materials-chemistry/ai produced the result/arXiv 2024 · v2
Machine-learned potentials trained to simulate copper ion conductor Cu7PS6 faster
Researchers fitted two machine-learned interatomic potentials to quantum-mechanical calculations on the copper compound Cu7PS6, then used them in place of those calculations to simulate atomic structure and vibrations.
spectrum · one line per step, placed by what the step does · bright lines used AI
Constructing accurate machine-learned potentials and performing highly efficient atomistic simulations to predict structural and thermal properties
arXiv, 2024
doi:10.48550/arxiv.2411.10911 · record aix-00144 v2 · checked 2026-10-09
- AI was for
- Simulation surrogate, Experimental design
- Model family
- Multilayer perceptron, Linear model
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
To predict how a material behaves — how its atoms sit, how they jostle, how heat moves through it — you need to know the force on every atom at every moment. The most trustworthy way is to work it out from quantum mechanics, using a method called density functional theory. It is accurate but slow. Running it step by step through time, which is called ab initio molecular dynamics, limits you to a few hundred atoms for a few thousandths of a nanosecond. Many properties of interest only show up in bigger samples over longer times.
The material here is Cu7PS6, a superionic conductor, meaning its copper ions can move through the solid almost as if it were a liquid. The researchers set out to build cheaper stand-ins for the quantum calculations and then check whether simulations driven by those stand-ins reproduced the same structural and vibrational behaviour.
Where AI came in
The AI is the force calculator itself. A moment tensor potential, a fitted mathematical form, and a neuroevolution potential, a small neural network, were both trained from scratch on energies, forces and stresses computed by density functional theory for 300 atomic arrangements drawn from ab initio molecular dynamics. An active-learning loop then let the moment tensor potential flag arrangements it was least sure about; those were sent for fresh quantum calculations and added, reaching 354 arrangements in total. One tenth of the data was held back to test the fit, where the neuroevolution potential's energies differed from the quantum reference by 0.45 meV per atom.
With the potentials trained, they stood in for the quantum calculations inside molecular dynamics runs of 56,000 atoms over 100 picoseconds — far beyond what the quantum method was used for here. From those runs the team extracted radial distribution functions, which describe how far apart atoms typically sit, and the phonon density of states, which describes the material's vibration frequencies, and compared both with the ab initio reference at 300 K. The neuroevolution potential is reported running about 41 times faster than the moment tensor potential and about 15 times faster than a GPU-accelerated DeePMD model.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
Two machine-learned interatomic potentials for the superionic conductor Cu7PS6 — a moment tensor potential and a neuroevolution potential — were fitted to energies, forces and stresses from DFT calculations on configurations drawn from ab initio molecular dynamics, with active learning used to extend the training set to 354 configurations. Reported root-mean-square errors on the held-out 10% of configurations are 0.59 meV/atom and 0.046 eV/Å for the MTP and 0.45 meV/atom and 0.037 eV/Å for the NEP. Radial distribution functions and phonon density of states computed from molecular dynamics with both potentials at 300 K are compared with ab initio molecular dynamics references, and the paper reports the NEP running approximately 41 times faster than the MTP and about 15 times faster than GPU-accelerated DeePMD.
How AI was used
Ab initio molecular dynamics on a 448-atom Cu7PS6 supercell at temperatures up to 700 K under the NVT ensemble, with a 1 fs time step and 1000 steps per temperature, supplied an initial 300 configurations sampled every ten steps. A moment tensor potential with a 6 Å cutoff was fitted to DFT energies, forces and stresses from these configurations, then used inside LAMMPS molecular dynamics at 300, 500 and 700 K with an extrapolation-grade criterion to select further configurations for DFT labelling and refitting; this active-learning loop yielded 354 representative configurations. A neuroevolution potential was trained on the same data with radial and angular cutoffs of 7.0 Å and 5.0 Å and separate loss weights for energies, forces and stresses, holding out 10% of configurations as a validation set. Both potentials were then run as surrogates for DFT in molecular dynamics — a 56000-atom supercell, Gaussian-initialised velocities, 1 fs time step, equilibration followed by a 100 ps NVE production run with a 20 ps correlation time — from which radial distribution functions and the phonon density of states (via Fourier transform of the velocity autocorrelation function) were computed and compared with ab initio molecular dynamics. Runtimes of the two potentials, a DeePMD model and an empirical potential were measured on GPU and CPU hardware.
The shape of the work
Structural · the record, drawn
no AI
Generate reference configurations by AIMD and DFT
Obtaining raw data, whether by measurement, download or retrieval.
Structures for initial training were sampled every ten steps, yielding 300 configurations.where the paper describes this · verbatim
AI
Fit moment tensor potential
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
The moment tensor potential (MTP) was constructed based on reference data generated through high-accuracy DFT calculations.where the paper describes this · verbatim
AI
Enrich training set by active learning
Iterative search over a space. The AI stood in for manual curation. Its result feeds back into an earlier step.
After active learning, a total of 354 representative configurations were selected for final MTP construction.where the paper describes this · verbatim
AI
Train neuroevolution potential
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
Radial and angular cutoffs were set to 7.0 Å and 5.0 Å, respectively, to define the interaction range.where the paper describes this · verbatim
AI
Evaluate energy and force errors on held-out set
Testing outputs against ground truth. The AI stood in for simulation.
To evaluate model performance, 10% of the configurations were excluded from training and used as a validation setwhere the paper describes this · verbatim
AI
Run machine-learning molecular dynamics
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
Phonon density of states (DOS) and radial distribution functions (RDF) were computed using molecular dynamics (MD) simulations in the GPUMD framework.where the paper describes this · verbatim
no AI
Compute RDF and phonon DOS and compare with AIMD
Testing outputs against ground truth.
RDFs were obtained for both total and partial pair correlations at 300 K using a 448-atom supercell.where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The machine-learned potentials are themselves the object of the study and produce every reported structural, vibrational and timing result; the compared AIMD/DFT calculations serve as reference
we introduce a neuroevolution potential (NEP), trained on a dataset generated from ab initio molecular dynamics (AIMD) simulationswhere the paper describes this · verbatim
NEP achieves RMSEs of 0.44 and 0.45 meV/atom for training and validation sets, respectivelywhere the paper describes this · verbatim
The computational speed of the NEP (running with one NVIDIA RTX 3080 implemented in GPUMD)where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of neuroevolution potential (NEP)Which version of the model was used is not stated.
- Version of moment tensor potential (MTP)Which version of the model was used is not stated.
- Version of DeePMD (DP)Which version of the model was used is not stated.
About this article
Record aix-00144, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error