COFA: counterfactual attention framework for trustworthy wafer map failure classification

Classifying wafer map failure pattern plays a crucial role in semiconductor manufacturing, as it can help identify the underlying cause of abnormalities, thus reducing production costs. Existing works have shown that deep learning methods have great advantages in recognizing failure patterns. However, recent studies mainly focus on utilizing attention mechanisms to pinpoint critical regions as salient features, while ignoring the imperceptible underlying features and the causal relationship between prediction results and attention. This paper introduces a model-agnostic classification framework that leverages counterfactual explanations to enhance attention. Our approach consists of two steps: counterfactual example generation (Explain) and attention-based classifier refinement (Reinforce). The counterfactual explainer is designed to identify key pixel-level features, the adjustment of which could lead to different predictions. These generated counterfactual examples reveal hidden causal factors in the classifier’s decision-making process. Then the classifier utilizes these pixel features as attention, conducting reliable classification under the guidance of counterfactual examples. Through extensive experiments on real-world datasets, we demonstrate the effectiveness of our proposed model. It achieves an accuracy of 98.125% in the defect classification task on the WM-811K dataset and 92.544% on the MixedWM38 dataset, outperforming state-of-the-art attention methods such as SENet, CBAM, and Vision Transformer by over 5%. Our results highlight the superiority of our approach and its potential for practical implementation in the semiconductor manufacturing domain.