Pipeline
Key results
Quick actions
Recent jobs
Measurements
Phase metrics, particles and transport on the current label overlay. Results update when you run an analysis.
Volume fractions
Phase metrics
Particle size
Particle shape
Each dot is one particle fully inside the volume. Sphericity 1 is a sphere.
Transport
Tortuosity factor per phase and direction from a steady diffusion solve; 1 is a straight path.
Triple-phase boundary
Table explorer
Click a column header to sort it and see its distribution.
Confidence
How far the numbers can be trusted: slice quality, sampling error, the volume a precise answer needs, and whether coarser sampling would have changed it.
Data health, slice by slice
Fractions with both error sources
Upper whisker: 95 % sampling range. Lower whisker: every boundary moved by one voxel.
Volume needed for ±5 %
Would coarser sampling change the answer?
Precision while milling
Relative error of each phase fraction as slices arrived in live mode. Dashed: ±5 % target.
Chemistry
Correlative EDS with the SE volume, and ways to map chemistry with a fraction of the EDS time: context-driven sparse EDS (deep kernel learning, after Kalinin and Ziatdinov) and coupled-NMF pan-sharpening with BSE (after Borodinov et al., 2019).
Reconstruction quality by method
What the benchmark shows
- Sparse EDS guided only by the SE image misses chemistry that SE cannot see.
- Adding the BSE image, whose signal follows composition, recovers most of it at a tenth of the EDS time.
- Pan-sharpening with BSE also finds the hidden region; with few element channels it is limited by EDS counting statistics.
EDS in this benchmark is simulated on the real segmentation. Run it with python examples/chemical_fusion.py.
Correlate real EDS maps
Import element-map exports (e.g. AZtec TIFFs), register them to the SE stack, find chemical phases and the composition of every segmented phase or particle.
Acquire
Follow an acquisition while the FIB mills, and run the self-driving agent on a digital twin of this sample before any beam time.
Self-driving run on the digital twin
The agent mills, refocuses, corrects drift, adapts dwell and slice thickness, takes sparse EDS and stops at the precision you ask for. Its Thermo Fisher AutoScript calls are recorded in dry-run only.
Live acquisition
Start it next to the microscope with nuscope-tomo watch <acquisition folder> -p <project> --voxel 20 5 5 --model rf.pkl. Studio follows the growing volume in the View workspace.
Precision against instrument time
Fraction error (95 %) of the phases that must reach the target. Dashed: target.
Slice thickness and dwell
What the run used at each slice.
Where the time goes
Decision log
Earlier comparison of policies
Recipe
Every action in Studio is recorded as a step. Run the same pipeline on another dataset with nuscope-tomo run studio_recipe.yaml.