Deep Learning in Optical Metrology: How Can DYnet++ Enhance Single-Shot Deflectometry for Complex Surfaces?

Fuel cells are electrochemical devices that convert the chemical energy from a fuel and an oxidizing agent like Oxygen into electrical energy through a chemical reaction. They are considered a promising and environmentally friendly technology for generating electricity, particularly for powering vehicles, homes, and portable electronics.

However, micro defects on the surfaces of fuel cells can have various implications depending on their size, nature, and location. These defects can include imperfections, irregularities, or anomalies in the materials that make up the fuel cell components, such as the electrodes, electrolyte, and catalyst layers. Micro defects disrupt the smooth flow of ions and electrons within the fuel cell. As a consequence, the resistance of the cell is increased, and the overall efficiency and output power of the cell is reduced. 

The traditional method to detect these defects is through Scanning Electron Microscopy (SEM). It involves the information about the morphology and topography of the surface to identify the defects. The Korean Research Institute of Standards and Science researchers have developed a technology based on deep learning techniques that enables real-time 3D measurements using a single-sot pattern projection method.

Their method of single-shot deflectometer uses a high carrier frequency pattern. However, the visibility of the captured fringe pattern using these methods is not feasible when projecting this pattern onto a metal surface with low polishing quality, such as a battery fuel. Due to low reflectivity, the quality of the captured image could be better, and the phase cannot be retrieved correctly. Many surfaces with highly deformed levels generate complex reflected fringe patterns that include closed-loop and opened-loop features, demonstrating a low-frequency composite pattern from which phase retrieval is difficult.

To overcome this limitation, the team built an AI algorithm for the pattern projection method inspired by the technique of DL in optical meteorology. They used DYnet++, trained with measurement data on thousands of surface shapes. This allows DYnet++ to perform real-time 3D morphology measurements of surfaces with low reflectivity or complex shapes. They added more convolution layers to the Ynet model based on the Unet++ architecture to generate a DYnet++ model or nested Y-net. Basically, their proposed concept is a standard encoder and decoder block to help the network learn better from fringe patterns.

Obtaining a good training dataset is essential in every DL task to ensure the best result. Training data in deflectometry can be generated by simulation and experimentally. However, the simulation data will only partially reflect the actual physical imaging process. This will lead to a problem with very good results with the simulation data but no good experimental results. They designed a Deformable Mirror (DM) to obtain experimental training data quickly. It is a specialized optical device used in adaptive optics systems to correct for distortions and aberrations in the incoming light.

In conclusion, their proposed method’s strong and novel point is that even when the surface has low reflectivity and a very complex topology that might generate closed- and opened-loop fringe patterns together, their DL network can still measure them in seconds. The model could predict the results quickly and automatically without human intervention. This is extremely useful for speeding up the manufacturing process of these surfaces in modern industry.

Check out theย Paper.ย All Credit For This Research Goes To the Researchers on This Project. Also,ย donโ€™t forget to joinย our 30k+ ML SubReddit,ย 40k+ Facebook Community,ย Discord Channel,ย andย Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..

๐Ÿ Join the Fastest Growing AI Research Newsletter Read by Researchers from Google + NVIDIA + Meta + Stanford + MIT + Microsoft and many others...