File size: 5,055 Bytes
bbe9b3e | 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 | """Create compact paper-ready figures for the incomplete-physics closure."""
from __future__ import annotations
import json
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import torch
from src import t2_graybox_discrete_energy as graybox
from src.generate_t2_graybox_closure import MATERIAL
from src.train_t2_graybox_discrete_energy import load_cohort
from src.train_t2_multiaxial_models import RecurrentStress
ROOT = Path(__file__).resolve().parents[1]
DATA = ROOT / "data" / "t2_graybox_closure_v1" / "cohort.h5"
MODELS = ROOT / "models" / "t2_graybox_closure_v1"
ARTIFACTS = ROOT / "artifacts" / "t2_graybox_closure_v1"
def main() -> None:
cohort = load_cohort(DATA)
denrm_checkpoint = torch.load(MODELS / "denrm.pt", map_location="cpu", weights_only=False)
denrm = graybox.NeuralHardeningLaw(channels=2)
denrm.load_state_dict(denrm_checkpoint["state_dict"])
denrm.eval()
gru_checkpoint = torch.load(MODELS / "gru.pt", map_location="cpu", weights_only=False)
gru = RecurrentStress(6, cell="gru", hidden=72)
gru.load_state_dict(gru_checkpoint["state_dict"])
gru.eval()
norm = gru_checkpoint["normalization"]
with torch.no_grad():
denrm_stress = graybox.rollout(
cohort.strain,
torch.full((len(cohort.strain),), float(MATERIAL["young_pa"])),
torch.full((len(cohort.strain),), float(MATERIAL["poisson"])),
torch.full((len(cohort.strain),), float(MATERIAL["yield_stress_pa"])),
denrm,
bisection_iterations=24,
)["stress"]
gru_stress = (
gru((cohort.strain - norm["strain_mean"]) / norm["strain_std"])
* norm["stress_std"]
+ norm["stress_mean"]
)
metrics = json.loads((ARTIFACTS / "model_metrics.json").read_text())
figure, axes = plt.subplots(2, 2, figsize=(11.0, 7.8), constrained_layout=True)
colors = {"truth": "#1f2937", "denrm": "#e11d48", "gru": "#2563eb"}
component_labels = ("xx", "yy", "zz", "xy", "yz", "xz")
for axis, family in zip(
axes[0], ("out_of_phase_lissajous", "random_direction_blocks"), strict=True
):
index = next(i for i, value in enumerate(cohort.families) if value == family)
strain = cohort.strain[index].numpy()
component = int(np.argmax(np.ptp(strain, axis=0)))
x = 100.0 * strain[:, component]
axis.plot(
x,
cohort.stress[index, :, component].numpy() / 1.0e6,
color=colors["truth"],
linewidth=2.1,
label="AgentFEM reference",
)
axis.plot(
x,
denrm_stress[index, :, component].numpy() / 1.0e6,
"--",
color=colors["denrm"],
linewidth=1.8,
label="DENIM (2-state closure)",
)
axis.plot(
x,
gru_stress[index, :, component].numpy() / 1.0e6,
color=colors["gru"],
linewidth=1.1,
alpha=0.85,
label="GRU",
)
axis.set_title(family.replace("_", " "))
axis.set_xlabel(f"strain {component_labels[component]} (%)")
axis.set_ylabel(f"stress {component_labels[component]} (MPa)")
axis.grid(alpha=0.22)
axes[0, 0].legend(frameon=False, fontsize=8)
names = ("Incomplete J2", "GRU", "DENIM")
keys = ("incomplete_j2", "gru", "denrm")
rmse = [metrics["models"][key]["test"]["rmse_mpa"] for key in keys]
bars = axes[1, 0].bar(
names, rmse, color=("#9ca3af", colors["gru"], colors["denrm"])
)
axes[1, 0].bar_label(bars, fmt="%.2f")
axes[1, 0].set_ylabel("held-out path RMSE (MPa)")
axes[1, 0].set_title("unseen loading-path accuracy")
axes[1, 0].grid(axis="y", alpha=0.22)
peeq = np.linspace(0.0, 0.10, 250)
with torch.no_grad():
learned = denrm.isotropic(
torch.tensor(peeq, dtype=torch.float32),
torch.full((len(peeq),), float(MATERIAL["yield_stress_pa"])),
).numpy()
reference = np.interp(
peeq,
np.asarray(MATERIAL["hardening_peeq"]),
np.asarray(MATERIAL["hardening_stress_pa"]),
) - float(MATERIAL["yield_stress_pa"])
axes[1, 1].plot(peeq, reference / 1.0e6, color=colors["truth"], linewidth=2.1, label="hidden table")
axes[1, 1].plot(peeq, learned / 1.0e6, "--", color=colors["denrm"], linewidth=1.8, label="learned monotone law")
axes[1, 1].set_xlabel("equivalent plastic strain")
axes[1, 1].set_ylabel("isotropic hardening radius (MPa)")
axes[1, 1].set_title("unknown hardening-law recovery")
axes[1, 1].grid(alpha=0.22)
axes[1, 1].legend(frameon=False, fontsize=8)
figure.suptitle(
"Incomplete-physics closure: 3-state tabulated reference → 2-state DENIM",
fontsize=13,
)
ARTIFACTS.mkdir(parents=True, exist_ok=True)
figure.savefig(ARTIFACTS / "closure_summary.png", dpi=220)
plt.close(figure)
if __name__ == "__main__":
main()
|