2_training/forces/train_force.py

#!/usr/bin/env python
"""Train NequIP force field for diamond; compare DFPT and ML phonons at Gamma.

Steps:
  1. Build forces/dataset.xyz from QE displaced configs
     (re-runs SCF with tprnfor=.true. for any disp without forces)
  2. Run DFPT (ph.x) on pristine UC at q=Gamma
  3. Train NequIP MLIP (nequip-train train.yaml) in deeph conda env
  4. Compile model to diamond_ase.nequip.pt2 (nequip-compile)
  5. Run phonopy + NequIP calculator to get ML Gamma frequencies
  6. Plot + print comparison: DFPT vs ML phonons

Usage:
  python train_force.py [params.json]
  python train_force.py [params.json] --skip-training   # use existing checkpoint
"""
import glob
import json
import os
import re
import subprocess
import sys

import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt

SCRIPT_DIR     = os.path.dirname(os.path.abspath(__file__))
DATA_DIR       = os.path.abspath(os.path.join(SCRIPT_DIR, '..', '..', '1_data_prepare', 'data'))
PARAMS_DEFAULT = os.path.abspath(os.path.join(SCRIPT_DIR, '..', '..', '1_data_prepare', 'params.json'))

DATASET_XYZ  = os.path.join(SCRIPT_DIR, 'dataset.xyz')
DFPT_DIR     = os.path.join(SCRIPT_DIR, 'dfpt')
OUTPUTS_DIR  = os.path.join(SCRIPT_DIR, 'outputs')
PT2_PATH     = os.path.join(SCRIPT_DIR, 'diamond_ase.nequip.pth')
PT2_REF_PATH = os.path.join(SCRIPT_DIR, 'diamond_ase_ref.nequip.pt2')
TRAIN_YAML   = os.path.join(SCRIPT_DIR, 'train.yaml')

# Reference QE phonon frequencies (phonopy+QE finite difference, not DFPT)
REF_QE_QPOINTS_YAML = (
    '/home/apolyukhin/scripts/ml/diamond-qe/diamond_epc/displacements/qpoints.yaml'
)

RY2EV         = 13.605693122994      # Rydberg → eV
RY_BOHR2EVANG = 25.71104309541616    # Ry/bohr → eV/Å


def load_params(path=None):
    with open(path or PARAMS_DEFAULT) as f:
        return json.load(f)


# ---------------------------------------------------------------------------
# Step 1: Build forces dataset
# ---------------------------------------------------------------------------

def has_forces_in_pwout(pw_out_path):
    """Return True if pw.out already contains force output."""
    try:
        with open(pw_out_path) as f:
            for line in f:
                if 'Forces acting on atoms' in line:
                    return True
    except FileNotFoundError:
        pass
    return False


def add_tprnfor_to_pwin(pw_in_path):
    """Patch pw.in to add tprnfor = .true. in &CONTROL (idempotent)."""
    with open(pw_in_path) as f:
        content = f.read()
    if 'tprnfor' in content:
        return
    content = content.replace('&CONTROL', '&CONTROL\n tprnfor = .true.')
    with open(pw_in_path, 'w') as f:
        f.write(content)


def read_qe_forces(pw_out_path):
    """Parse energy (eV) and forces (eV/Å) from a QE pw.out with tprnfor."""
    with open(pw_out_path) as f:
        text = f.read()

    energy_match = re.search(r'!\s+total energy\s+=\s+([-\d.]+)\s+Ry', text)
    if not energy_match:
        raise ValueError(f'No converged energy in {pw_out_path}')
    energy_ev = float(energy_match.group(1)) * RY2EV

    # Locate "Forces acting on atoms" header and read only the TOTAL forces block.
    # QE also prints decompositions (non-local, ionic, ...) with the same
    # "atom N type T force = fx fy fz" format, so we stop at the first blank line
    # or "The ... to forces" line after the initial block.
    m = re.search(r'Forces acting on atoms.*?\n', text)
    if not m:
        raise ValueError(f'No forces in {pw_out_path}')
    block_start = m.end()

    forces = []
    for line in text[block_start:].split('\n'):
        if not line.strip():
            if forces:
                break   # blank line ends the total-force block
            continue
        if line.strip().startswith('The '):
            break       # start of decomposition sub-section
        parts = line.split()
        if len(parts) >= 9 and 'atom' in parts:
            forces.append([float(parts[-3]), float(parts[-2]), float(parts[-1])])
    if not forces:
        raise ValueError(f'Could not parse force lines in {pw_out_path}')
    return energy_ev, np.array(forces) * RY_BOHR2EVANG


def build_forces_dataset(params):
    """Build dataset.xyz from all disp-XX QE outputs; re-run SCF if forces missing."""
    if os.path.exists(DATASET_XYZ):
        print(f'  Dataset exists: {DATASET_XYZ}')
        return

    from ase.io import read as ase_read, write as ase_write
    from ase.calculators.singlepoint import SinglePointCalculator

    qe_setup  = params['paths']['qe_setup']
    np_qe     = params['execution']['mpi_np']
    disp_dirs = sorted(glob.glob(os.path.join(DATA_DIR, 'disp-*')))
    print(f'  Processing {len(disp_dirs)} displaced configs...')

    atoms_list = []
    for disp_dir in disp_dirs:
        label   = os.path.basename(disp_dir)
        scf_dir = os.path.join(disp_dir, 'scf')
        pw_in   = os.path.join(scf_dir, 'pw.in')
        pw_out  = os.path.join(scf_dir, 'pw.out')

        if not has_forces_in_pwout(pw_out):
            print(f'  [{label}] Re-running SCF with tprnfor=.true. ...')
            add_tprnfor_to_pwin(pw_in)
            bash_cmd = (f"source {qe_setup} 2>/dev/null; "
                        f"mpirun -np {np_qe} pw.x < pw.in > pw.out 2>&1")
            subprocess.run(['bash', '-c', bash_cmd], cwd=scf_dir, check=True)
        else:
            print(f'  [{label}] Forces already in pw.out')

        energy_ev, forces_evang = read_qe_forces(pw_out)
        atoms = ase_read(pw_out, format='espresso-out')
        atoms.calc = SinglePointCalculator(
            atoms, energy=energy_ev, forces=forces_evang)
        atoms_list.append(atoms)
        fmax = np.abs(forces_evang).max()
        print(f'  [{label}] E={energy_ev:.4f} eV  |F|max={fmax:.4f} eV/Å')

    ase_write(DATASET_XYZ, atoms_list, format='extxyz')
    print(f'  Wrote {len(atoms_list)} structures → {DATASET_XYZ}')


# ---------------------------------------------------------------------------
# Step 2: DFPT at Gamma
# ---------------------------------------------------------------------------

def run_dfpt(params):
    """Run ph.x on pristine UC at q=Gamma; skip if ph.out already exists."""
    ph_out = os.path.join(DFPT_DIR, 'ph.out')
    if os.path.exists(ph_out):
        print(f'  DFPT output exists: {ph_out}')
        return

    uc_scf_dir = os.path.join(DATA_DIR, 'bands', 'uc', 'scf')
    os.makedirs(DFPT_DIR, exist_ok=True)

    ph_in_text = (
        "Phonons at Gamma\n"
        "&inputph\n"
        "  prefix = 'diamond'\n"
        f"  outdir = '{uc_scf_dir}/'\n"
        f"  fildyn = '{DFPT_DIR}/dynmat.dat'\n"
        "  ldisp = .false.\n"
        "/\n"
        "0.0 0.0 0.0\n"
    )
    with open(os.path.join(DFPT_DIR, 'ph.in'), 'w') as f:
        f.write(ph_in_text)

    qe_setup = params['paths']['qe_setup']
    np_qe    = params['execution']['mpi_np']
    bash_cmd = (f"source {qe_setup} 2>/dev/null; "
                f"mpirun -np {np_qe} ph.x < ph.in > ph.out 2>&1")
    print('  Running ph.x at q=Gamma ...')
    subprocess.run(['bash', '-c', bash_cmd], cwd=DFPT_DIR, check=True)
    print(f'  DFPT done: {ph_out}')


def parse_dfpt_gamma_freqs():
    """Parse Gamma phonon frequencies from ph.out → (thz_array, cm1_array)."""
    ph_out = os.path.join(DFPT_DIR, 'ph.out')
    freqs_thz, freqs_cm1 = [], []
    with open(ph_out) as f:
        for line in f:
            m = re.search(
                r'freq\s*\(\s*\d+\)\s*=\s*([\d.]+)\s*\[THz\]\s*=\s*([\d.]+)\s*\[cm-1\]',
                line)
            if m:
                freqs_thz.append(float(m.group(1)))
                freqs_cm1.append(float(m.group(2)))
    return np.array(freqs_thz), np.array(freqs_cm1)


def parse_phonopy_qpoints_yaml(yaml_path=None):
    """Read Gamma phonon frequencies from a phonopy qpoints.yaml → (thz, cm1)."""
    import yaml as _yaml
    path = yaml_path or REF_QE_QPOINTS_YAML
    with open(path) as f:
        data = _yaml.safe_load(f)
    cm1 = np.array([b['frequency'] for b in data['phonon'][0]['band']])
    thz = cm1 / 33.35641
    return thz, cm1


# ---------------------------------------------------------------------------
# Step 3: NequIP training
# ---------------------------------------------------------------------------

def find_best_checkpoint():
    """Return (ckpt_path, run_dir) for latest best.ckpt, or (None, None)."""
    for run_dir in sorted(glob.glob(os.path.join(OUTPUTS_DIR, '*', '*')),
                          reverse=True):
        ckpt = os.path.join(run_dir, 'best.ckpt')
        if os.path.exists(ckpt):
            return ckpt, run_dir
    return None, None


def run_nequip_training(params):
    """Run nequip-train train.yaml in the deeph conda env."""
    t          = params.get('forces', {})
    conda_env  = t.get('conda_env', 'deeph')
    conda_base = params['paths']['conda_base']
    activate   = (f'source {conda_base}/etc/profile.d/conda.sh'
                  f' && conda activate {conda_env}')
    print(f'  Running nequip-train (conda: {conda_env}) ...')
    # nequip-train uses Hydra: pass --config-path (dir) and --config-name (no .yaml)
    subprocess.run(
        ['bash', '-c',
         f'{activate} && nequip-train --config-path {SCRIPT_DIR} --config-name train'],
        cwd=SCRIPT_DIR, check=True)


# ---------------------------------------------------------------------------
# Step 4: Compile model
# ---------------------------------------------------------------------------

def compile_model(params):
    """Compile best.ckpt → diamond_ase.nequip.pt2 via nequip-compile."""
    if os.path.exists(PT2_PATH):
        print(f'  Compiled model exists: {PT2_PATH}')
        return

    ckpt, _ = find_best_checkpoint()
    if ckpt is None:
        print('  ERROR: No checkpoint found; training may not have completed.')
        return

    t          = params.get('forces', {})
    conda_env  = t.get('conda_env', 'deeph')
    conda_base = params['paths']['conda_base']
    activate   = (f'source {conda_base}/etc/profile.d/conda.sh'
                  f' && conda activate {conda_env}')

    device = t.get('device', 'cuda')
    cmd = (f'{activate} && nequip-compile --mode torchscript'
           f' --device {device} --target ase {ckpt} {PT2_PATH}')
    print(f'  Compiling {os.path.basename(ckpt)} → diamond_ase.nequip.pt2 ...')
    subprocess.run(['bash', '-c', cmd], cwd=SCRIPT_DIR, check=True)
    print(f'  Compiled: {PT2_PATH}')


# ---------------------------------------------------------------------------
# Step 5: Phonopy + NequIP at Gamma
# ---------------------------------------------------------------------------

def run_phonopy_ml(params, model_path=None, freqs_json=None):
    """Compute ML Gamma phonon frequencies using phonopy + NequIP calculator.

    Runs in the deeph conda env via a subprocess launcher.
    Returns freqs_thz as a (nbands,) numpy array.
    """
    model_path = model_path or PT2_PATH
    freqs_json = freqs_json or os.path.join(SCRIPT_DIR, '_phonopy_freqs.json')

    t          = params.get('forces', {})
    conda_env  = t.get('conda_env', 'deeph')
    conda_base = params['paths']['conda_base']
    device     = t.get('device', 'cuda')

    # Use the original scf.in (ibrav=0, CELL_PARAMETERS angstrom) — the bands
    # pw.in uses 'alat' units which newer ASE versions cannot parse.
    uc_pw_in = os.path.abspath(os.path.join(DATA_DIR, '..', 'scf.in'))

    launcher = os.path.join(SCRIPT_DIR, '_phonopy_launcher.py')
    with open(launcher, 'w') as f:
        f.write(f"""import json, numpy as np
import phonopy
from phonopy.structure.atoms import PhonopyAtoms
from nequip.ase import NequIPCalculator
from ase import Atoms
from ase.io import read as ase_read

uc = ase_read('{uc_pw_in}', format='espresso-in')
calc = NequIPCalculator.from_compiled_model(
    '{model_path}', device='{device}', chemical_species_to_atom_type_map=True)

sc_matrix = [[4, 0, 0], [0, 4, 0], [0, 0, 4]]
ph = phonopy.Phonopy(
    unitcell=PhonopyAtoms(
        symbols=uc.get_chemical_symbols(),
        cell=uc.get_cell(),
        scaled_positions=uc.get_scaled_positions(),
    ),
    supercell_matrix=sc_matrix,
    primitive_matrix='auto',
)
ph.generate_displacements(distance=0.01)
supercells = ph.supercells_with_displacements
print(f'  Phonopy: {{len(supercells)}} displacements ({{len(supercells[0])}} atoms each)')

forces_list = []
for i, sc in enumerate(supercells):
    a = Atoms(symbols=sc.get_chemical_symbols(),
              positions=sc.get_positions(),
              cell=sc.get_cell(), pbc=True)
    a.calc = calc
    f = a.get_forces()
    forces_list.append(f)
    if (i + 1) % 5 == 0 or i == 0:
        print(f'  [{{i+1}}/{{len(supercells)}}] |F|max={{np.abs(f).max():.4f}} eV/A')

ph.forces = forces_list
ph.produce_force_constants()
ph.run_qpoints([[0, 0, 0]], with_eigenvectors=False)
freqs_thz = ph.get_qpoints_dict()['frequencies'][0].tolist()

with open('{freqs_json}', 'w') as fp:
    json.dump({{'freqs_thz': freqs_thz}}, fp)
print(f'  Gamma freqs written to {freqs_json}')
""")

    activate = (f'source {conda_base}/etc/profile.d/conda.sh'
                f' && conda activate {conda_env}')
    print(f'  Running phonopy + NequIP in {conda_env} env ...')
    subprocess.run(['bash', '-c', f'{activate} && python {launcher}'],
                   cwd=SCRIPT_DIR, check=True)

    with open(freqs_json) as fp:
        data = json.load(fp)
    return np.array(data['freqs_thz'])


# ---------------------------------------------------------------------------
# Step 6: Compare and plot
# ---------------------------------------------------------------------------

def print_phonon_table(dfpt_thz, dfpt_cm1, ml_thz):
    """Print mode-by-mode comparison table."""
    cm1_per_thz = 33.35641
    ml_cm1 = ml_thz * cm1_per_thz
    n = max(len(dfpt_thz), len(ml_thz))
    print(f"{'Mode':>5} {'DFPT (THz)':>11} {'DFPT (cm-1)':>12} "
          f"{'ML (THz)':>10} {'ML (cm-1)':>11} {'Err (cm-1)':>12}")
    print('-' * 65)
    for i in range(n):
        d_thz = dfpt_thz[i] if i < len(dfpt_thz) else 0.0
        d_cm1 = dfpt_cm1[i] if i < len(dfpt_cm1) else 0.0
        m_thz = ml_thz[i]   if i < len(ml_thz)   else 0.0
        m_cm1 = ml_cm1[i]   if i < len(ml_cm1)   else 0.0
        print(f'{i+1:>5} {d_thz:>11.4f} {d_cm1:>12.2f} '
              f'{m_thz:>10.4f} {m_cm1:>11.2f} {m_cm1-d_cm1:>+12.2f}')
    opt_d = dfpt_thz[dfpt_thz > 1.0]
    opt_m = ml_thz[ml_thz > 1.0]
    if len(opt_d) and len(opt_m):
        err_pct = (np.mean(opt_m) - np.mean(opt_d)) / np.mean(opt_d) * 100
        print(f'\nOptical: DFPT={np.mean(opt_d):.2f} THz, '
              f'ML={np.mean(opt_m):.2f} THz, err={err_pct:+.1f}%')


def plot_phonon_comparison(dfpt_thz, ml_thz, outpath):
    """Bar chart of DFPT vs ML Gamma phonon frequencies."""
    cm1_per_thz = 33.35641
    dfpt_cm1 = dfpt_thz * cm1_per_thz
    ml_cm1   = ml_thz   * cm1_per_thz
    n = max(len(dfpt_cm1), len(ml_cm1))
    modes = np.arange(1, n + 1)
    d = np.pad(dfpt_cm1, (0, n - len(dfpt_cm1)))
    m = np.pad(ml_cm1,   (0, n - len(ml_cm1)))

    fig, ax = plt.subplots(figsize=(6, 4))
    w = 0.35
    ax.bar(modes - w/2, d, w, label='DFPT (ph.x)', color='steelblue')
    ax.bar(modes + w/2, m, w, label='ML (NequIP)',  color='tomato')
    ax.set_xlabel('Mode')
    ax.set_ylabel('Frequency (cm$^{-1}$)')
    ax.set_xticks(modes)
    ax.set_title('Diamond $\\Gamma$-point phonons: DFPT vs ML (NequIP)')
    ax.legend()
    fig.tight_layout()
    fig.savefig(outpath, dpi=150)
    plt.close(fig)
    print(f'  Saved: {outpath}')


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------

def main():
    skip_training = '--skip-training' in sys.argv
    use_ref_model = '--ref-model' in sys.argv
    params_path = next((a for a in sys.argv[1:] if not a.startswith('--')), None)
    params = load_params(params_path)

    # Diagnostic: run phonopy with reference model and compare to reference QE phonopy+QE data
    if use_ref_model:
        print('=== DIAGNOSTIC: reference NequIP model ===')
        freqs_json = os.path.join(SCRIPT_DIR, '_phonopy_freqs_ref.json')
        if os.path.exists(freqs_json):
            print(f'  Using cached: {freqs_json}')
            with open(freqs_json) as fp:
                ml_freqs_thz = np.array(json.load(fp)['freqs_thz'])
        else:
            print('  Running phonopy with reference model...')
            ml_freqs_thz = run_phonopy_ml(params, model_path=PT2_REF_PATH,
                                          freqs_json=freqs_json)
        qe_thz, qe_cm1 = parse_phonopy_qpoints_yaml()
        print('\nGamma phonons — reference model vs reference QE (phonopy+QE):')
        print_phonon_table(qe_thz, qe_cm1, ml_freqs_thz)
        outpath = os.path.join(SCRIPT_DIR, 'phonon_comparison_ref.png')
        plot_phonon_comparison(qe_thz, ml_freqs_thz, outpath)
        print('\ntrain_force.py --ref-model done.')
        return

    print('Step 1: Building forces dataset...')
    build_forces_dataset(params)

    print('\nStep 2: Running DFPT at q=Gamma...')
    run_dfpt(params)

    print('\nStep 3: NequIP training...')
    ckpt, _ = find_best_checkpoint()
    if skip_training and ckpt:
        print(f'  --skip-training: using existing checkpoint: {ckpt}')
    elif ckpt and os.path.exists(PT2_PATH):
        print(f'  Using existing checkpoint: {ckpt}')
    else:
        run_nequip_training(params)
        ckpt, _ = find_best_checkpoint()
        if ckpt is None:
            print('ERROR: Training did not produce a checkpoint.')
            sys.exit(1)

    print('\nStep 4: Compiling model...')
    compile_model(params)

    print('\nStep 5: Running phonopy + NequIP at Gamma...')
    ml_freqs_thz = run_phonopy_ml(params)

    print('\nStep 6: Comparing phonons...')
    dfpt_thz, dfpt_cm1 = parse_dfpt_gamma_freqs()
    print_phonon_table(dfpt_thz, dfpt_cm1, ml_freqs_thz)

    outpath = os.path.join(SCRIPT_DIR, 'phonon_comparison.png')
    plot_phonon_comparison(dfpt_thz, ml_freqs_thz, outpath)

    print('\ntrain_force.py done.')


if __name__ == '__main__':
    main()