PySlice: Routine Vibrational Electron Energy Loss Spectroscopy Prediction with Universal Interatomic Potentials

Vibrational spectroscopy in the electron microscope can reveal phonon excitations with nanometer spatial resolution, yet routine prediction remains out of reach due to fragmented workflows requiring specialized expertise. Here we introduce PySlice, the first publicly available implementation of the Time Autocorrelation of Auxiliary Wavefunction (TACAW) method, providing an automated framework that produces momentum- and energy-resolved vibrational electron energy-loss spectra directly from atomic structures. By integrating universal machine learning interatomic potentials with TACAW, PySlice eliminates the bottleneck of per-system potential development. Users input atomic structures and obtain phonon dispersions, spectral diffraction patterns, and spectrum images through a unified workflow spanning molecular dynamics, GPU-accelerated electron scattering, and frequency-domain analysis. We outline the formulation behind the code, demonstrate its application to canonical systems in materials science, and discuss its use for advanced analysis and materials exploration. The modular Python architecture additionally supports conventional electron microscopy simulations, providing a general-purpose platform for imaging and diffraction calculations. PySlice makes vibrational spectroscopy prediction routine rather than specialized, enabling computational screening for experimental design, systematic exploration of phonon physics across materials families, and high-throughput generation of simulated data for training of future machine learning models.

💡 Research Summary

The manuscript introduces PySlice, an open‑source Python framework that makes routine prediction of vibrational electron energy‑loss spectroscopy (VEELS) feasible for a broad community. The authors identify three major bottlenecks that have prevented widespread use of VEELS simulations: (1) the need for system‑specific interatomic potentials, (2) the lack of an integrated, GPU‑accelerated multislice engine, and (3) the fragmented implementation of the Time Autocorrelation of Auxiliary Wavefunction (TACAW) formalism, which historically required stitching together separate MD, multislice, and analysis codes.

Core innovations

1. Universal Machine‑Learning Interatomic Potentials (uMLIPs) – PySlice interfaces ASE with a suite of pre‑trained neural‑network potentials (ORB, MACE, CHGNet, Allegro‑FM, UMA, etc.) that have been trained on massive DFT databases covering most of the periodic table. Users simply supply a crystal structure (CIF, XYZ, LAMMPS trajectory, or ASE Atoms) and obtain DFT‑quality forces and energies without any per‑system fitting. The MDCalculator automates equilibration detection using temperature, potential‑energy variance, and sliding‑window criteria, then switches from an NVT Langevin stage to an NVE production run (or low‑friction Langevin) to generate a thermally populated trajectory.

2. GPU‑accelerated multislice propagation – The electron‑scattering engine is built on PyTorch, allowing automatic selection of CUDA, Apple Metal (MPS), or CPU back‑ends. The multislice algorithm follows the standard slice‑by‑slice phase‑shift and Fresnel propagation scheme, using Kirkland’s parameterized atomic scattering factors for accurate projected potentials. Exit wavefunctions for each MD timestep are cached as NumPy arrays, enabling restart capability and post‑processing without recomputation.

3. TACAW analysis – Instead of the frequency‑filtered FRFPMS approach, TACAW performs a single multislice simulation on the full time‑dependent atomic trajectory. The exit wave ψ(k,t) is mean‑subtracted and Fourier‑transformed in time to obtain ψ̃(k,ω). The vibrational EELS intensity I(k,ω)=|ψ̃(k,ω)|² directly yields phonon peaks across the entire Brillouin zone, providing momentum‑resolved spectra, dispersion curves, and spectrum‑image maps.

Software architecture – The workflow is modular: (1) structure input, (2) optional MD, (3) multislice, (4) TACAW analysis. Data containers (Trajectory, WFData, TACAWData, HAADFData) manage information flow and enable flexible recombination of stages. The code also supports conventional frozen‑phonon, HAADF, and 4D‑STEM simulations, making PySlice a general‑purpose TEM/STEM platform.

Demonstrations – The authors showcase four transition‑metal dichalcogenides (MoS₂, WS₂, WSe₂, MoSe₂), generating phonon dispersions and diffraction patterns with a single script. They further produce vibrational spectrum images for a bulk Si/Ge heterostructure interface and for a silicon substitution defect in graphene, illustrating spatial‑resolution capabilities.

Strengths and limitations – PySlice dramatically lowers the expertise barrier, offers GPU speed‑ups, and leverages transferable ML potentials. However, accurate VEELS still depends on sufficient MD sampling (trajectory length, timestep) and on the fidelity of the chosen uMLIP for a given chemistry; exotic systems (strongly correlated oxides, magnetic materials) may require additional validation or custom potentials.

Future directions – The authors propose extending the framework to longer, anharmonic MD runs, expanding the uMLIP training set to cover more chemistries, and using the large simulated datasets to train inverse‑problem ML models for experimental VEELS interpretation.

In summary, PySlice unifies state‑of‑the‑art interatomic potentials, high‑performance multislice scattering, and the TACAW formalism into a single, user‑friendly package. By automating equilibration, caching, and Fourier‑domain analysis, it transforms vibrational EELS simulation from a specialist task into a routine computational tool, opening new possibilities for high‑throughput phonon screening, experimental design, and data‑driven materials discovery.