iso-Eg-optimisation

Bayesian optimisation workflow for fitting Hubbard-U parameters to ternary alloy band-gap targets (and tracking auxiliary metrics such as effective mass and short-band agreement).

What This Repository Does

The main driver, process_ternary.py, does the following:

1. Selects one ternary system (InGaAs, InGaSb, InAsSb, or GaAsSb).
2. Uses data_int.hdf5 as a discrete U-parameter grid for the two binaries in that ternary.
3. Proposes a new grid point with Bayesian optimisation.
4. Builds orbital-resolved U values and writes them into QE input files under intermediate/<ternary>/....
5. Reads resulting QE XML outputs to compute:
   • band-gap mismatch loss (loss1),
   • binary short-band deviation (loss0, via delta_*),
   • optional effective-mass metrics (meff_*).
6. Appends results to data_<ternary>.hdf5.

Key Files

• process_ternary.py: Main orchestration script.
• run_test.sh: Slurm launcher example.
• data_int.hdf5: Precomputed U-grid data used for Bayesian search bounds and lookup.
• binary.hse.data.hdf5: Binary HSE reference eigenvalues used in find_delta.
• cube_donati_<ternary>.csv: Experimental/reference Eg points for each ternary.
• intermediate/<ternary>/<composition>/...: QE inputs/outputs for each composition point.
• local/: Helper modules called by process_ternary.py.

Main Local Modules Used

process_ternary.py directly imports and uses:

• local/U_standard_main.py
  • Convert between nested U dictionaries and flattened forms.
• local/make_U.py
  • Interpolate/assemble U values and inject them into .in files.
• local/bandgaps_alloy_bands.py
  • Parse QE XML files and extract Eg values.
• local/do_loss.py
  • Compute sorted pointwise Eg differences and weighted L2 loss.
• local/bayes_class.py
  • Wrapper around bayes_opt for parameter suggestion/registration.
• local/meff.py
  • Parse XML eigenvalues and estimate effective mass.

Requirements

Typical Python dependencies:

• numpy
• h5py
• bayesian-optimization (imported as bayes_opt)

Runtime/external dependencies:

• Quantum ESPRESSO (pw.x) available in PATH.
• MPI launcher (mpirun).
• Cluster environment scripts used in this project:
  • ~/tools/setup_data_qe7.3.1.sh
  • setup_conda.sh

Quick Start

1. Activate environment

Example (adapt to your cluster setup):

source ~/tools/setup_data_qe7.3.1.sh
source setup_conda.sh
conda activate alloy_bayes

2. Run locally

./process_ternary.py InGaAs

3. Run via Slurm

sbatch run_test.sh

Command-Line Usage

python process_ternary.py [TERNARY]

• TERNARY options: InGaAs, InGaSb, InAsSb, GaAsSb
• Default if omitted: InGaAs

Current Execution Mode Note

At the moment, ternary_proc() calls iteration_test(...), which does not launch QE runs (local/run_all.sh is commented there). It assumes output files already exist under intermediate/<ternary>/... and only evaluates/parses them.

If you want full run-and-evaluate behavior, switch to iteration(...) in ternary_proc().

Inputs and Outputs

Inputs (required)

• cube_donati_<ternary>.csv
• data_int.hdf5
• binary.hse.data.hdf5
• Prepared intermediate/<ternary>/... folder structure and QE files

Generated/updated outputs

• data_<ternary>.hdf5
  • Group lambda/<binary>/index
  • Group lambda/<binary>/delta
  • Group Eg/<material> with attrs (Eg_exp, x, y)
  • Group ell/ell1 (loss1) and ell/ell0 (loss0)
• intermediate/<ternary>/data_Eg.csv
• intermediate/<ternary>/intermediate.csv

Troubleshooting

• Missing XML parsing data:
  • Ensure QE XML exists in tmp/ or fallback .save/data-file-schema.xml paths.
• MPI/Slurm launch failures:
  • Verify partition/QoS/time settings in run_test.sh.
  • Confirm SLURM_NPROCS is set and consistent with allocation.
• Empty or unexpected HDF5 output:
  • Check that composition directories and input templates exist under intermediate/<ternary>/.
  • Confirm placeholder replacement by local/make_U.py is occurring.
• pseudopotential directory is incorrect for the user's system
  • change the directory in the QE input files and re-tar them

Repository Notes

• .gitignore excludes many generated files (*.csv, *.txt, intermediate directories, cache directories, etc.).
• Several helper and analysis scripts are present (loss/, U_valley/, conversion utilities), but the core optimisation entry point is process_ternary.py.