Proof

Published methods, open-source software, and industrial delivery experience.

NeuralSoftX is new as a commercial entity, but the methods and software behind it are not. This page maps the open-source platforms in active use, the peer-reviewed work that grounds them, and the broader industrial-ML track record that supports commercial delivery.

Open-source platform developed by the founder

The simulation platform behind NeuralSoftX's synthetic-data work.

MULTEM

Multislice simulation for TEM, STEM, CBED, EFTEM, and EELS

Open-source C++ / CUDA simulation platform for electron microscopy, developed by Dr Lobato since 2014. Used as the simulation engine for synthetic training data, method validation, and physics-backed model development when industrial labelled data is scarce.

  • GitHub: 78 stars, 27 forks, GPL-3.0, actively maintained (last push July 2025)
  • Published: Ultramicroscopy 2015 (177 citations) and 2016 progress paper (81 citations)
  • Delivery: simulation-supported workflows or synthetic-data pipelines

NeuralSoftX capabilities

What NeuralSoftX delivers today — described by capability, not package.

The company's engineering line is in-house, led by the author of the underlying published methods, and built as the next generation of workflows that address industrial electron-microscopy bottlenecks. Capabilities are organised below using the EM-literature vocabulary, not generic machine-learning jargon.

Restoration & preprocessing

Physics-informed image restoration, registration, and diffraction preprocessing.

Acquisition-quality repair before downstream analysis ever runs: single-shot denoising for low-dose and fast-scan acquisitions, rigid and non-rigid registration for drifting time-series, tomography preparation for tilt-series, and sub-pixel centring / restoration for CBED and 4D-STEM data.

  • Modalities: SEM, STEM, TEM, HAADF-STEM, HRSTEM, CBED, 4D-STEM, tomography
  • Typical delivery: ONNX-packaged inference, on-prem preprocessing step, or pipeline integration
  • Runs on: CPU, NVIDIA GPU, or DirectML

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 workflows across the full set, using the EM-literature vocabulary each category demands rather than generic computer-vision jargon.

  • Atom-column detection and localisation — HAADF-STEM, HRSTEM, HRTEM — for catalysts, alloys, ferroelectrics, and semiconductor R&D.
  • Phase, grain, and domain segmentation — HRTEM / HRSTEM at atomic resolution; SEM microstructure at mesoscale — for multi-phase alloys, perovskites, 2D heterostructures, and additive-manufacturing QA.
  • Nanoparticle / particle segmentation — SEM, STEM, TEM — for catalyst, pharmaceutical, battery, and pigment sizing.
  • Fibre and filament segmentation — SEM — for electrospun mats, fibre-reinforced composites, filter media, and nanocellulose.
  • Crack, void, and fractography segmentation — SEM — for additive manufacturing, fatigue-life studies, and post-mortem failure analysis.
  • Technical-cleanliness particle analysis — SEM-EDS under ISO 16232 / VDA 19 workflows — for automotive, aerospace, and precision-manufacturing QC.
  • Defect and damage segmentation in 2D materials — STEM — for monolayer semiconductors and quantum-material engineering.
  • Semiconductor defect inspection and metrology — CD-SEM, e-beam inspection (EBI), cross-sectional TEM / STEM — for patterned-wafer triage, LER / LWR metrology, and FinFET / GAA failure analysis.
  • Biological and volume-EM segmentation — SBF-SEM, FIB-SEM, serial-section TEM — capability, with cryo-EM single-particle picking flagged as adjacent.

Simulation-driven model development

Physics-accurate synthetic training data, end-to-end.

Industrial EM datasets rarely come with enough labelled examples to train robust models directly. NeuralSoftX closes that gap with MULTEM-generated synthetic training data — electron–specimen interaction simulated at the multislice level, with detector-stage physics applied on top — so the network sees training inputs that match the client's instrument, sample, and acquisition regime.

  • Simulation engine: MULTEM (C++/CUDA multislice, open-source, actively maintained since 2014)
  • Covers: TEM, STEM, HAADF-STEM, CBED, 4D-STEM, EFTEM, EELS
  • Outputs: training datasets, pretrained models, or full simulation-driven method-development workflows

Why synthetic data

Industrial EM labs don't need to label more data — they need a simulator that matches their microscope.

The binding constraint in industrial electron microscopy is labelled data scarcity: real-world datasets are proprietary, expensive to annotate, and often live under NDA. NeuralSoftX's signature approach is physics-based synthetic data generation — every simulated training image is the output of a full forward model, from electron–specimen scattering through the detector itself.

Electron–specimen physics

Multislice simulation via MULTEM covers elastic and inelastic scattering, frozen-lattice thermal diffuse scattering (essential for quantitative HAADF-STEM), and full aberration-function modelling. Scattering potentials use the Lobato–Van Dyck 2014 parameterisation (Acta Crystallographica A) — roughly an order of magnitude more accurate than Doyle–Turner or Kirkland parameterisations across the periodic table.

Detector recording physics

The detector forward model is instrument-specific, not a generic Gaussian. It covers the point-spread function (PSF), modulation transfer function (MTF), detective quantum efficiency (DQE(0) and DQE(ν)), Poisson electron shot noise, Gaussian readout noise, gain and dark reference, and — for STEM — fast-scan and slow-scan distortion following Ophus, Ciston & Nelson (Ultramicroscopy 2016).

Domain randomisation

The simulator sweeps dose, defocus, spherical aberration, PSF width, gain, and contamination distributions so that networks trained on synthetic data generalise to real experimental micrographs. Sim-only training is used where labels are impossible (e.g. defect detection, atom segmentation); pretrain-then-finetune is used where a small amount of experimental data is available.

Selected publications

Peer-reviewed work grounding the solutions.

Dr Lobato has 30+ peer-reviewed publications and a total of 1,800+ citations (h-index 21) across Science, Nano Letters, PNAS, Physical Review Letters, Ultramicroscopy, Acta Crystallographica, and npj Computational Materials. Selected items below; full list on Google Scholar.

Ultramicroscopy · 2015 · 177 citations

MULTEM: A new multislice program for accurate and fast electron diffraction and imaging simulations

Introduces the MULTEM simulation platform used today as a foundation for synthetic training data and method development across TEM, STEM, CBED, and EELS.

npj Computational Materials · 2024 · 48 citations

Deep convolutional neural networks to restore single-shot electron microscopy images

Physics-informed deep-learning restoration for low-dose and fast-acquisition SEM, STEM, and TEM images — the published method that informs NeuralSoftX's current in-house workflow development.

Acta Crystallographica A · 2014 · 109 citations

An accurate parameterisation for scattering factors, electron densities, and electrostatic potentials

Foundational physics work that underlies quantitative electron microscopy simulation and interpretation.

Physical Review Letters · 2018 · 44 citations

Single-atom detection from low contrast-to-noise-ratio electron microscopy images

Quantitative detection at the single-atom limit on images below the standard interpretability threshold.

Ultramicroscopy · 2016 · 81 citations

Progress and new advances in simulating electron microscopy datasets using MULTEM

The follow-up paper documenting expanded simulation capabilities and performance improvements.

Science · 2020 · 417 citations

Micelle-directed chiral seeded growth on anisotropic gold nanocrystals

Co-author. High-impact collaborative work illustrating electron-microscopy contribution to nanomaterials research.

Recognition

Cosslett Award — Best Invited Paper, Microscopy & Microanalysis 2018.

Awarded at M&M 2018, the largest annual international microscopy conference, for invited work on quantitative electron microscopy.

Broader industrial delivery

Industrial ML track record outside electron microscopy.

NeuralSoftX also collaborates with Liquisens on industrial ML workflows outside core electron microscopy — water-quality prediction (chemical oxygen demand, legionella detection) and process-monitoring regression and embeddings for production-line data. Client names under NDA. This experience supports commercial execution credibility without changing NeuralSoftX's microscopy-first positioning.

Next step

Does any of this look close to your workflow problem?

Start a conversation See solutions detail