deviation on Poisson redraws of the fitted model (nominal 68%), and 95% within two (nominal 95%).

Perspective

The demonstrated scope is a thin WSe2 bilayer about 0.15 depths of field thick, plus pixel-object fits on thick specimens such as PrScO3, PZT and PdH; per-atom error bars are calibrated only when the model is well specified.

Each bar is a model-conditional bound, and per-atom misfit on the real bilayer exceeds it until diffuse scattering and coherence are better modeled; at published doses, counts constrain single-atom depth no better than layer geometry.