MatterSim fine-tuned MLIP for BaSnO₃ and Y-doped BaSnO₃ perovskites with oxygen vacancies

The checkpoint of a fine-tuned MatterSim MLIP [https://doi.org/10.48550/arXiv.2405.04967] based on the pretrained model MatterSim-v1.0.0-5M. The potential was trained on the DFT calculations performed in Quantum Espresso. The training set consisted of the configurations including BaSnO3 and BaSn1-xYxO3 with perovskite structure.

Attached files

bso_bsy_2.0.extxyz - training set with positions, forces and potential energies calculated with DFT, in extended XYZ format. Contains 347 configurations.

bsy_test.extxyz - test set with forces and potential energies calculated with DFT, in extended XYZ format. Contains ... configurations.

BaSnYO3_model.pth - Fine-tuned MatterSim PyTorch weights.

Contents of the training set (bso_bsy_2.0.extxyz)

Training set includes:

• pure BaSnO3 supercells (2x2x2 unit cells) with random displacement of all atoms (sigma = 0.05 angstrom) - 100 configurations
• BaSnO3 supercells (2x2x2) with single oxygen vacancy, with random displacements of atoms - 20 configurations
• BaSnO3 cell with varied lattice parameter (+/- 10% around equilibrium lattice parameter obtained in DFT) - 40 configurations
• Ba8Sn7YO24 cells (2x2x2 BaSnO3 supercell with 1 Sn replaced with Y) with random displacement of all atoms - 97 configurations
• Ba8Sn7YO24 cells with varied lattice parameter (+/- 10% around equilibrium lattice parameter obtained in DFT) - 20 configurations
• Ba8Sn7YO23 cells (same as above with 1 random oxygen vacancy) with random displacement of all atoms - 20 configurations
• 3x3x3 BaSnO3 supercell with 3 random Sn atoms replaced with Y atoms, and 2 random oxygen vacancies (Ba27Sn24Y3O79); the structure was initially relaxed with the MLIP, fine-tuned on all the configurations listed above, and random displacements with sigma=0.05 applied afterwards.  - 50 configurations

The final training set contains 347 configurations. The full training set is attached.

Configurations were generated with the scripts available at https://github.com/alexey-rulev/qe_to_ff

Statistics of the training set is provided in the attached files (distance_distribution_violin.png, force_distribution_per_element.png)

Contents of the Test Set (bsy_test.extxyz)

• 3x3x3 BaSnO3 supercell with 3 random Sn atoms replaced with Y atoms, and 2 random oxygen vacancies (Ba27Sn24Y3O79); the structure was relaxed with the fine tuned MLIP. This way, 20 configurations were generated; then, for each configuraiton 3 configurations with random displacements (sigma=0.05 angstrom) were created, making a total of 60 configurations

Parameters of DFT calculations

All DFT calculations were performed in Quantum Espresso:

ecutrho = 600
ecutwfc = 80
occupations = smearing #needed smearing for correct handling of defect systems
smearing = gaussian
degauss = 0.01
conv_thr = 1.0e-10
Hubbard (ortho-atomic): U O-2p 7
K-points: 6 6 6  1 1 1 for 2x2x2 supercells;
          4 4 4  1 1 1 for 3x3x3 supercells

Core electrons were treated with projector augmented wave pseudopotentials available in the standard solid-state pseudopotentials (SSSP) library (http://materialscloud.org/sssp)

DFT calculations parameters (Hubbard U) were selected to best fit the experimental vibrational data of BaSnO3 (see https://doi.org/10.3390/cryst15050440)

Fine tuning parameters

Fine-tuning was performed using MatterSim training script (https://github.com/microsoft/mattersim/blob/main/src/mattersim/training/finetune_mattersim.py).

The training parameters were following:

• Learning rate lr: 5e-5
• batch_size = 4
• epochs: 200 (early stopping achieved at ~150 epochs)
• re_normalize: True

Other parameters were kept at default values, namely: 

• cutoff: 5.0
• threebody_cutoff: 4.0
• early_stop_patience: 10
• include_forces: True 
• include_stress: False 

Other default values are listed in: https://github.com/microsoft/mattersim/blob/d0a52e64fc1a27c89e60ca41fa818768a67585be/src/mattersim/training/finetune_mattersim.py (accessed 17 Nov 2025, last change 2 Apr 2025)

Convergence of loss function and MAE (energy, forces, stress) is given in learning curves.png

Validation

Values are calculated on the Test Set:

Energy MAE: 0.0074 eV/atom 

Force MAE: 0.0374 eV/Å

Parity plots are given in the attached files (validation_parity_plots.png)

Attached figure (S4.png) demonstrates the phonon dispersion curves of BaSnO3, calculated with pre-trained model MatterSim-v1.0.0-5M, fine-tuned model and DFT.

Intended use / domain of validity

Model is intended for BaSnO₃ and Y-doped BaSnO₃ perovskites with small displacements, lattice strains (±10%), oxygen vacancies, and dilute Y substitution patterns similar to the training set. Extrapolation to other chemistries, high defect concentrations, large deformations, or extreme thermodynamic conditions should be done with caution. The primary goal of this model is to simulate phonon properties of the relevant materials.

Citation

If you use this model, please cite:

1. MatterSim: https://doi.org/10.48550/arXiv.2405.04967

2. This Zenodo record.

3. The experimental reference for BaSnO₃ vibrational data: https://doi.org/10.3390/cryst15050440

Changes in this version: added the test set.