In recent publications, our group has presented simulation studies that investigated the applicability of five ultraviolet inspection wavelengths for defect detection using a signal-to-noise ratio (SNR) based defect metric. All computations were performed using our in-house finite-difference time-domain (FDTD) Maxwell’s equations solver. Initial simulations contained no perturbations except for the defect. In the absence of noise, the defects are readily apparent, and we therefore added Poisson noise both to the defect and no-defect simulated images that yield the differential image. For these noisy differential images, signal and noise were separated to calculate the SNR with two additional variables: an area threshold A_min and a noise threshold σ. A wavelength dependent A_min (λ) that scaled linearly with wavelength was required to enable comparison across all five wavelengths. All types of defects were best resolved using an inspection wavelength of λ=47 nm assuming ideal conditions (i.e. no variation in numerical aperture or source intensity as a function of wavelength.)
In this work, we add line edge roughness as “wafer noise” on both defect and no-defect geometries to test this SNR approach in a more realistic way. Various permutations of defect types, polarizations, illumination angles, focal planes and wavelengths affect the differential images produced differently. For every combination of the above mentioned experimental setups we individually maximize the SNR as a function of A_min and σ to improve defect detection, and trends across wavelengths are assessed.
Finally, we present an alternative approach to the defect detection that is based on a neural network. We address questions of feature selection and its relation to the previous used SNR approach.