Solutions

Six commercial solution areas for industrial electron microscopy.

Every engagement starts from a concrete workflow problem — not from a generic software template. NeuralSoftX works where existing tools become unreliable: drift, low dose, difficult diffraction data, scarce labels, or instrument-specific variation. Each solution below names the underlying machinery (open-source packages and simulation platform) that makes the work possible.

1 — Registration & alignment

Rigid and non-rigid registration for datasets that change during acquisition.

Time-series image stacks drift, distort, or evolve in appearance during acquisition — particularly in fast-process studies, unstable acquisitions, and semiconductor SEM datasets. Classical rigid registration stops being enough. NeuralSoftX combines classical registration with ML-based correction to handle stacks where morphology itself is changing.

Typical data

Time-series SEM, STEM, or TEM images with rigid and non-rigid motion, scan artefacts, or evolving morphology.

Machinery

Classical registration plus custom ML models; simulation-backed evaluation where needed.

2 — Tomography preprocessing

Preprocessing support for tilt-series data before 3D reconstruction quality is lost.

Tilt-series reconstructions degrade quickly when raw projections are noisy, misaligned, distorted, or affected by acquisition instabilities. NeuralSoftX builds pre-reconstruction pipelines that restore, align, and compensate tilt-series data before the volume is built — including support for low-dose and high-resolution acquisitions.

Typical data

STEM or TEM tomography projections, low-dose and high-resolution datasets, beam-sensitive specimens.

Machinery

Physics-informed per-projection restoration, custom alignment and compensation, and MULTEM for simulation-backed validation.

3 — Low-dose and fast-scan restoration

Image restoration when repeat acquisition is too expensive or damaging.

Single-shot images can be too noisy or too distorted for reliable interpretation, but longer acquisitions cost time, throughput, or specimen integrity. NeuralSoftX delivers restoration as lightweight ONNX-based inference tools that can run on CPU, NVIDIA GPU, or DirectML — adaptable to instrument-specific conditions. The underlying method is published in npj Computational Materials (2024).

Typical data

SEM, STEM, and TEM images acquired under low-dose, low-signal, or fast-scan conditions.

Machinery

Physics-informed single-shot restoration delivered as ONNX-packaged inference (CPU, NVIDIA GPU, or DirectML).

4 — CBED and 4D-STEM preprocessing

Sub-pixel centring and restoration for quantitative diffraction workflows.

CBED and 4D-STEM pipelines become unstable when diffraction patterns are miscentred, noisy, or difficult to process with standard routines. NeuralSoftX provides preprocessing that centres, restores, and conditions diffraction data before it enters ptychography or quantitative diffraction analysis.

Typical data

CBED and 4D-STEM datasets for ptychography, diffraction analysis, and quantitative microscopy.

Machinery

Sub-pixel centring, restoration, and joint preprocessing recipes for difficult diffraction data.

5 — Segmentation across the EM stack

Segmentation is not one task — it is a family of tasks that depend on the modality and length scale.

NeuralSoftX builds segmentation workflows across the full EM stack, matching each category to the EM-literature vocabulary fab and R&D teams already use. The list below is organised by what the network is actually locating — atoms, phases, particles, fibres, cracks, contaminants, device features — not by generic computer-vision jargon.

  • Atom-column detection and localisation — HAADF-STEM, HRSTEM, HRTEM — for catalyst nanoparticles, high-entropy alloys, ferroelectric thin films, and interfacial reconstructions in semiconductor and battery R&D.
  • Phase, grain, and domain segmentation — HRTEM / HRSTEM at atomic resolution for crystallographic phases and polymorphs; SEM / BSE-SEM at mesoscale for microstructural constituents (ferrite / pearlite / martensite, alloy phases, additive-manufacturing microstructures).
  • Nanoparticle / particle segmentation — SEM, STEM, TEM — instance-level segmentation for catalyst sizing, pharmaceutical nanoparticle QC, powder-metallurgy feedstock, pigment analysis, and battery-electrode particle statistics.
  • Fibre and filament segmentation — SEM predominantly — diameter, orientation, and network statistics for electrospun mats, fibre-reinforced composites, filter media, and nanocellulose products.
  • Crack, void, and fractography segmentation — SEM — porosity and crack QA for L-PBF / DED additive manufacturing, fatigue-life failure analysis, weld-root inspection, and post-mortem fractography.
  • Technical-cleanliness particle analysis — SEM-EDS under ISO 16232 / VDA 19 workflows — automotive fluid-circuit, brake, and transmission release testing, aerospace cleanliness, precision hydraulics.
  • Defect and damage segmentation in 2D materials — HAADF-STEM, ABF-STEM, aberration-corrected HRTEM — point and line defects in MoS2, WSe2, graphene, and h-BN for monolayer-semiconductor device physics and quantum-emitter engineering.
  • Semiconductor defect inspection and metrology — CD-SEM contour / LER-LWR extraction, e-beam inspection (EBI) nuisance and killer-defect triage, EUV stochastic-defect detection, and cross-sectional TEM / STEM FinFET / GAA interface delineation.
  • Biological and volume-EM segmentation — SBF-SEM, FIB-SEM, serial-section TEM — organelle and membrane segmentation for phenotyping and connectomics-adjacent workflows (listed as capability; cryo-EM single-particle picking flagged as adjacent).

Typical data

SEM, BSE-SEM, HRSTEM, HAADF-STEM, HRTEM, 4D-STEM, SBF-SEM, FIB-SEM — across atomic, mesoscale, and volumetric length scales.

Delivery

Workflow per task category: dense semantic segmentation, instance segmentation (Mask R-CNN class), keypoint heatmap regression (atom-column detection), or flood-filling segmentation (volume EM). Browser-based review interfaces integrated where operator-in-the-loop review matters.

6 — Simulation-driven model development

Synthetic training data, physics-accurate end-to-end.

Industrial electron-microscopy datasets rarely carry enough labelled examples to train robust models directly. NeuralSoftX closes that gap by simulating training data at full physical fidelity — and this works because both the electron–specimen interaction and the detector recording process are well-understood physical processes. The generated data is not a textbook approximation; it matches a specific instrument, detector, and acquisition regime.

Every simulated image is the output of a full forward model: multislice electron–specimen scattering (elastic and inelastic, frozen-lattice thermal diffuse scattering for HAADF) via MULTEM; aberration-function modelling and probe simulation for STEM; detector physics covering point-spread function (PSF), modulation transfer function (MTF), detective quantum efficiency (DQE), Poisson electron shot noise, Gaussian readout noise, gain and dark reference; and scan-distortion modelling for fast-scan / slow-scan STEM acquisitions. Scattering potentials use the Lobato–Van Dyck 2014 parameterisation (Acta Crystallographica A), roughly an order of magnitude more accurate across the periodic table than Doyle–Turner or Kirkland.

Typical data

TEM, STEM, HAADF-STEM, CBED, 4D-STEM, EFTEM, and EELS acquisition regimes where simulation can approximate the relevant scattering and detection physics.

Delivery

Synthetic training corpora, pretrained models, sim-only or sim-pretrain-then-real-finetune pipelines, or full simulation-driven method development. Engine: MULTEM (C++/CUDA multislice, GPL-3.0, actively maintained since 2014, published in Ultramicroscopy 2015 and 2016).

Delivery modes

Engagements happen at the algorithm, tool, or workflow level.

Algorithm package

A callable method that internal teams integrate themselves. Lowest-friction starting point when the client already has ML infrastructure.

On-prem inference tool

A practical operator-facing workflow without cloud dependence — typically ONNX-packaged for CPU or GPU inference.

Workflow integration

Method embedded into a larger pipeline for instrument vendors or internal software teams. Tight fit to existing data formats and acquisition controls.

Custom model development

End-to-end method development for problems specific to one product line, detector setup, or sample family — often combining simulation-backed training with targeted experimental validation.

Next step

Which of these is closest to your current workflow problem?