Adaptive Transfer Learning With Deformable Convolution for Wafer Defect Pattern Recognition

The convolutional neural network (CNN) based model has been widely used for wafer map defect pattern recognition. However, it often fails to detect patterns effectively after a few months of deployment due to shifts in data distribution, usually linked to aging inspection tools or recipe changes. To address this issue, researchers have turned to transfer learning techniques. However, most methods require labeled data for the target dataset, which is time-consuming and labor-intensive. Unsupervised domain adaptation (UDA), a transfer learning technique, requires no human labeling and performs better than conventional methods. This study introduces deformable convolutional encoder adversarial discriminator domain adaptation (DC-ADDA), which adapts the deformable convolutional network (DC-Net) as an encoder in the adversarial discriminative domain adaptation method. Using MixedWM38 and synthesized multi-bin wafer bin maps (SWBMs) as source datasets with WM-811K as the target, DC-ADDA outperforms other methods across transfer learning, encoders, and domain adaptation scenarios. The proposed DC-ADDA method achieves an accuracy of 66.46%, a precision of 68.13%, a recall of 69.63%, and an F1-score of 68.41% when applied from MixedWM38 to WM-811K. When applied from SWBMs to WM-811K, it achieves an accuracy of 75.61%, a precision of 75.20%, a recall of 74.81%, and an F1-score of 74.81%.