Hidden Clues in WM-811K: Uncovering the Spatial Signals Behind Wafer Map Prediction
Year: 2026
Type: preprint
Abstract: WM-811K (also known as LSWMD) is a wafer map dataset first made publiclyavailable in a 2015 paper by Wu et al. It contains 811,457 real-world wafer mapscollected from TSMC fabs.When Wu et al. introduced WM-811K in 2015, they used classical machine learningtools like SVMs. Later, as CNNs became popular, researchers applied them to WM-811Kas well.The problem with CNNs and neural networks in general is that they suffer froma certain opacity, making it harder to understand what’s going on inside them unlikesome of the more explainable classifiers like decision trees.So while it’s possible to achieve high performance with CNNs, at the same time it’snot easy to tell which features of the dataset are driving that performance.This study attempts to answer this specifically for WM-811K: how much of thepredictive performance results from the wafer layout, the count of failed dies, theirdistance from the center and their exact 2-D location.After analyzing duplicate wafer maps, removing same wafer maps with conflictinglabels and ensuring that the same manufacturing lot or exact same map doesn’t appearin training, validation and test data, we were left with 25,513 defect-labeled wafer mapsfrom eight classes.We then constructed four variations of each wafer map. The Original was theactual wafer map present in the dataset. The Layout-only kept only the layout of thenon-background dies. The Global-shuffle kept the count of failed dies while shufflingthem around. The Radial-shuffle randomized the locations of failed dies only in theirradial bin, which was chosen out of 8 radial bins based on distance from the center.The data split, preprocessing and CNN architecture remained the same for all fourof them.Macro-F1 progressed like this: 0.3829 (Layout-only), 0.5803 (Global-shuffle), 0.7439(Radial-shuffle) and 0.8563 (Original). The three stepwise gains were +0.1974, +0.1636,and +0.1123.These results show that WM-811K contains useful information at several levels:wafer layout, number of failures, distribution of failed dies in a given radius, and exact2-D locations of the failed dies. Therefore, high classification performance on thisbenchmark does not come only from learning the exact defect configuration
Author: Tolani, Kabir
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236147059
