Structure-file-driven OpenMX/ABACUS-to-HamGNN training pipeline. Use when users provide a crystal structure file such as CIF or POSCAR and want to build a perturbed CIF dataset, run DFT preprocessing (OpenMX by default, ABACUS when explicitly requested), and train a HamGNN model, stopping at the trained checkpoint.

2026-04-207

name	abacus-data-processing
description	Prepare ABACUS calculation inputs from CIF/POSCAR with default settings and generate per-structure INPUT/KPT/STRU run directories.

ABACUS Data Processing

Reference usage policy:

Do not bulk-read reference/* during normal execution.
Before finalizing generated config files, quickly scan reference/ROUTING.md for strong intent matches.
If a very close route is found, open only the mapped markdown(s) (and minimal companion example files if needed), then adapt parameters to that pattern instead of blindly keeping defaults.
If no strong route is found, keep the template/script defaults.
When discussing setup choices with the user, you may reference relevant files under reference/* as supporting evidence.
Keep reference loading minimal and targeted; do not ingest unrelated documents.

Prepare one split-local ABACUS workdir per dataset split. Do not run train and test generation in one shared mutable directory.
Seed each split-local workdir from <skill_root>/templates/abacus/:

abacus_input_gen.yaml
run_abacus.sh

Render the copied template into the split-local workdir. Do not edit skill-owned templates in place.
In abacus_input_gen.yaml, render split-specific fields only:

structure_glob
output_root
optional system_name

In run_abacus.sh, only render split-specific values such as SYSTEM_NAME, INPUT_ROOT, and DATASET_ROOT. Preserve the runtime environment contract.
Keep default settings unless the user explicitly requests overrides. Default baseline:

calculation=scf
basis_type=lcao
ecutwfc=100, scf_thr=1e-7, scf_nmax=200
out_mat_hs2=1 (export sparse H/S matrices for downstream graph generation)
kpt=Gamma 4x4x4

Runtime environment policy:

run via <skill_root>/scripts/data_processing.sh
for input generation, default conda env is ${ABACUS_CONDA_ENV:-OpenMX} (contains pymatgen and pyyaml); if unavailable, fall back to base
for ABACUS execution via run_abacus.sh, default conda env is ${ABACUS_CONDA_ENV:-abacus_env}
only override ABACUS_CONDA_ENV or PYTHON_BIN when explicitly needed
logs are written to <workdir>/logs/abacus_input_gen.log

Use the split-local entry command:

<skill_root>/scripts/data_processing.sh --workdir <abacus_split_root> --config abacus_input_gen.yaml

For ABACUS batch execution against generated case directories, use:

bash ./run_abacus.sh [jobs] [omp_threads]
default command is ABACUS_CMD=${ABACUS_CMD:-${ABACUS_BIN}} and can be overridden for site-specific launchers.

Input policy:

CIF input is first-class
POSCAR/CONTCAR/vasp-like files are also allowed when readable by pymatgen
fail fast when structure_glob resolves to zero files
placeholders like __CIF_GLOB__ / __ABACUS_INPUT_ROOT__ must be rendered before execution

Generation policy:

output is one run folder per structure under output_root
each run folder contains INPUT, KPT, STRU, and source.* for traceability
default naming is index_stem
default overwrite=false; existing run dirs must fail instead of silent overwrite
default copy_source_file=true
default move_flags=[1,1,1]

Mapping and safety guardrails:

validate pseudopotential mapping (PP_DICT) for all species
when basis_type=lcao, validate orbital mapping (ORB_DICT) for all species
if any element mapping is missing, stop and report missing symbols; do not guess

Multi-split isolation:

train/test (or other splits) must use independent output_root
do not let one split overwrite another split's generated run folders

Returnables for downstream stages:

generated output_root per split
generated run-folder list
generation log path (logs/abacus_input_gen.log)
when run_abacus.sh is executed, return ABACUS run log path (logs/run_abacus_*.log)
clear failure reason when generation stops