Project
Atomic STEM Modelling & Image Analysis
Research software comparing an interpretable Gaussian STEM surrogate with abTEM/multislice reference simulations for atomic-resolution image modelling and validation.
Overview
This ongoing research project asks whether an inexpensive, interpretable Gaussian forward model can preserve the image and candidate-ranking information needed for atomic-structure inference while reducing simulation cost. The Gaussian model is treated as the M0 surrogate and abTEM/multislice as the high-fidelity HAADF-STEM reference.
The work builds on prior research reconstructing three-dimensional nanocluster structures from aberration-corrected STEM data.
Problem
Atomic-resolution STEM analysis depends on comparing experimental or candidate images with physically meaningful simulations. High-fidelity multislice calculations can be expensive, while faster approximations are useful only if they preserve the structural information relevant to downstream inference.
Approach
The experimental workflow starts from the same ASE atomic structure and generates both a Gaussian surrogate image and a matched abTEM HAADF-STEM reference. Physical sampling and downstream preprocessing are kept explicit so differences can be attributed to the forward models rather than mismatched grids or hidden transformations.
Validation is designed across complementary levels: pixel/intensity fidelity, structural image agreement, candidate-ranking behaviour, and runtime. Separate convergence and detector-refinement experiments characterize the reference calculation before broader surrogate comparisons.
Outcome
The repository now contains the Gaussian simulator, abTEM reference adapter, shared grid and processing logic, Au benchmark structures, image metrics, integration tests, and a sequence of abTEM grid, convergence, and detector-refinement experiments.
The current milestone is a controlled validation framework rather than a claim that the surrogate is already an adequate replacement for multislice simulation. The next scale-up tests whether the surrogate preserves structure-ranking behaviour across larger candidate sets and orientations.
Tradeoffs
A faster model is useful only when its domain of validity and failure modes are measured against the scientific task that will use it. Pixel similarity alone is therefore insufficient: scale, local structure, ranking behaviour, and reference-convergence choices must be evaluated explicitly.
Professional profile
Continue from the evidence
See how this work fits into my broader computational science and software engineering background, or get in touch about relevant work.