forgottenl5/esc-fault-benchmark

ESC-Fault: a simulated benchmark for electrostatic chuck fault diagnosis

Simulated sensor data from electrostatic chucks (ESCs), the part of a semiconductor etch or deposition tool that holds the wafer with static electricity and cools it with backside helium. Two releases:

• v1: 140 chucks, about 125,000 wafers, 7 classes (one fault per chuck), 14 signals at 1 Hz.
• v2 (new, folder v2/): 552 chucks, about 490,000 wafers; new arcing fault; two faults at once on some chucks (multi-label); intermittent leaks and slow onsets; a recipe that appears mid-life; and a test_shift split from a second tool profile.

Code, simulator, task definitions, baselines and scoring: https://github.com/forgottenl5/esc-fault-benchmark

Classes

healthy, contamination, wear, he_leak, residual_charge, heater_zone, sensor_drift (a decoy: the sensor is wrong, the chuck is fine), and in v2 arcing. Normal events that are not faults (plasma cleans, seasoning, sensor recalibrations, coolant drift, recipe changes, slow normal wear) appear in every chuck.

Files

Folder	Contents
wafers/, v2/wafers/	One row per wafer: recipe, counters, release (dechuck) time, summary features. All splits.
timeseries/	v1 raw 1 Hz signals, all splits.
v2/timeseries/	v2 raw 1 Hz signals for test and test_shift. Rebuild train/val signals with generate2.py --raw.
labels/, v2/labels/	Per-wafer labels, severity and remaining life, plus per-chuck metadata. Train and validation only.
tasks/, v2/tasks/	Task files. Test files have no answers.

v1 splits by chuck: 98 train, 14 validation, 28 test. v2: 360 train, 48 validation, 96 test, 48 test_shift. Test answers are held back.

Tasks

1. Diagnose each (chuck, end wafer) window. v1: 7 classes. v2: which of 7 faults are active (multi-label). Macro-F1.
2. Detect early: one alarm per chuck with the fault type. Detection rate, median delay, false alarms.
3. Remaining life for chucks that wear out. Mean absolute error in wafers.

All tasks are causal: use only wafers up to the end wafer.

Baselines

v2 (gradient-boosted trees):

	val	test (tool A)	test_shift (tool B)
Task 1 macro-F1 over 7 faults	0.902	0.872	0.847
Task 2 detection rate / median delay (wafers)	95% / 77	94% / 85	93% / 83
Task 3 MAE in wafers (guess-the-mean)	60.2 (117.7)	55.8 (125.8)	69.7 (115.9)

In a data-size study (5 to 40 training chucks per class), v2 test macro-F1 rises from 0.80 to 0.87 and the second tool costs about 0.03 at every size; sensor_drift stays hardest (test F1 0.20 to 0.44). Which sensor drifts matters most: a drifting zone-temperature sensor is caught only 11% of the time on test (breakdown by sensor in the GitHub README).

Raw signals barely help: adding the shape of each 90-second wafer step (30 blocks x 11 signals) to the summary features changes v2 macro-F1 by +0.000 / +0.003 / +0.005 (val / test / test_shift), and zone-temperature drift is still missed (recall 0.06 on test, 0.00 on test_shift). Script: baseline_raw_profiles.py on GitHub. Finding drifting zone-temperature sensors is the open problem.

v1 (test set):

Model	Task 1 macro-F1	Task 2 detection	Median delay (wafers)	Task 3 MAE (wafers)
Threshold rules	0.667	87.5%	152	n/a
Random forest	0.812	83.3%	73	65.5
Small neural net	0.782	83.3%	74.5	n/a
Random forest + trend and consistency features	0.861	not run	not run	not run

Load it

from datasets import load_dataset

repo = "forgottenl5/esc-fault-benchmark"
wafers = load_dataset(repo, "v2_wafers", split="train").to_pandas()   # v1: "wafers"
labels = load_dataset(repo, "v2_labels", split="train").to_pandas()   # v1: "labels"
df = wafers.merge(labels, on=["chuck_id", "wafer_idx"])
print(df.label.value_counts().head(10))

Configs: wafers, labels, timeseries (v1) and v2_wafers, v2_labels, v2_timeseries (v2). The raw signals are large; stream them with streaming=True. A ready-to-run v1 starter notebook is in the GitHub repo: https://github.com/forgottenl5/esc-fault-benchmark/blob/main/notebooks/esc-fault-starter.ipynb

v2 starter notebook (loads the v2/ files, trains one classifier per fault, scores it, shows which drifting sensors are missed, writes test predictions): https://github.com/forgottenl5/esc-fault-benchmark/blob/main/notebooks/esc-fault-v2-starter.ipynb

Limits

Simulated from a simplified physics model. Some constants come from public patents and papers; the rest are assumptions, listed in the dataset card on GitHub. Good scores here do not prove a method works on a real tool. Corrections from people who work with ESCs are welcome.

Citation

@misc{escfault2026,
  title  = {ESC-Fault: a simulated benchmark for electrostatic chuck fault diagnosis},
  author = {forgottenl5},
  year   = {2026},
  url    = {https://github.com/forgottenl5/esc-fault-benchmark},
  doi    = {10.5281/zenodo.24185993}
}

DOI: 10.5281/zenodo.24185993 (Zenodo archive of the code and simulator).

License

CC BY 4.0. Author: forgottenl5.