WaferVista

WaferVista contains 30,332 samples from 16 wafer-pattern classes, each with four numeric views on a 64 x 64 grid. It is derived from WM-811K. The additional multi-bin, profile and defect views are synthesized representations.

Paper: WaferVista: A Benchmark Multi-View Dataset for Wafer Map Analysis Constructed using Multi-Stage Visual Synthesis (under submission).

Download and Load

Download WaferVista.pkl from Files and versions. This release uses a pandas pickle; load it with pandas rather than datasets.load_dataset().

pip install numpy==2.3.5 pandas==2.3.3

import pandas as pd

df = pd.read_pickle("WaferVista.pkl")
print(df.shape)  # (30332, 7)
sample = df.iloc[0]
print(sample[["sample_id", "split", "failureType"]])
print(sample["multiBinMap"])  # numpy.ndarray, uint8, (64, 64)
train = df.loc[df["split"] == "train"]

Fields and Encoding

Each row represents one sample. df.attrs is empty; all encoding information is documented here.

Column	Type	Meaning
sample_id	string	Sample identifier.
split	string	train (21,232), val (4,550), or test (4,550).
failureType	string	Defect-pattern class name.
waferMap	uint8 [64, 64]	0 = outside wafer, 1 = pass, 2 = fail.
multiBinMap	uint8 [64, 64]	Categorical test-result bin values; 0 = outside, 127 = pass.
profileMap	int16 [64, 64]	Unitless relative values representing a continuous physical measurement: 0-255 are valid values; -1 denotes background.
defectMap	float32 [64, 64]	D64: a compact relative defect-density representation of the 6400 x 6400 source defect map.

The four views correspond to complementary information used in semiconductor production:

• Single-bin map (waferMap) records the pass/fail test result of each die, showing the spatial distribution of functional and failed dies across a wafer.
• Multi-bin map (multiBinMap) distinguishes die-level test-result categories, providing finer detail about different failure categories and their spatial distributions.
• Profile map (profileMap) describes spatial variation in a continuous physical quantity measured across a wafer. In this benchmark, the synthesized profile represents that variation using unitless relative values, supporting analysis of spatial uniformity and local deviations.
• Defect map (defectMap) describes the locations and spatial concentration of defects detected during wafer inspection. The 6400 x 6400 source maps are too large for a compact release, so they are summarized into the 64 x 64 D64 representation to retain coarse defect-density patterns at a practical size. D64 values provide an image-derived relative density representation.

bin_to_rgb below is one example palette provided for visualization. The tuples are in RGB order; you may replace the colors to suit your visualization without changing the stored bin values or their meaning. Colors are display choices, while bin values remain categorical test-result codes.

bin_to_rgb = {
    0: (127, 127, 127), 127: (255, 248, 248),
    3: (140, 180, 210), 8: (128, 0, 128), 9: (116, 162, 216),
    10: (42, 42, 165), 11: (0, 128, 0), 12: (0, 0, 255),
    15: (255, 0, 255), 18: (0, 255, 255), 19: (92, 92, 205),
    20: (230, 216, 173), 26: (0, 128, 128), 28: (0, 140, 255),
    30: (127, 255, 0), 31: (221, 160, 221), 32: (50, 205, 154),
    33: (255, 255, 0), 35: (255, 0, 0), 41: (0, 218, 184),
    42: (112, 25, 25), 43: (112, 81, 20), 44: (192, 80, 165),
    48: (222, 30, 30), 52: (149, 209, 255), 55: (30, 144, 30),
    60: (70, 113, 171), 61: (222, 179, 138), 62: (128, 116, 180),
    64: (153, 48, 112), 65: (111, 215, 232), 88: (230, 245, 250),
    90: (158, 79, 121), 99: (79, 158, 121),
}

class_to_label = {
    "ArrayBlock": 0, "Center": 1, "Checkerboard": 2, "Cruciform": 3,
    "Donut": 4, "DripStriping": 5, "EdgeLoc": 6, "EdgeRing": 7,
    "HorizontalBanding": 8, "HorizontalStriping": 9, "Loc": 10,
    "Nearfull": 11, "Random": 12, "Scratch": 13, "ThickEdgeRing": 14,
    "VerticalStriping": 15,
}

Use profile == -1 to identify background; 0 is a valid relative value. The following example restores the original JET colors and gray background. Install opencv-python and matplotlib if needed; df is the DataFrame loaded above.

import cv2
import numpy as np
import matplotlib.pyplot as plt

profile = df.iloc[0]["profileMap"]
background = profile == -1
bgr = cv2.applyColorMap(np.clip(profile, 0, 255).astype(np.uint8), cv2.COLORMAP_JET)
rgb = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
rgb[background] = (128, 128, 128)
plt.imshow(rgb, interpolation="nearest")
plt.axis("off")
plt.show()

For numerical analysis, exclude background with profile[profile >= 0].

Acknowledgements and Reference

We acknowledge M.-J. Wu, J.-S. R. Jang, and J.-L. Chen for their work on WM-811K, the source dataset used to construct WaferVista. Please cite the original WM-811K paper when using this dataset:

M.-J. Wu, J.-S. R. Jang, and J.-L. Chen, "Wafer Map Failure Pattern Recognition and Similarity Ranking for Large-Scale Data Sets," IEEE Transactions on Semiconductor Manufacturing, vol. 28, no. 1, pp. 1-12, 2015, doi: 10.1109/TSM.2014.2364237.

License

WaferVista is released under Creative Commons Attribution 4.0 International (CC BY 4.0). Provide appropriate attribution, link to the license, and indicate any changes. See LICENSE for the full terms. This license covers our contributions to WaferVista; the original WM-811K dataset retains its own licensing terms.