Enhancing Optics-Based Detection of Semiconductor Patterning Defects Using Convolutional Neural Networks

Semiconductor patterning defects have technological and financial consequences in nanoelectronics fabrication.

Optical methods offer fast, non-destructive defect identification, but separating signals from measurement noise is daunting. An asymmetry exists between the impact of misclassifying a defect as pristine and vice versa. Data-driven machine learning is a key approach for the classification of patterning defect images, even if unresolved.

We have optimized a dataset of experimental images collected at 193 nm wavelength for enhanced processing using supervised neural networks. 30,000 images of scattering from bridges and line breaks had been collected using a scatterfield microscope using orthogonal linear polarizations aligned to an intentional defect array.

Large datasets are desirable to resolve overfitting in neural networks; data augmentation techniques and filtering are applied to images of pristine and defect patterning, increasing the training set size by about an order of magnitude. We compare the effects of wavelet compressions including the 2-D Haar wavelet transform on the training of the neural network. Initial results show the successful binary classification of the defect and no-defect samples for specific defect types, enabling future work.