Differentiable Modeling for Computational Imaging

This thesis explores advancements in computational imaging methods that enable lensless, phase-sensitive microscopy, wavefront sensing, and metrology applications. By using overlapping regions illuminated by a probe beam, interference pattern measurements allow for computational reconstruction of objects through iterative algorithms. The focus is on enhancing these reconstructions using differentiable modeling of the underlying physics. This approach allows for the joint optimization of various parameters and the integration of deep learning techniques. The goal is to increase the flexibility and robustness of computational imaging for various applications, particularly under challenging experimental conditions. The significance is particularly evident in the semiconductor industry, where nanoscale imaging is critical due to shrinking feature sizes in computer chips. With transistor dimensions approaching a few nanometers, traditional metrology faces significant challenges. Extreme Ultraviolet (EUV) light, with wavelengths around 13.5 nm, is used in semiconductor manufacturing but is absorbed or poorly refracted by most materials, making lenses impractical. Computational lensless imaging provides high-resolution, non-destructive imaging of nanoscale structures on semiconductor wafers, addressing a key metrology need.