Source-linked AI summary
Online Open-World Wafer Defect Discovery using Hyperdimensional Computing
Quanling Zhao, Ava Emami, Tsuyoshi Idé, Rajiv Joshi, Steven Holmes, Michael Passow, Tajana Rosing
TL;DR
Wafer defect analysis lacks a way to discover and track recurring unknown signatures while production continues. WaferHDC uses evolving hyperdimensional memories to screen normal wafers and group recurring anomalies without labels or a predefined defect count. It improves ARI by 13.3 percentage points and NMI by 10.5 percentage points over existing statistical-detector and streaming-clustering baselines.
Problem
Existing approaches detect anomalies or group defects offline, leaving recurring unknown signatures untracked while production is ongoing.
Source: §2 · p. 2Method
WaferHDC uses hyperdimensional memory to model normal wafer patterns and incrementally discover recurring defect modes without labels or a predefined class count.
Source: §1 · pp. 1–2Results
13.3 percentage points higher ARI and 10.5 percentage points higher NMI than the strongest existing statistical detectors and streaming-clustering baselines.
Source: §1 · p. 2Takeaways & Limitations
On WM-811K, WaferHDC supports continuous monitoring by providing an expandable memory framework for emerging wafer failures rather than only recognizing predefined defects.
Source: §5 · p. 7Online operation assumes a reference set of only non-defective wafers, with no labeled defect examples or advance knowledge of the number of defect modes.
Source: §3 · p. 3
Abstract
from arXiv · showhide
Conventional wafer defect analysis relies on offline, closed-world classifiers that fail to detect novel, unseen defect signatures emerging during continuous semiconductor manufacturing. To address this, we propose WaferHDC, an online open-world wafer defect discovery framework based on hyperdimensional computing (HDC). The memory-centric architecture of WaferHDC enables rapid anomaly detection and grouping of recurring failures into compact hypervector representations. This allows the memory space to dynamically scale with new defect modes without retraining a static classifier or archiving entire raw wafer history. WaferHDC first models normal spatial distributions, then dynamically groups recurring failures into persistent defect memory without labels or predefined class counts. It outperforms the strongest online baseline by 13.3 p.p. in adjusted Rand index and 10.5 p.p. in normalized mutual information on the industry-standard WM-811K dataset. These results demonstrate the effectiveness of WaferHDC for continuous and open-world wafer defect analysis.
1 Introduction
WaferHDC addresses online discovery of recurring, previously unseen wafer defects by combining normal-pattern screening with evolving defect-mode memories. It improves clustering alignment over existing online approaches while avoiding defect labels and predefined class counts.
1 Introduction
- WaferHDC unifies normal-wafer screening and discovery of recurring defect modes through hyperdimensional computing.It rejects wafers inconsistent with normal patterns, accumulates similar anomalies, and promotes recurring spatial signatures into persistent memories.Source: §1 · p. 1
- The online setting addresses previously unseen defects that emerge as wafers arrive and gradually form recurring modes.Figure 1 illustrates unlabeled abnormal-wafer detection and organization of recurring patterns; displayed labels are illustrative.Source: Figure 1 · p. 2
- Its evolving memory can add defect modes without labels, a predefined class count, full wafer-history storage, or classifier retraining.WaferHDC represents wafers as binary hypervectors compared with Hamming distance and incrementally combined in associative memory.Source: §1 · p. 2
- 13.3 percentage points higher ARI and 10.5 percentage points higher NMI than the strongest existing statistical detectors and streaming-clustering baselines.WaferHDC forms compact memories of recurring, semantically coherent defect patterns rather than only local microclusters.Source: §1 · p. 2
2 Related Work
Prior approaches address classification, anomaly detection, or offline grouping, but the cited work identifies a gap in tracking recurring unknown failure signatures during production. WaferHDC processes wafers online and consolidates repeated spatial signatures into persistent defect-mode memories.
2 Related Work
- Closed-set classifiers assume predefined categories and labeled examples, while anomaly detectors flag patterns outside modeled normal behavior.Neither step alone organizes multiple abnormal wafers into coherent groups that may share a recurring failure mechanism.Source: §2 · p. 2
- Most unsupervised wafer-analysis methods operate offline; prior methods either detect unseen patterns or discover groups offline, not track emerging signatures during production.Table 1 defines open-world handling, online causal processing, and mode discovery as distinct capabilities.Source: §2 · p. 2, Table 1 · p. 2
- WaferHDC processes wafers as they arrive and consolidates repeated spatial signatures into persistent defect-mode memories.The paper states that recurring failures can then be recognized, tracked, and prioritized for timely root-cause analysis and corrective action.Source: §2 · p. 2
3 WaferHDC Design
WaferHDC processes wafers online through a shared hypervector representation, first screening for normal behavior and then accumulating recurring abnormalities into persistent defect modes. It operates without defect labels or a predefined number of defect classes.
3 WaferHDC Design
- WaferHDC screens each incoming wafer for normal behavior, then accumulates recurring abnormalities into persistent defect memories without labels or a predefined number of modes.Unmatched patterns remain candidates until recurrent, coherent, and clearly abnormal enough for promotion.Source: §3 · p. 3, §3.4 · p. 5
3.1 Distance-sensitive binary wafer encoding
- Thermometer-encoded projected features form binary hypervectors whose Hamming distances retain graded differences between wafer patterns.This representation supports efficient comparisons and incremental updates without storing every previously observed wafer.Source: §3.1 · p. 3
3.2 Stage 1: modeling normal production wafer behavior
- Stage 1 models normal variation with multiple Hamming-space prototypes and checks both nearest-mode resemblance and support from nearby observed normal wafers.The two distance signals are normalized and combined; wafers exceeding the calibrated threshold proceed to Stage 2.Source: §3.2 · p. 4
3.3 Memory for recurring abnormal patterns
- Temporary candidate memories accumulate recurring abnormal patterns using fixed-size summaries rather than retaining the complete wafer history.Bit counts provide a majority prototype and measure within-candidate agreement; the same compact representation is used for persistent modes.Source: §3.3 · pp. 4–5
3.4 Recurrence-driven defect discovery
- A candidate becomes a persistent defect mode only after sufficient recurrence, spatial consistency, and abnormality; later matching wafers update that mode directly.Other unfamiliar wafers continue to form or update temporary candidates as production proceeds.Source: §3.4 · p. 5
4 Experimental Evaluation
On WM-811K, the evaluation examines WaferHDC’s screening and online discovery stages separately and together. In the full pipeline, its variants offer different coverage–quality operating points while improving clustering alignment over the strongest online baseline.
4.1 Experimental Setup
- The WM-811K protocol treats the none class as normal and the remaining eight classes as unknown defect modes, with labels and mode counts unavailable during processing.All methods receive the same phased stream; at most 5,000 non-defect wafers are used for training, and threshold selection does not use defect labels.Source: §4.1 · pp. 5–6
4.3 Full-Pipeline Open-World Discovery
- 13.3 and 10.5 percentage points higher ARI and NMI than the strongest online baseline, respectively, WaferHDC-Pure reaches 56.77% ARI and 69.38% NMI in the full pipeline.WaferHDC-Bal. assigns 58.66% of unknown wafers while exceeding baselines in NMI; WaferHDC-Agg. assigns 76.42% and retains 50.11% ARI, higher than the strongest baseline.Source: Figure 5 · p. 6, §4.3 · p. 6
4.2 Stage-Wise Evaluation
- 93.48% of defect wafers are rejected by Stage 1, with 97.37% AUROC; its distance-sensitive hypervectors and Hamming-space memory are reused for Stage 2 discovery.The screening performance is competitive with the strongest classical anomaly detectors across rejection rate, AUROC, and AUPRC.Source: Figure 3 · p. 5, §4.2 · p. 6
- 58.07% ARI and 70.68% NMI for WaferHDC-Pure exceed online clustering baselines’ approximately 22–23% ARI in the Stage 2-only evaluation.WaferHDC-Bal. raises assignment to 64.20% while retaining 60.31% NMI; WaferHDC-Agg. assigns more defect wafers than all three baselines while achieving substantially higher ARI.Source: Figure 4 · p. 5, §4.2 · p. 6
4.3 Full-Pipeline Open-World Discovery
- Normal-wafer contamination is 0.14% for WaferHDC-Pure, 0.66% for WaferHDC-Bal., and 3.24% for WaferHDC-Agg.The first two rates are substantially lower than the baselines; the Aggressive variant emphasizes broader defect coverage.Source: §4.3 · pp. 6–7
- WaferHDC processes each wafer in approximately 0.35–0.81 ms, corresponding to about 1,200–2,800 wafers per second.The authors report this speed as sufficient for online deployment given industrial probing and testing cycles.Source: §4.3 · p. 7
5 Conclusion
WaferHDC unifies normal screening and incremental defect-mode discovery in an evolving HDC memory design. On WM-811K, it improves full-pipeline ARI and NMI over the strongest baseline.
5 Conclusion
- On WM-811K, WaferHDC improves full-pipeline ARI and NMI by 13.3 and 10.5 percentage points over the strongest baseline.The conclusion presents HDC as an expandable memory framework for continuously learning emerging wafer failures rather than recognizing only predefined defects.Source: §5 · p. 7