BO

Bayesian-optimization-driven active learning for the automated construction of the effective Hamiltonian of BaTiO₃.

Parameterizing an effective Hamiltonian has traditionally required hand-built distorted structures that satisfy symmetry constraints. Here the distorted configurations are proposed by Bayesian optimization (TuRBO trust regions with BoTorch), evaluated with VASP, and the Hamiltonian coefficients are refitted by linear regression after each step. Fewer than 30 distorted configurations were enough to fit the Hamiltonian for BaTiO₃, with Monte Carlo transition temperatures within 10% of experiment.

This is the research code used for the paper below, kept as it was run on the cluster.

Paper

M. Dai et al., Active learning-based automated construction of Hamiltonian for structural phase transitions: a case study on BaTiO₃, J. Phys.: Condens. Matter (2025). doi:10.1088/1361-648X/ad882a

Contents

Path	Content
main.py	active-learning loop: refit coefficients → BO proposes a configuration → VASP evaluates it
bo.py	TuRBO state, candidate generation and acquisition (UCB)
vasp.py	wraps a VASP calculation of a distorted BaTiO₃ cell as a BoTorch test problem
BaTiO3/	reference structure (POSCAR.ori), VASP input files, and the recorded run in Data/

Running

conda create --name BO python=3.11.5
conda activate BO
conda install -c pytorch pytorch torchvision torchaudio
conda install -c conda-forge botorch pymatgen

The scripts submit VASP calculations through Slurm (sbatch job-vasp.sh) and expect a cluster environment: set root_path, path and ssh_root at the top of main.py to your own working directories before running python main.py.

License

MIT