Changelog

Changelog#

The changes of each release, newest first. The changes of the next release go under a heading ## Upcoming: <version>, which is renamed to the version when it is released.

1.0.11#

A bugfix release on top of 1.0.10 with fixes ported from dev; no features and no new defaults. Plain potentials, plane-wave and probe multislice with default settings, ensembles of unweighted parameters, zarr round trips and the configuration defaults are bit-identical to 1.0.10 at the default float32 precision (checked on 66 scenarios).

Fixed: results that change#

Magnitudes are from small test systems and vary with the data.

4D-STEM and measurements

Centre of mass and iCOM are divided by the captured intensity (#417); an all-zero pattern gives NaN instead of 0.
Default diffraction patterns are unshifted: their coordinates, azimuthal_average, bandlimit and centre of mass now agree with that (#463; azimuthal averages differ by up to 80 %).
Radial bins are exactly step_size wide (#469; default ADF/HAADF move by 0.2 to 0.8 %, radial_binning by 13 to 33 %).
A FlexibleAnnularDetector without outer follows the waves of each call, not the first ones it saw (#434).
PolarMeasurements.integrate azimuthal limits (#459), rad/deg conversions (#458, off by 1e6), depth_profile and Images.coordinates (#461), interpolate_line and 2 / images (#470), periodic spline interpolation (#506).
AnnularDetector(offset=...) is applied (#324); get_detector_region returns percent as a float (#283).
block_direct=True blocks the whole bright-field disk instead of 1 mrad (#484); block_direct() without cutoff blocks only the zero-angle pixel (#556).
Reciprocal-space waves measure like real-space ones (#565, #566); lazy gaussian_source_size is periodic like the eager one (#485); integrate_gradient takes the minimum per image (#556).
Seeded Poisson and scan noise give another realisation for the same seed (#464, #475). Float32 stays float32 in transfer functions, integrate_gradient and center_of_mass (#604).
PixelatedDetector(reciprocal_space=False) has real-space axes, WavesDetector(gpts=...) declares the right shape, and block_direct no longer grows its metadata (#499).

PRISM

Aberrated SMatrix.reduce/scan match Probe (#318; 17 to 75 %). An eager frozen-phonon reduce is averaged like the lazy one (#576), and eager SMatrix.scan/reduce reduce the CTF ensemble axes (#466). PRISM no longer leaves its outer angle on the caller’s detector (#491).

Potentials and atoms

orthogonalize_cell followed by Potential no longer drops atoms near cell faces (#345, #390, #555; wurtzite GaN lost 2 of 8 atoms, projected potential +33 %). Non-periodic cells keep displaced atoms (#460, #555).
CrystalPotential honours first_slice/last_slice (#269); GPAWPotential(repetitions=...) was wrong by about 100 % (#494); GPAWParametrization applies charge (#329).
GPAW and Bloch slices are integrated centred on their slices, which shifts each slice by dz/2 (5 % for GPAW, 17 to 30 % for Bloch; whole-cell sums unchanged) (#574); ChargeDensityPotential slices are sampled at midpoints (#472).
IAM magnetic fields change by ~1e-6 and their slices are labelled ÅT, Å²T (#543); projection="finite" with sigmas for only some elements dropped the others (#480); the GPAW magnetic real_space projection ignored first_slice (#544); wrapped_crop_2d outside the cell (#389); eager multislice of potentials with two or more ensemble axes scrambled the results (#333).

Bloch waves

Non-centrosymmetric structures were computed for the negated structure (#379; wurtzite exit waves change by up to 1.27).
The symmetrising metric is applied consistently (#448): the eigendecomposition and matrix-exponential paths now agree to rounding. For Si at 100 keV, default intensities change by R = 1.5e-3 to 2.2e-3 on the zone axis and 3e-4 to 1.3e-3 at a 7° tilt (R = Σ|ΔI| / ΣI).
Eager rotation ensembles (#450), A/B/C centring (#487) and CuPy scattering matrices (#449).

HRTEM and CTF

set_aberrations('scherzer') C10 sign and ordering (#471); apply_ctf no longer writes the waves’ energy into the CTF (#489).

Core-loss

Continuum normalisation (#364), the (2l+1) double count (#368; L2,3 ÷3, M4,5 ÷5), site threshold and a spurious 1.02 factor (#468; K +3 to 4 %, L2,3 −22 to −27 %) and the vxc vacuum tail (−4 % at 3 eV). Single-channel frozen-phonon core loss was num_configs times too large or zero (#387).

Backscattering

Back-propagation is corrected (#562) and ensembles work (#499).

Weighted distributions

Distribution weights are probabilities applied when the ensemble is reduced (#466). In 1.0.10 a STEM focal spread was effectively a top-hat, an HRTEM Gaussian spread had a standard deviation of σ/√2, and tilt and aperture weights were ignored. The peak of the reduced intensity of a 200 kV, 25 mrad probe with gaussian(σ, num_samples=11) defocus is 1.76 times the 1.0.10 value at σ = 25 Å and 2.21 times at 50 Å. Frozen-phonon results are unchanged.
New: OrdinalAxis.weights/from_distribution and reduce_ensemble(axis=). gaussian weights sum to one; normalize='intensity' and 'amplitude' are deprecated; .sum()/.mean() over a weighted axis warn; negative or wrong-length weights raise ValueError. Bloch rotation spreads and noise axes carry no weights, as before. Zarr files with a weighted axis cannot be read by 1.0.10.

Distributed computations

The client’s configuration reaches the workers of a distributed client, also for to_tiff and to_zarr(compute=False) (#346). Workers used to see only the YAML defaults, so for example precision='float64' runs were computed in float32. Distributed results obtained with a non-default configuration are worth repeating.

Other

FFTW fftn over three axes transformed the wrong axes and strided input with overwrite_x=True raised (#558); a lazy fft2_convolve had the wrong shape for long kernels (#577).
Waves.phase_shift relabelled reciprocal-space waves and widened complex64 (#550), as did tile(renormalize=True) (#414); 1D interpolate lost float64/complex dtypes (#427); bandlimit and block_direct(radius=0) kept the direct beam on some grids (#506).
StructureFactorArray.get_projected_potential honours sampling= and an integer gpts and IndexedDiffractionPatterns.positions_dict is complete (#456).

Fixed: errors and crashes#

No longer raising: Images.scan_noise, apply_ctf at a second energy, number - Images, lazy gaussian_filter with a constant boundary, lazy normalize, single-point GridScan, noisy near-orthogonal Potential cells, tilted eager Bloch exit waves with scalar thickness, IndexedDiffractionPatterns.sort(), core loss on 16 or 17 pixel grids, AnnularDetector over a LineScan with an extra ensemble axis, GPAW 26 calculators, frozen-phonon GPAWPotential, ensemble backscattering, A/B/C centring, graphene repeated along c, and numba typing on some numba/numpy pairings.
TransitionPotentialArray.filter_sites with Atoms of several elements and a threshold no longer raises IndexError; it returns the kept atoms of the element of the transition potential, as Atoms for Atoms input and as positions for positions (#375).
show(overlay=...) of an ensemble that has no base axes, such as the integrated core-loss measurements of the core-loss tutorial, gave more line labels than lines, and an ensemble of complex lines fewer. Matplotlib 3.11 raises an error for this (“label must be scalar or have the same length as the input data”); older versions took the whole list as the label. The lines are now labelled by the axes that are overlaid.
excitations_weights fails no longer on NumPy 2 (np.math). On CuPy 14.2 a lazy FFT in a dask worker thread no longer fails with CUFFT_EXEC_FAILED (see cupy.fft-cache-size).

Changed: behaviour#

abTEM reads ~/.config/abtem and $ABTEM_CONFIG as well as dask’s configuration files (#416); 1.0.10 ignored them. A leftover cupy.fft-cache-size: 0 MB there now disables the cuFFT plan cache.
cupy.fft-cache-size, ignored in 1.0.10, is read and defaults to auto (25 % of the device memory per thread); 0 MB disables the cache, -1 removes the bound, and a plan larger than the bound runs uncached with a warning.
LineScan.angle reports degrees (#470). SegmentedDetector(to_cpu=...) defaults to True as documented (#371). Exploded show() grids use 0.25 between panels (#354).
New errors instead of wrong results: backscattering with exit planes that do not start at the entrance plane or are not after every slice (#562, #499); a built TransitionPotentialArray at another energy or grid (#386); PixelatedDetector(resample='uniform') without reciprocal space (#499); transition_potential_scan with expansion_scope="full" (#506).
Zarr arrays over 512 MiB use a multi-chunk layout and a failed write removes the partial file (#363, #369). GPAWParametrization for Pm, Sm, Eu and Pu falls back to an unpolarised atom with a warning (#467). Importing abtem no longer fails if cupyx.scipy.ndimage cannot be loaded (#381).

1.0.10#

Features:

Expanded real-space multislice with propagator- and fully-corrected algorithms, and backscattered waves (PR #236)
- Related internal function name change: standard multislice is now properly called FourierMultislice
Updated BullseyeAperture to use smoothed aperture edges/corners (PR #266)
Logarithmic scale display for DiffractionPatterns and images (PR #303)
Finite-projection integrals 7x faster on CPU (PR #309)

Documentation:

Expanded tutorial on real-space multislice
Depth-profile visualization of potentials in the walkthrough
Single-file Zarr zip storage, logarithmic display scaling, soft Bullseye apertures, anisotropic Debye-Waller factors and B-factor conversion helpers
All published notebooks verified to run with this release

Dependencies:

NumPy 2.0 or newer is now required (PR #245)
GitHub actions based on uv and now cover more versions (including Python 3.14) (PR #308)
New branching structure: dev for development, main for releases (only via PRs from dev)
Support for Zarr 3 ZipStore (requiring zarr>=3.1)
- Added Zstandard compression of the arrays in the ZipStore at default level 4 (hat tip: quantEM) (PR #252)
Deprecated [gpu] optional dependency (as just specifying cupy will not install the correct CUDA version)
Moved testing, docs, and dev from optional dependencies to groups
Narrow Dask version exclusion to !=2025.12.*,!=2026.1.0,!=2026.1.1 (PR #285)
Declared sympy as an optional dependency (core-loss extra), required for core-loss EELS form factors (PR #320)

Bugfixes:

Frozen-phonon ensemble handling (PR #267 & PR #292)
- May also have resulted in incorrect behavior with ensemble_mean = False for e.g. defocus distributions
CrystalPotential with frozen phonons bugs (especially bad on – luckily rare – eager compute) (PR #306)
Minor bugs, unsafe patterns, and dead code (PR #265)
Anisotropic Debye-Waller factors for Bloch waves (PR #271)
- Added helper functions to convert between crystallographic B-factors and thermal sigmas
Added missing .calculate_exit_waves for BlochwaveEnsemble (PR #294)
Silent atom drop when z-position lands in SliceIndexedAtoms blind spot (PR #273)
Early-exit bug in orthogonalize_cell (PR #291)
Fixed broken tutorial workflow for core-loss filtered imaging (PR #284)
- Minor performance improvements for transition_potential_scan (PR #286)
Bullseye aperture: ring_width and spoke_width are now validated, rejecting values that previously produced a silently wrong (solid disk) aperture; docstring corrected to describe the actual fractional units introduced by the soft-edge redesign (PR #319)
Unified the task-level progress bar config key on diagnostics.task_progress (Bloch-wave code paths previously read a different, non-functional key) (PR #321)

1.0.9#