File size: 7,313 Bytes
ce33d98 74e2a2e ce33d98 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 | """Per reviewed MP row: compare geometry with each candidate task and choose the energy source.
Writes decisions.parquet (one row per reviewed LeMat row) and replacements.jsonl (the new
per-row payload for every row whose DFT record is replaced).
"""
import sys, json, glob, numpy as np, pandas as pd
from pathlib import Path
from scipy.optimize import linear_sum_assignment
W = Path(sys.argv[1])
T_OK = 0.2 # eV/atom: |DFT - MLIP median| bound for "MLIP-consistent" (99.5th pct on clean rows)
REF_GAIN = 0.05 # eV/atom: a live row switches only if the candidate is this much closer to the MLIPs
SAME, NEAR = 1e-3, 0.05 # A, lattice and site deviation
f = pd.read_parquet(W / "identify.parquet")
f = f[f.review != ""].set_index("immutable_id")
cand = pd.read_parquet(W / "candidates.parquet")
cand_by = {k: g for k, g in cand.groupby("immutable_id")}
ext = {}
for fn in glob.glob(str(W / "s3/*.jsonl")):
for line in open(fn):
d = json.loads(line)
ext[d["task_id"]] = d
L = pd.read_parquet(W / "lemat_review_rows.parquet",
columns=["immutable_id", "lattice_vectors", "cartesian_site_positions",
"species_at_sites"]).drop_duplicates("immutable_id").set_index("immutable_id")
def compare(lat, sp, xyz, d):
"""Max lattice deviation, permutation (task sites -> LeMat order), max site deviation."""
tl = np.asarray(d["lattice"], float); tsp = np.asarray(d["species"]); txyz = np.asarray(d["xyz"], float)
sp = np.asarray(sp)
if len(tsp) != len(sp) or sorted(tsp) != sorted(sp):
return np.inf, None, np.inf
dlat = float(np.abs(tl - lat).max())
if dlat > NEAR:
return dlat, None, np.inf
inv = np.linalg.inv(lat); perm = np.empty(len(sp), int)
for el in np.unique(sp):
i = np.flatnonzero(sp == el); j = np.flatnonzero(tsp == el)
fr = (txyz[j][None, :, :] - xyz[i][:, None, :]) @ inv; fr -= np.round(fr)
rr, cc = linear_sum_assignment(np.linalg.norm(fr @ lat, axis=2)); perm[i[rr]] = j[cc]
fr = (txyz[perm] - xyz) @ inv; fr -= np.round(fr)
return dlat, perm, float(np.linalg.norm(fr @ lat, axis=1).max())
records, payloads = [], []
for mid, r in f.iterrows():
lr = L.loc[mid]
lat = np.stack(lr.lattice_vectors); xyz = np.stack(lr.cartesian_site_positions); sp = list(lr.species_at_sites)
mlip = r.epa - r.d_med if pd.notna(r.d_med) else np.nan
a = abs(r.d_med) if pd.notna(r.d_med) else np.nan
cs = []
for c in (cand_by.get(mid).itertuples() if mid in cand_by else []):
d = ext.get(c.task_id)
if d is None or "error" in d or d.get("energy") is None:
continue
dlat, perm, dpos = compare(lat, sp, xyz, d)
geo = "same" if max(dlat, dpos) <= SAME else "near" if max(dlat, dpos) <= NEAR else "different"
e = d["energy"] / len(d["species"])
cs.append(dict(task_id=c.task_id, dep=c.task_deprecated, is_lemat=c.is_lemat_task, is_mp=c.is_mp_task,
calc_type=c.calc_type, geo=geo, dlat=dlat, dpos=dpos, epa=e,
b=abs(e - mlip) if geo != "different" else np.nan, perm=perm, doc=d))
live = [c for c in cs if not c["dep"] and not (c["is_lemat"] and r.identified)
and (not isinstance(r.mp_run, str) or c["calc_type"].split(" ")[0] == r.mp_run)]
live_geo = [c for c in live if c["geo"] != "different"]
mp_c = next((c for c in live if c["is_mp"]), None)
choice, action = None, None
if r.review in ("control_same", "control_live"):
action = "control"
choice = next((c for c in cs if c["is_lemat"]), None)
elif r.review == "stale":
if mp_c is not None and mp_c["geo"] != "different" and (np.isnan(a) or mp_c["b"] <= max(T_OK, a + REF_GAIN)):
choice, action = mp_c, "replace_energy"
elif live_geo:
best = min(live_geo, key=lambda c: (c["b"], c["calc_type"].endswith("Optimization")))
if np.isnan(a) or best["b"] <= max(T_OK, a + REF_GAIN):
choice, action = best, "replace_energy"
if choice is None and mp_c is not None and mp_c["geo"] == "different":
# MLIPs were evaluated on LeMat's geometry; the re-relaxed one is close (median dV 2e-4),
# so the same referee applies with the MLIP energy at LeMat's geometry as the proxy.
b_s = abs(mp_c["epa"] - mlip) if pd.notna(mlip) else np.nan
if np.isnan(a) or np.isnan(b_s) or b_s <= max(T_OK, a + REF_GAIN):
choice, action = mp_c, "replace_structure"
if choice is None:
action = "keep_stale_mlip_consistent" if (pd.notna(a) and a <= T_OK and r.material_deprecated is False) else "remove"
else: # live_big
if live_geo and pd.notna(a) and a > T_OK:
best = min(live_geo, key=lambda c: c["b"])
if best["b"] < a - REF_GAIN and best["b"] <= T_OK:
choice, action = best, "replace_energy"
if action is None and mp_c is not None and mp_c["geo"] == "different" and pd.notna(a) and a > T_OK:
b_s = abs(mp_c["epa"] - mlip)
if b_s < a - REF_GAIN and b_s <= T_OK:
choice, action = mp_c, "replace_structure"
if action is None:
action = "keep"
rec = dict(immutable_id=mid, review=r.review, grp=r.grp, action=action, n_cand=len(cs), n_live_geo=len(live_geo),
old_epa=r.epa, mlip_epa=mlip, old_vs_mlip=r.d_med, lemat_task=r.lemat_task, lemat_calc_type=r.lemat_calc_type,
lemat_batch=r.lemat_batch, lemat_task_deprecated=r.lemat_task_deprecated, mp_task=r.mp_task,
mp_calc_type=r.mp_calc_type, material_deprecated=r.material_deprecated)
if choice is not None:
d = choice["doc"]
rec.update(new_task=choice["task_id"], new_calc_type=choice["calc_type"], new_epa=choice["epa"],
new_vs_mlip=choice["epa"] - mlip, geo=choice["geo"], dlat=choice["dlat"], dpos=choice["dpos"],
new_is_mp_task=choice["is_mp"])
if action in ("replace_energy", "replace_structure", "control"):
perm = choice["perm"] if choice["perm"] is not None else np.arange(len(d["species"]))
take = lambda v: None if v is None else [v[i] for i in perm]
payloads.append(dict(immutable_id=mid, action=action, task_id=choice["task_id"], energy=d["energy"],
perm=[int(i) for i in perm],
lattice=d["lattice"], species=take(d["species"]), xyz=take(d["xyz"]),
forces=take(d["forces"]) if d.get("forces") else None, stress=d.get("stress"),
magmom=take(d["magmom"]), total_magnetization=d.get("total_magnetization"),
efermi=d.get("efermi"), last_updated=d.get("last_updated"),
potcars=d.get("potcars"), hubbards=d.get("hubbards")))
records.append(rec)
out = pd.DataFrame(records)
out.to_parquet(W / "decisions.parquet")
with open(W / "replacements.jsonl", "w") as fh:
for p in payloads:
fh.write(json.dumps(p) + "\n")
pd.set_option("display.width", 220)
print(pd.crosstab([out.review, out.grp], out.action, margins=True))
print(pd.crosstab(out.action, out.geo.fillna("-")))
|