materials-chemistry/ai produced the result/arXiv 2024 · v2
Neural network force field simulates how a ferroelectric crystal's molecules begin to spin
Researchers trained a neural network to predict the forces between atoms in the molecular ferroelectric HdabcoClO4, then used it to run molecular dynamics on crystals of 11,232 atoms at temperatures from 120 to 500 K.
spectrum · one line per step, placed by what the step does · bright lines used AI
Dynamical Disorder in the Mesophase Ferroelectric HdabcoClO4: A Machine-Learned Force Field Study
arXiv, 2024
doi:10.48550/arxiv.2410.15746 · record aix-00091 v2 · checked 2026-10-08
- AI was for
- Simulation surrogate
- Model family
- Multilayer perceptron
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Some crystals are built not from single atoms but from whole molecules, stacked in a repeating pattern. In a ferroelectric crystal, those building blocks are arranged so that the material carries an electrical polarity of its own, which can be flipped by an applied field. When such a crystal is warmed, its molecules start to move: they rattle, then rotate, and the neat ordering that gave the crystal its polarity breaks down. Working out exactly which molecules start turning, and at what temperature, is hard because the motion is messy and happens at the scale of single molecules over very short times.
The salt studied here, HdabcoClO4, is made of two kinds of molecular ion: a positively charged Hdabco+ unit that carries an extra proton, and a negatively charged perchlorate, ClO4-. The researchers set out to follow, in simulation, how each of these two parts begins to rotate as the temperature rises, and to see whether the protons hop between neighbouring Hdabco+ molecules. They then compared the temperatures at which disorder set in with phase transition temperatures already reported from experiment.
Where AI came in
Simulating atoms honestly means calculating the forces between them from quantum mechanics, using a method called density functional theory. That is accurate but slow, so it can only follow a few hundred atoms for a very short time. The team instead used those expensive calculations as worked examples and trained a neural network, NeuralIL, to predict the same energies and forces from an atom's surroundings. The network learned from a training set that grew to 2,180 configurations.
Part of the training was itself steered by the models. Ten copies of the network were trained from different starting points, and configurations where they disagreed most about the forces were singled out for a fresh quantum calculation and added to the training set. The finished force field then stood in for the quantum method during the molecular dynamics runs, which is what allowed the 11,232-atom crystals and longer timescales. Every dynamical result reported, including the thermal expansion, the rotations and the proton hops, comes from those simulations.
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
The authors trained a neural-network force field with NeuralIL on density-functional-theory molecular dynamics data for the molecular ferroelectric HdabcoClO4, expanding the training set through committee-based active learning and through strained and atom-swapped configurations until it contained 2180 configurations. Using the force field, they ran 19 molecular dynamics simulations with 11,232-atom supercells at fixed temperatures between 120 and 500 K. The simulations show a low-temperature transition near 200 K coinciding with the onset of occasional ClO4- rotation and a higher-temperature transition near 380 K where both ClO4- and Hdabco+ rotate, which the authors compare with the reported experimental transition temperature of 377 K. Proton transfer between Hdabco+ molecules was observed from 150 K upwards, with chain-orientation switching about two orders of magnitude less frequent than individual proton transfer.
How AI was used
A machine-learned force field was fitted to DFT reference data so that molecular dynamics could be run at supercell sizes and timescales beyond ab initio molecular dynamics. DFT-MD simulations in VASP with the vdW-DF-cx functional on a 416-atom cell supplied initial training configurations. NeuralIL, a residual neural network built on Jax and Flax with core widths 64:32:16, encoded each atom's environment within a 4.0 A cutoff as spherical Bessel descriptors with element embeddings, and was trained for 25 epochs with batch size eight using the VeLO optimiser, weighting energies 0.4 and forces 0.6, with 20% of configurations held out for validation in each iteration. Training proceeded in four stages: a crude model, six iterations adding DFT-MD configurations with the largest force deviations, four rounds of committee-based active learning in which ten models trained from different random initialisations flagged the 200 sampled configurations with the largest force standard deviations for DFT recomputation, and a final stage adding 480 cell-scaled and 200 atom-swapped configurations. The resulting force field drove NPT molecular dynamics in Jax-MD with a Nose-Hoover chain thermostat and barostat at 1 bar, on 6x6x6 supercells of phase II, with 30 ps thermalisation and 180 ps production runs.
The shape of the work
Structural · the record, drawn
no AI
Generate DFT-MD reference data
Numerical or physics simulation, including where a learned surrogate replaces it.
Initial DFT-MD simulations were carried out at different temperatures and volumes to obtain diverse yet physically representative starting training datawhere the paper describes this · verbatim
AI
Train machine-learned force field
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
The first MLFF model was based on 400 configurations randomly selected from DFT-MD simulations C and Dwhere the paper describes this · verbatim
AI
Committee-based active-learning selection and DFT labelling
Iterative search over a space. The AI stood in for manual curation. Its result feeds back into an earlier step.
The standard deviation in the force predictions for the predictions of the committee was then used to identify atomic configurationswhere the paper describes this · verbatim
no AI
Construct strained and atom-swapped training configurations
Cleaning, filtering, normalising or labelling data already obtained. Its result feeds back into an earlier step.
480 configurations were constructed by scaling the unit cell parameters with a factor ranging from 0.9 to 1.1 in increments of 0.1where the paper describes this · verbatim
AI
Run MLFF molecular dynamics across temperatures
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
In total, 19 simulations with fixed temperatures in the range between 120 and 500 K were performedwhere the paper describes this · verbatim
no AI
Analyse disorder, displacement and proton transfer
Extracting understanding from model behaviour.
The rotational disorder of the ClO4− -molecules is evaluated using a rotational autocorrelation functionwhere the paper describes this · verbatim
no AI
Compare computed transitions with reported experiment
Testing outputs against ground truth.
The computational results are consistent with experimental measurements overall.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.
All reported dynamical results (thermal expansion, order parameter, rotational autocorrelation, proton transfer) come from molecular dynamics driven by the trained machine-learned force field
The machine-learned force field was trained using the neural-network-based NeuralILwhere the paper describes this · verbatim
20% of the configurations in the training set were randomly set aside for validation in each iteration of the trainingwhere the paper describes this · verbatim
All training data can be accessed through the Nomad database with DOI:10.17172/NOMAD/2024.09.19-1.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.
- How many were testedThe paper gives no count of what was tested.
- Version of NeuralIL machine-learned force fieldWhich version of the model was used is not stated.
About this article
Record aix-00091, version 2, checked by a person on 2026-10-08. 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