On the use of physics in machine learning for manufacturing process inspection

Abstract

We discuss the use of machine learning in computational imaging for manufacturing process inspection and control. In a recent article we described a physics-enhanced auto-correlation based estimator (Peace) for quantitative speckle. We derived an explicit forward relationship between the Particle Size Distribution (PSD) and the speckle autocorrelation for particle sizes significantly larger than the wavelength (x100 to approximately x1,000). We subsequently trained a machine learning kernel to invert the autocorrelation and obtain the PSD, using the explicit forward model to reduce the number of experimentally acquired examples. In this talk, we present an expanded discussion of Peace and its properties, including spatial and temporal sampling and accuracy, and more general applications.