materials-chemistry/ai produced the result/ · v2
Neural network potential used to simulate how calcium oxide melts under pressure
Researchers trained a neural network to reproduce quantum-mechanical forces between atoms in calcium oxide, then used it to run molecular dynamics on supercells of 10,648 to 17,280 atoms and compute melting temperatures up to 20 GPa.
spectrum · one line per step, placed by what the step does · bright lines used AI
Melting Behavior and Phase Stability of CaO from Neural Network Potentials: a Molecular Dynamics Study
doi:not-stated · record aix-00119 v2 · checked 2026-10-08
- AI was for
- Simulation surrogate
- Model family
- Multilayer perceptron
- Checked by
- Held-out
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Calcium oxide, or lime, is a simple compound of calcium and oxygen that stays solid to very high temperatures. Knowing exactly where it melts matters for furnace linings, cements and models of the Earth's interior, but such temperatures are awkward to measure in the laboratory. One alternative is to simulate the atoms directly. Methods that solve the quantum equations for every electron give reliable forces, yet they are so costly that only a few hundred atoms can be followed for a few picoseconds. Melting, however, is a collective affair: it needs enough atoms for a liquid region to form and enough time for solid and liquid to settle against each other.
The authors set out to bridge that gap for calcium oxide. They wanted melting temperatures at ordinary pressure and along a curve up to 20 gigapascals, together with the energy absorbed on melting and the volume change that accompanies it, all from simulations large and long enough for the question to be meaningful.
Where AI came in
The artificial intelligence here is a stand-in for the quantum calculation. The team first ran costly first-principles simulations on small cells to produce about 12,000 snapshots of calcium oxide, covering solid, liquid, solid-liquid interfaces, cells containing voids and cells squeezed or stretched to different pressures, each labelled with the energies and forces on its atoms. Each atom's surroundings were encoded as a list of 900 numbers describing neighbours within five angstroms. A three-layer feed-forward neural network was then trained to predict energies and forces from that description, reaching an average error below 5 meV per atom on energy.
Once trained, the network replaced the quantum calculation inside standard molecular dynamics. It supplied the forces for runs on supercells of 10,648 to 17,280 atoms, stepping forward half a femtosecond at a time for at least 300 picoseconds. Those runs produced the melting temperatures, by watching crystals melt outwards from a void and by holding solid and liquid in contact, as well as the heating and cooling curves used for the energy and volume of melting. Small first-principles runs on a 216-atom cell, and an older hand-built model of the forces, were kept as points of comparison.
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 interatomic potential for calcium oxide on about 12,000 ab initio molecular dynamics configurations covering solid, liquid, interfacial, voided and pressure-varied structures, then used it to run molecular dynamics on supercells of 10,648 to 17,280 atoms. Melting at ambient pressure was computed by void-nucleated melting, giving Tm = 3055 ± 11 K, and by two-phase coexistence, giving Tm = 2847 ± 15 K. The enthalpy of fusion was 73.46 kJ/mol at 3055 K and 71.59 kJ/mol at 2847 K, with a volume expansion of about 29% at both reference melting temperatures. A high-pressure melting curve was computed to 20 GPa, over which the overheating ratio relative to the thermal instability temperature rose from about 17% at ambient pressure to 24% at 20 GPa.
How AI was used
A machine-learning interatomic potential was fitted to reproduce ab initio energies and forces so that classical molecular dynamics could be run at the system sizes and durations the melting calculations required. Reference data came from ab initio molecular dynamics in Quantum ESPRESSO with the PBEsol functional on 3 × 3 × 3 and 2 × 2 × 4 supercells, sampling solid structures from 300 to 3200 K, liquid structures from 2000 to 6000 K, solid-liquid interfaces near the expected melting point, cells with central voids from 1% to 40% in volume, and expanded or compressed cells spanning pressures to 20 GPa. The resulting dataset of roughly 12,000 configurations was split into training, validation and test portions of about 75%, 15% and 10%. Local environments were encoded with the LATTE descriptor, built from five tensor contraction terms summing to 900 features within a 5.0 Å cutoff, and passed to a three-layer feed-forward network of shape 256:124:1 with Gaussian hidden activations and a linear output, trained for 3000 epochs on an exponentially decaying learning rate from 1e-4 to 1e-7. The trained potential then drove LAMMPS molecular dynamics with a 0.5 fs time step for the void-nucleated melting runs on an 11 × 11 × 11 supercell, the two-phase coexistence runs on elongated supercells up to 6 × 6 × 60, the thermal instability runs on a defect-free supercell, and the stepwise heating and cooling runs used for the caloric and thermal expansion curves; separate ab initio runs on a 216-atom supercell provided comparison data.
The shape of the work
Structural · the record, drawn
no AI
Generate ab initio reference configurations
Numerical or physics simulation, including where a learned surrogate replaces it.
To build the dataset, we carried out AIMD simulations using GPU-enabled Quantum ESPRESSOv6.8 packagewhere the paper describes this · verbatim
no AI
Assemble and split training dataset
Cleaning, filtering, normalising or labelling data already obtained.
The dataset is composed of ∼ 12,000 structural configurationswhere the paper describes this · verbatim
AI
Encode local atomic environments as descriptors
Encoding data into features, descriptors, embeddings or graphs. The AI stood in for conventional algorithm.
In this work, we employ LATTE descriptor, which is based on Cartesian tensor contractionswhere the paper describes this · verbatim
AI
Train the interatomic potential
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
The potential uses a fully-connected, deep, feed-forward Neural Network (FFNN) architecture consisting of three layers, 256:124:1.where the paper describes this · verbatim
AI
Run void-nucleated melting simulations
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
A void is created in the relaxed crystal structure of an 11 × 11 × 11 supercellwhere the paper describes this · verbatim
AI
Run two-phase coexistence simulations
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
The second method we applied to calculate the melting temperature of CaO is the Two-Phase Coexistence (TPC) method.where the paper describes this · verbatim
no AI
Refine melting temperatures to convergence
Iterative search over a space. Its result feeds back into an earlier step.
the process is repeated using the new refined melting temperature as the estimated melting pointwhere the paper describes this · verbatim
AI
Compute caloric and thermal expansion curves and compare with reference data
Testing outputs against ground truth. The AI stood in for simulation.
we employed both ab-initio molecular dynamics (AIMD) and classical molecular dynamic (CMD) with the MLIP to calculate the caloric curve at ambient pressurewhere 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 melting temperatures, enthalpy and volume curves and the high-pressure melting curve are produced by molecular dynamics driven by the machine-learned interatomic potential trained in this study; the ab initio and empirical-potential results serve as reference
We employ a multilayer perceptron (MLP) architecture, chosen for its computational efficiency in large-scale MD production runswhere the paper describes this · verbatim
The training performance was verified in terms of the Mean Absolute Error (MAE) on energy and forceswhere the paper describes this · verbatim
The dataset generated during the current study is fully available in the Zenodo repositorywhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- How many were testedThe paper gives no count of what was tested.
About this article
Record aix-00119, 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