Efficient Large-Scale STEM-EELS Simulations With Torched-TACAW

arXiv:2607.02236 · cond-mat.mtrl-sci, physics.comp-ph · Submitted 2026-07-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Efficient Large-Scale STEM-EELS Simulations With Torched-TACAW".

Mira: Here's the gist of the paper. Efficient large-scale STEM-EELS simulations with torched-TACAW presents a practical methodology for performing near ab initio quality simulations of ultra-low-loss electron energy loss spectra (EELS) arising from vibrational and magnon excitations in large material systems.

Kai: This approach is significant because it enables the simulation of realistic, large structure models containing defects or interfaces with tractable memory use and data flow.

Mira: The overall approach combines molecular dynamics based on foundational machine-learned interatomic potentials, partitioning of elongated supercells, and on-the-fly processing of multislice outputs to enable near ab initio quality simulations with tractable memory use and data flow.

Kai: The workflow involves several key steps: 1.

Title and authors: Kai: So this paper is called "Efficient Large-Scale STEM-EELS Simulations With Torched-TACAW". I mean, that's a mouthful for a title. It sounds like they're talking about making something really big and complicated work fast.

Mira: Yeah, it’s about taking electron energy loss spectra from materials, especially those involving vibrations or magnons—that’s EELS—and doing them at the atomic level using this method called TACAW. The title suggests the main focus is on efficiency for large systems.

Lev: Efficiency in what context? Because when you start pushing up to really large models, like with defects or interfaces, the standard TACAW calculations become very slow and demanding on hardware.

Kai: Exactly. And this paper is proposing torched-TACAW as a practical way to handle those large systems without needing astronomical amounts of memory or data throughput. It’s about making something that used to be too big for us feasible on real hardware.

Mira: It’s essentially a practical methodology for performing simulations that aim for near ab initio quality, which means they should be very accurate, but the way they do it is optimized so it doesn't crash our systems when we look at materials with defects or grain boundaries.

Lev: I'm interested in how they handle the computational cost. Because molecular dynamics and multislice calculations for these kinds of problems usually generate a massive data flow that just overwhelms you.

The paper's summary: Kai: So the core idea is combining molecular dynamics with a way to evaluate those auxiliary wave functions, which are called TACAW, using this new approach called torched-TACAW. It’s about getting those electron energy loss spectra I(q, E) right.

Mira: The paper explains that they use time auto-correlation of these auxiliary wave functions to get the scattering intensities given by that central equation, I(q, E) proportional to beta E minus the exponential term involving phi MD(q, E) squared <ref:2607.02236#pg1>.

Lev: But what's the catch? The paper points out that when you move to realistic material systems—say, hundreds of thousands of atoms—the standard TACAW workflow becomes very challenging because of the memory use and data flow.

Kai: That's where torched-TACAW comes in. It’s a freely available implementation that focuses specifically on making the data flow tractable by using on-the-fly processing and computational batches to keep memory usage manageable.

Mira: They are also focusing heavily on the way they partition elongated supercells, breaking them down into smaller segments of lower effective thickness so you don't have to simulate one giant, impossibly large unit cell in the molecular dynamics simulation.

Lev: I like that partitioning idea. If you can reduce the size of the supercell being simulated in MD, it directly cuts down on the computational load for generating those trajectories, right?

Kai: Right. And they use machine-learned interatomic potentials like ORB for those molecular dynamics simulations because they offer near ab initio accuracy while being much faster than traditional first-principles methods when you need to simulate large systems.

The paper's improvements: Mira: The paper lists several specific improvements they made to this methodology. One major point is leveraging machine-learned interatomic potentials, like ORB, which the authors say are the recommended choice because they provide near ab initio accuracy with better performance than traditional methods for large systems.

Kai: And another big improvement is how they handle the data flow. They implemented on-the-fly processing and aggregating of data through computational batches to avoid storing terabytes of intermediate wave functions, which is a huge memory saving.

Lev: That sounds good for the memory side, but what about the accuracy when you are actually measuring those diffraction patterns? I mean, how do they ensure spectral leakage isn't an issue?

Mira: They address that by accurately predicting energy resolved diffraction patterns I(q, E) using torched-TACAW. Crucially, they use a Hann window as the default choice for window functions to suppress spectral leakage during that prediction.

Kai: So it’s not just about simulation speed; they are also tightening up the post-processing part so the results are clean and reliable, which is important for experimental comparison.

Lev: From my side, I'm looking at how this would run on actual hardware. Because they focus on data flow management, it suggests that this approach is designed to be computationally efficient enough to run without needing massive distributed computing clusters for every single calculation.

Conclusion: Kai: So to wrap up the "Efficient Large-Scale STEM-EELS Simulations With Torched-TACAW" paper, they show how you can use molecular dynamics with machine learning potentials and this torched-TACAW workflow to handle very large material systems like the eight times eight times sixty-four unit cell of rutile TiO2.

Mira: The main implication is that we can get near ab initio quality simulations for vibrational and magnon excitations in complex materials with defects or interfaces, which is something was previously computationally prohibitive due to memory demands.

Lev: I think what this means for us is that if we want to simulate real-world nanoscale structures that have imperfections, this method gives us a path to do it on current hardware instead of waiting for much more expensive resources just to get those initial trajectories.

Kai: Right. It’s about combining foundational machine-learned interatomic potentials with a multislice workflow in a way that keeps things physically transparent and relatively easy to implement, which is key for adoption.

Mira: The authors also note that they are providing a freely available implementation of the TACAW part of the workflow, which makes it accessible to more researchers looking at this area.

Lev: For real hardware deployment, I'd say the focus on managing data flow and using MLIPs is what matters most for making this actually runnable beyond just theoretical validation.

Kai: So that’s how they set up torched-TACAW to be a practical tool for large-scale simulations of EELS spectra in complex materials.

Department of Physics and Astronomy, Uppsala University · Materials Science & Engineering Department, University of Washington

cond-mat.mtrl-sci, physics.comp-ph

Submitted: 2026-07-02

Updated: 2026-10-28

Code: https://github.com/HamishGBrown/py

Importance score: 97/100

The gist: Efficient large-scale STEM-EELS simulations with torched-TACAW presents a practical methodology for performing near ab initio quality simulations of ultra-low-loss electron energy loss spectra (EELS)

Key concepts

Torched-TACAW
A technique used to evaluate the TACAW signal by propagating it through a time sequence of atomic configurations. It is designed to be efficient for calculating EELS spectra in large systems while maintaining near ab initio accuracy.
Machine-Learned Interatomic Potentials (MLIPs)
Foundational AI models, such as ORB, used to accurately predict interatomic forces and energies. These potentials allow the simulation of very large material systems by providing high-quality physical descriptions without requiring expensive traditional quantum mechanical calculations.
Supercell Partitioning
A strategy where a large material structure is divided along one axis (the z-direction) into smaller segments. This reduces the size of the supercell that needs to be simulated in molecular dynamics, making the simulation computationally tractable for very large systems.
On-the-fly Processing
A data management technique where intermediate wave function calculations are processed immediately as they are generated, rather than storing them all. This keeps memory usage low and manageable when dealing with the massive data flows from multislice calculations.

Terminology

Summary

Efficient large-scale STEM-EELS simulations with torched-TACAW presents a practical methodology for performing near ab initio quality simulations of ultra-low-loss electron energy loss spectra (EELS) arising from vibrational and magnon excitations in large material systems. This approach is significant because it enables the simulation of realistic, large structure models containing defects or interfaces with tractable memory use and data flow.

Methodology Overview

The overall approach combines molecular dynamics based on foundational machine-learned interatomic potentials, partitioning of elongated supercells, and on-the-fly processing of multislice outputs to enable near ab initio quality simulations with tractable memory use and data flow. The workflow involves several key steps:

  1. Molecular dynamics simulations are used to obtain MD trajectories, utilizing machine-learned interatomic potentials (MLIPs) like ORB for their near ab initio accuracy and computational efficiency.

  2. The structure model is partitioned along the z-direction into segments of lower effective thickness, stacked together to reduce the size of the supercell that needs to be simulated in MD.

  3. The TACAW signal is evaluated by performing multislice propagation through the time sequence of atomic configurations using torched-TACAW.

Improvements for AI systems

  1. Improved AI system can perform near ab initio quality simulations of vibrational and magnon excitations in large systems containing defects, interfaces, or grain boundaries by utilizing torched-TACAW for efficient STEM-EELS calculations.

  2. The improved system can handle realistic material systems with up to hundreds of thousands of atoms, such as the 8x8x64 unit cell of rutile TiO2 mentioned in the paper, by employing machine-learned interatomic potentials like ORB and supercell partitioning strategies.

  3. The AI system will accurately predict energy resolved diffraction patterns I(q, E) using the torched-TACAW method, ensuring spectral leakage is suppressed by using a Hann window as the default choice for window functions.

  4. The system can manage large data flows associated with multislice calculations by implementing on-the-fly processing and aggregating of data through computational batches to keep memory usage tractable, avoiding the need to store TBs of intermediate wave functions.

  5. The improved system can maintain physical transparency and relative ease of implementation by combining molecular dynamics based on foundational machine-learned interatomic potentials with a multislice workflow, as described in the torched-TACAW methodology.

Sources

Related papers