WaferDiSC: Distribution-enhanced scoring and consistency modeling for unknown defect detection in wafer bin maps

Abstract Wafer bin maps (WBMs) provide important spatial information for semiconductor process monitoring and yield diagnosis. However, most existing wafer-map analysis methods are developed under a closedset assumption and may fail to detect previously unknown defect patterns. Diffusion-based reconstruction methods have shown promising performance for unknown-defect detection, but their iterative sampling process is computationally expensive. To address these issues, this paper proposes WaferDiSC, a distribution-enhanced scoring and consistency modeling framework for unknown defect detection in WBMs. From a measurement perspective, WaferDiSC provides a quantitative score for measuring the deviation of a test WBM from the known wafer-level defect distribution. WaferDiSC distills a featureguided flow-matching teacher into a dual-end consistency student, enabling four-step reconstruction with three Euler flow-matching updates and one final consistency jump. To improve detection robustness, four complementary anomaly responses are constructed from reconstruction discrepancy, feature-distribution deviation, trajectory inconsistency, and morphology-aware structural response, and are fused using a reliability-aware adaptive Softmax strategy. Experiments on 36 open-set scenarios constructed from WM-811K and MixedWM38 show that WaferDiSC achieves an overall AUROC of 0.9404, outperforming the strongest baseline by 6.99 percentage points. In addition, WaferDiSC reduces the inference time to 361.12 ms per WBM, achieving 4.28× and 4.36× speedups over Flow Matching and WigDM, respectively.