A PoDFA Benchmarking Study Between Manual and AI-supervised Machine Learning Methods to Evaluate Inclusions in Wrought and Foundry Aluminum Alloys

Abstract

The PoDFA inclusion measurement is achieved by identifying the inclusions and their concentration in the melt for each type with a trained operator. The standard technique is realized by using a square grid with an optical microscope to count the total area with each detected square. This manual and non-efficient method requires a lot of time and effort and can generate important variations in PoDFA results for reproducibility and repeatability. In the past, there were many unsuccessful attempts to automatically detect, count, and classify all inclusion types due to the complexity of the application. Disc sampling, image artifacts, polishing defects, and metallurgical constituents are some examples that can interfere with the inclusion detection and the measurement methodology. Commercial image analysis systems with threshold options and Boolean logical operations are not sufficient to automate the solution. The implementation of artificial intelligence technologies such as supervised machine learning algorithms are necessary to automate this complex method. The benchmarking study was achieved between the standard PoDFA methodology compared to the artificial intelligent way. Results show that the new technique exhibits a good correlation and a high potential for industrial use.