quantem.gpu

GPU-accelerated 4D-STEM analysis for scientists using Python. Load measurements, select diffraction patterns, make detector images, and reconstruct phase with single-sideband ptychography (SSB).

Active development. These examples use the current source checkout. APIs may change; follow the installation guide and record git rev-parse HEAD with results.

Start with Python

Start here	What you will do
Install	Set up Apple Silicon MPS or NVIDIA CUDA
From acquisition to images	One Gold acquisition: I/O → DPC → SSB
Advanced Gold tutorial	Inspect aberrations and the model probe; compare 1× and 4×
Save and share your data	Preserve measurements, calibration, and processing history in QEM
Python API reference	Check parameters, units, return values, and backend limitations

Your first images

from quantem.gpu import io, detector
from quantem.core.visualization import show_2d

data = io.load("gold_master.h5")
show_2d(
    [detector.bf(data), detector.adf(data)], title=["BF", "ADF"],
    norm="power_sqrt", cmap="inferno",
)

Each image covers a 512 × 512 scan and uses independent square-root contrast. This source has no recorded physical sampling, so no physical scale bar is shown.

The detector fits the bright-field disk automatically. Replace gold_master.h5 with your file; the acquisition is not bundled. Keep data open while working, then call data.close(). The README contains the same short workflow and additional static image examples.

SSB in quantem.gpu

quantem fits SSB aberrations with a physical model of the measurement, not an image-quality score: for each trial set of aberrations, sample tilt and depth spread it predicts the bright-field data the detector should have recorded and keeps the parameters whose prediction agrees best (least squares). This model-based, tilt-corrected fit is the default everywhere, and it is where SSB in quantem is heading.

Call	Method	Use it for
fit_aberrations()	model-based, tilt-corrected, refined through fifth order	the default: every fit
fit_aberrations(search="primary", tilt=False)	model-based, tilt and depth spread held at 0	thin, untilted samples, or to see what tilt correction changes

Both refine on exact derivatives of the model agreement. The earlier phase-variance search (objective="phase_variance"), which scored how well the bright-field pixels’ phase images agree instead of predicting the data, is retired. Backend support and evidence: SSB fit methods.

What stays consistent

Python is the interface; MPS and CUDA are execution backends. Supported options differ by operation, so limitations are listed beside their API. Ordinary GPU operations fail explicitly if the requested path is unavailable.

Data use (scan_rows, scan_cols, detector_rows, detector_cols) order, written \(I[R_r,R_c,k_r,k_c]\) in the scientific guides. data[row, col] returns a selected GPU Torch tensor. .sampling, .units, and .origin describe the axes; missing calibration stays unknown. Detector reductions return small NumPy images while the acquisition remains ANS encoded.

For developers

The Developer guide groups equations, backend kernels, native integration, file formats, and verification. Backend coverage lists what each runtime implements. The verified benchmark results retain revision and hardware provenance. Historical measurements are not promises for every machine.

Citing quantem.gpu

If quantEM’s interactive widgets, GPU-accelerated I/O, or data processing and reconstruction tools contributed to your research, please consider citing:

Sangjoon Lee et al., “Interactive Framework for Real-Time 4DSTEM Analysis and Reconstruction,” Microscopy and Microanalysis 32 (Supplement 1), ozag053.941 (2026). https://doi.org/10.1093/mam/ozag053.941

Questions and bug reports belong in the issue tracker.