materials-chemistry/ai produced the result/ · v2
Two neural networks predict how heat shrinks the band gap of silver-based crystals
Researchers combined a fine-tuned machine-learning model of atomic forces with a graph neural network trained on quantum-chemistry data to estimate how the light-absorbing properties of mixed silver antiperovskites change with temperature.
spectrum · one line per step, placed by what the step does · bright lines used AI
Machine Learning Modeling of Temperature-Dependent Optoelectronic Properties of Anharmonic Solid Solutions
doi:not-stated · record aix-00046 v2 · checked 2026-10-08
- AI was for
- Property prediction, Simulation surrogate
- Model family
- Graph neural network
- Checked by
- Replication
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Semiconductors absorb and emit light according to their band gap, the energy step an electron must climb to move freely through the crystal. The size of that step sets what a material is useful for, from solar cells to detectors. But a crystal is not a still object: its atoms vibrate, and the hotter it gets the more they move. Those vibrations shift the band gap, so the room-temperature value can differ from the one calculated for a motionless crystal. Working this out from first principles means simulating many atoms jiggling over time and solving the quantum-mechanical electronic problem again and again, which is costly.
The materials here are antiperovskites built from silver and sulphur, in which bromine and iodine are mixed in varying proportions. Such mixtures are disordered: the two halogen atoms can sit in many different arrangements, and each must be considered. These crystals are also anharmonic, meaning their vibrations are not the gentle, spring-like motions that simpler theories assume. The authors set out to map how the band gap of these mixtures varies with both composition and temperature, and to identify which particular vibrations are responsible.
Where AI came in
Two learned models stood in for the expensive quantum calculations. The first, a general-purpose machine-learning interatomic potential called MACE, was fine-tuned on density-functional theory data for thermally displaced 40-atom cells. A potential of this kind predicts the energies and forces on atoms, so it can drive simulations that would otherwise need quantum chemistry at every step. It was used to relax and rank the possible halogen arrangements, compute vibrational spectra at zero and finite temperature, and run molecular dynamics.
The second model, a graph neural network, treats a crystal as a network of atoms joined by bonds and was trained to predict band gaps directly from that structure, first on the cheaper PBEsol data and then on a smaller set computed with the more demanding HSEsol method. It was applied to snapshots from the dynamics and the results averaged to give temperature-dependent gaps, and to individually frozen vibrational patterns to see which ones moved the gap most. Reported test-set errors were below 0.1 eV, and the predictions were compared with previously published simulations and measurements.
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 fine-tuned the MACE interatomic potential on a DFT-PBEsol dataset of 2,200 thermally displaced Ag3SBrxI1-x configurations, and trained a graph neural network to predict band gaps, first on the PBEsol data and then on a further 1,000 configurations computed with the hybrid HSEsol functional. The potential was used to relax chemically disordered configurations, compute phonon dispersions at zero and finite temperature, and run molecular dynamics, while the GNN predicted band gaps of MD snapshots that were averaged to give temperature-renormalised values. Reported test-set R2 values were above 0.8 with mean absolute errors below 0.1 eV, and the predicted band gaps decreased with temperature, with relative reductions of about -38% at 300 K and -44% at 600 K in the compositional range x around 0.6-0.8. Mode-resolved analysis with frozen Gamma-point distortions attributed the largest band-gap reductions, approximately -30%, to low-frequency modes at or below 2 THz dominated by Ag displacements.
How AI was used
Two learned models were combined with first-principles data. A pre-trained universal machine-learning interatomic potential (MACE, mace-mpa-0-medium) was fine-tuned on a PBEsol dataset of total energies, atomic forces and stress tensors computed on 40-atom 2x2x2 supercells carrying quasi-harmonic thermal displacements and Br/I substitutions, using a 95/5 train/validation split, 256 equivariant messages, an 8 A cutoff and energy, force and stress weights of 1, 150 and 10. In parallel, a graph neural network implemented in PyTorch Geometric, operating on radius-cutoff graphs (5.5 A) with atomic number, mass, radius and electronegativity as node features and interatomic distance as the edge feature, was trained on the PBEsol band gaps and then fine-tuned on the smaller HSEsol band-gap dataset; architecture choices were set by a hyperparameter exploration. The fine-tuned potential performed ionic relaxations and energy ranking of symmetry-inequivalent substitution patterns, zero-temperature phonon calculations, normal-mode-decomposition anharmonic phonons from MD trajectories, and 50 ps Langevin MD runs in ASE. The GNN was then evaluated on statistically uncorrelated MD snapshots, and the resulting band gaps were averaged over configurations to obtain temperature-renormalised values; the same model was applied to frozen Gamma-point phonon distortions for mode-resolved electron-phonon analysis, with selected configurations recomputed in DFT for comparison.
The shape of the work
Structural · the record, drawn
no AI
Build DFT datasets on thermally displaced configurations
Obtaining raw data, whether by measurement, download or retrieval.
For each displaced configuration, total energies, atomic forces, mechanical stresses, and electronic band gaps were computed using the semilocal PBEsol exchange-correlation functionalwhere the paper describes this · verbatim
AI
Fine-tune MACE interatomic potential
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
The MACE model was subsequently fine-tuned on our PBEsol dataset comprising total energies, atomic forces, and stress tensorswhere the paper describes this · verbatim
AI
Train GNN band-gap model in two stages
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
The model was first trained on the larger PBEsol dataset and subsequently fine-tuned on the smaller HSEsol datasetwhere the paper describes this · verbatim
AI
Sample chemical disorder, relax and rank configurations
Reducing a candidate set by filtering or ranking, in a single pass. The AI stood in for simulation.
all symmetry-inequivalent configurations were fully relaxed using the fine-tuned MLIP and ranked according to their equilibrium energieswhere the paper describes this · verbatim
AI
Compute phonons and run finite-temperature molecular dynamics
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
The relaxed lowest-energy configurations are subsequently assessed for vibrational stability by computing its phonon frequencies with the MLIPwhere the paper describes this · verbatim
AI
Predict band gaps of MD snapshots and ensemble-average
Running a trained model over new data to predict, classify or score. The AI stood in for simulation.
electronic band gaps are computed for statistically uncorrelated MD snapshots using the GNN modelwhere the paper describes this · verbatim
no AI
Compare predictions with prior DFT-AIMD and experimental band gaps
Testing outputs against ground truth.
The calculated band gaps at 0, 200, 400, and 600 K show excellent agreement, within the numerical uncertainties, with previous DFT-AIMD resultswhere the paper describes this · verbatim
AI
Mode-resolved electron-phonon analysis with frozen phonon distortions
Extracting understanding from model behaviour. The AI stood in for simulation.
we quantified the band-gap variations induced by frozen Γ -point phonon distortions of the reference antiperovskite structurewhere 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 reported temperature-renormalised band gaps of the solid solutions are produced by the fine-tuned MLIP (molecular dynamics, relaxations, phonons) and the GNN band-gap model; no non-ML route to these numbers is reported for the disordered compositions.
By integrating machine-learning interatomic potentials (MLIP) with predictive graph neural networks (GNN) models trained on DFT datawhere the paper describes this · verbatim
the agreement between first-principles predictions and experiment is excellent across all investigated compositionswhere the paper describes this · verbatim
All relevant scripts, including model training workflows and data preprocessing routines, are freely accessible in the GitHub 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.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of GNN band-gap model (PyTorch Geometric graph convolutional model)Which version of the model was used is not stated.
About this article
Record aix-00046, 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