File size: 5,562 Bytes
239a1be | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 | """Evaluation metrics in physical units."""
from __future__ import annotations
from typing import Dict, List
import numpy as np
from .data import TARGETS, TARGET_KEYS
# Targets used for the headline macro average (see docs/data_card.md).
# The 6 dislocation-type breakdown targets are reported separately: several are
# annotated in only 3-6 of the 33 samples and have near-zero variance, which
# makes R2 numerically meaningless as a macro-average component.
PRIMARY = ("sigma", "xi_BCC", "xi_FCC", "xi_HCP", "xi_other", "rho_total", "twin")
def _r2(y, p):
if len(y) < 2:
return float("nan")
ss_res = float(np.sum((y - p) ** 2))
ss_tot = float(np.sum((y - y.mean()) ** 2))
if ss_tot < 1e-12:
return float("nan")
return 1.0 - ss_res / ss_tot
def regression_metrics(raw_true: np.ndarray, raw_pred: np.ndarray,
mask: np.ndarray) -> Dict[str, Dict[str, float]]:
"""raw_*: (N, Ft) physical units; mask: (N, Ft) 0/1."""
out = {}
for j, t in enumerate(TARGETS):
m = mask[:, j] > 0
n = int(m.sum())
if n == 0:
out[t.key] = {"n": 0, "r2": float("nan"), "mae": float("nan"),
"rmse": float("nan"), "unit": t.unit, "group": t.group}
continue
y, p = raw_true[m, j], raw_pred[m, j]
sd = float(np.std(y))
mae = float(np.mean(np.abs(y - p)))
r2_log = float("nan")
if t.group in ("dislocation", "twin") or t.transform == "asinh":
ly, lp = np.log1p(np.maximum(y, 0)), np.log1p(np.maximum(p, 0))
r2_log = _r2(ly, lp)
out[t.key] = {
"n": n,
"r2": _r2(y, p),
"r2_log": r2_log,
"mae": mae,
"nmae": mae / sd if sd > 1e-12 else float("nan"),
"rmse": float(np.sqrt(np.mean((y - p) ** 2))),
"bias": float(np.mean(p - y)),
"y_std": sd,
"unit": t.unit,
"group": t.group,
}
return out
def curve_metrics(strain: np.ndarray, sig_true: np.ndarray, sig_pred: np.ndarray,
eps_list=(0.02, 0.05, 0.10)) -> Dict[str, float]:
"""Per-sample macroscopic curve descriptors."""
out = {}
pk_t = float(np.max(sig_true))
pk_p = float(np.max(sig_pred))
out["peak_true"] = pk_t
out["peak_pred"] = pk_p
out["peak_rel_err"] = abs(pk_p - pk_t) / max(abs(pk_t), 1e-6)
for e in eps_list:
i = int(np.argmin(np.abs(strain - e)))
out[f"sig@{e:g}_true"] = float(sig_true[i])
out[f"sig@{e:g}_pred"] = float(sig_pred[i])
out[f"sig@{e:g}_rel_err"] = abs(sig_pred[i] - sig_true[i]) / max(abs(sig_true[i]), 1e-6)
out["curve_rmse"] = float(np.sqrt(np.mean((sig_pred - sig_true) ** 2)))
return out
def summarise(per_target: Dict[str, Dict[str, float]]) -> Dict[str, float]:
"""Macro-average R2 / MAE over the targets that actually have labels."""
r2s = [v["r2"] for v in per_target.values() if v["n"] > 0 and np.isfinite(v["r2"])]
r2l = [v["r2_log"] for v in per_target.values()
if v["n"] > 0 and np.isfinite(v.get("r2_log", float("nan")))]
nm = [v["nmae"] for v in per_target.values()
if v["n"] > 0 and np.isfinite(v.get("nmae", float("nan")))]
core = [v["r2"] for v in per_target.values()
if v["n"] >= 1500 and np.isfinite(v["r2"])]
prim = [per_target[k]["r2"] for k in PRIMARY
if k in per_target and per_target[k]["n"] > 0
and np.isfinite(per_target[k]["r2"])]
return {
"macro_r2": float(np.mean(r2s)) if r2s else float("nan"),
"macro_r2_primary": float(np.mean(prim)) if prim else float("nan"),
"n_primary": len(prim),
"core_r2": float(np.mean(core)) if core else float("nan"),
"median_r2": float(np.median(r2s)) if r2s else float("nan"),
"macro_r2_log": float(np.mean(r2l)) if r2l else float("nan"),
"macro_nmae": float(np.mean(nm)) if nm else float("nan"),
"n_targets_scored": len(r2s),
}
def sigma_decomposition(raw_true, raw_pred, mask, sid, j=0):
"""Split stress accuracy into an alloy-level term and a curve-shape term.
level_oracle_r2 : R2 of a predictor that knows each sample's *mean* stress
exactly but has no shape -> upper bound of a level-only model.
shape_r2 : R2 after removing each sample's own mean from both curves
-> how well the *shape* (hardening/softening) is captured.
"""
m = mask[:, :, j] > 0
y, p = raw_true[:, :, j], raw_pred[:, :, j]
yv, pv = y[m], p[m]
if len(yv) < 4:
return {}
r2 = _r2(yv, pv)
lv_t, lv_p = [], []
ys, ps = [], []
for i in range(y.shape[0]):
s = m[i]
if s.sum() < 4:
continue
lv_t.append(y[i, s].mean()); lv_p.append(p[i, s].mean())
ys.append(y[i, s] - y[i, s].mean()); ps.append(p[i, s] - p[i, s].mean())
lv_t, lv_p = np.array(lv_t), np.array(lv_p)
ys, ps = np.concatenate(ys), np.concatenate(ps)
return {
"r2": r2,
"level_oracle_r2": _r2(yv, np.concatenate([np.full(int(m[i].sum()), y[i, m[i]].mean())
for i in range(y.shape[0]) if m[i].sum() >= 4])),
"shape_r2": _r2(ys, ps),
"level_corr": float(np.corrcoef(lv_t, lv_p)[0, 1]) if len(lv_t) > 2 else float("nan"),
"level_mae_GPa": float(np.mean(np.abs(lv_t - lv_p))),
"level_bias_GPa": float(np.mean(lv_p - lv_t)),
"n_samples": len(lv_t),
}
|