Automated Image Analysis of Metallurgical Grade Samples Reinforced with Machine Learning

Abstract

Controlling metal cleanliness in primary and secondary aluminum production is critical for ensuring quality in commercial sales and for effective process optimization. Solidified aluminum melt samples are today typically analyzed using established techniques such as LiMCA and PoDFA, however, these techniques rely on heavy and expensive equipment, extensive running times, and high heterogeneity of the results. The primary bottleneck of PoDFA analyses, the current standard approach, is the manual analysis of melt micrographs by human operators. In the present study, an image analysis platform based on a machine learning algorithm capable of quantifying contaminants in PoDFA micrographs was developed and tested. Machine learning models enable improved performance in heterogeneous datasets compared to common image analysis techniques using minimal computational resources and are envisioned to enable superior cost-scaling in metal cleanliness assessments. Future implementations will expand on the quantitative differentiation of relevant inclusion types.