MULTEM
MULTEM supports HRTEM, STEM, ISTEM, diffraction, PED, CBED, EFTEM and EELS workflows. It is used to study electron–specimen interaction and to generate controlled data for method development.
Open the repositoryTrack record
Established methods behind a new company.
NeuralSoftX is built on peer-reviewed computational microscopy, industrial software engineering, and over a decade of quantitative electron-microscopy research.
- 1,800+ citations and h-index 21
- 30+ peer-reviewed publications
- Cosslett Award, M&M 2018
- Open-source software maintained since 2014
Open-source software
From physical simulation to neural-network inference.
Three public repositories demonstrate complementary experience in electron-microscopy physics, deep-learning restoration, and learned surrogate modelling.
tk_r_em
Six pretrained convolutional networks restore and enhance single-shot high- and low-resolution SEM, STEM and TEM images. The package supports CPU, NVIDIA GPU and Windows DirectML inference.
Open the repositoryrt_ppiscs
A densely connected neural network predicts ADF-STEM probe-position-integrated scattering cross-sections across common fcc crystals, principal zone axes, microscope parameters and thermal displacement values.
It replaces lengthy GPU-based simulations with real-time prediction on a standard desktop computer.
Open the repositoryEvidence and maturity
Established expertise is separated from active product development.
Publications and open-source software demonstrate the scientific foundation. New NeuralSoftX-owned tomography, ptychography and neural exit-wave systems are being implemented and validated independently.
Simulation, restoration and inverse methods
Peer-reviewed work and maintained software establish experience; they are not automatically represented as NeuralSoftX-owned IP.
New reconstruction technology
Tomography and ptychography workflows are being packaged, while the neural exit-wave system is still completing its validation record.
Suitability must be demonstrated
A method becomes client-validated only after testing representative held-out measurements across the contracted operating domain.
Representative method
Simulation-trained restoration, validated on experimental data.
The 2024 npj Computational Materials work demonstrates the core NeuralSoftX principle: use physical simulation to create training evidence, then test the method against real acquisitions.
Deep convolutional neural networks to restore single-shot electron microscopy images
The published workflow addresses noisy or distorted single acquisitions in SEM, STEM and TEM. Dr Lobato co-designed the study, created the mathematical models for undistorted and distorted images, and implemented, trained and evaluated the neural networks. The method was then tested on experimental acquisitions rather than stopping at synthetic benchmarks.
Read the publicationComputational reconstruction
Inverse methods for incomplete and distorted microscopy data.
NeuralSoftX focuses on the computational part of tomography: how data is corrected, represented, constrained and optimised to recover defensible three-dimensional information.
Atomic tomography
A first-author method combined a simulated HAADF-STEM forward model, inter-atomic distance constraints, Tikhonov regularisation and simulated annealing for limited-angle atomic reconstruction.
Read the proceedings paper3D models from a single projection
A co-authored study reconstructed three-dimensional atomic models from a single Z-contrast (HAADF-STEM) projection using statistical atom counting, verified against full electron tomography.
Read the publicationMolecular dynamics
First-author molecular-dynamics work modelled thermal evolution, rapid cooling and structural transitions in silver nanoparticles, including quantitative structural and electronic analysis.
Read the publicationSelected publications
Computational research connected to the services.
Exact citation totals change continuously, so this page presents stable publication records and links to Google Scholar for current metrics.
- Deep convolutional neural networks to restore single-shot electron microscopy images
- Obtaining 3D atomic reconstructions using a Bayesian genetic algorithm
- Real-time simulations of ADF-STEM scattering cross-sections using a densely connected neural network
- Artefact reduction based on sinogram interpolation for 3D electron-tomography reconstruction
- Three-dimensional atomic models from a single projection, verified by electron tomography
- Precision limits for locating nanocluster atoms in 3D from an ADF-STEM tilt series
- MULTEM: Accurate and fast electron diffraction and imaging simulations using GPUs with CUDA
- Atomic and electronic structure transformations of silver nanoparticles under rapid cooling
Commercial execution
Delivery experience beyond publication.
Research credibility matters only when it can be translated into a reliable scope, implementation and handover.
Semplor
Current commercial SEM work provides direct exposure to instrument, acquisition and user-workflow constraints.
Liquisens collaboration
Delivered prediction, regression and representation-learning workflows for industrial process data. Client identities remain confidential.
Cosslett Award
Best Invited Paper at Microscopy & Microanalysis 2018 for work on accurate and fast MULTEM simulation.
Apply the track record