Automated classification of microstructure data in additive manufacturing via low-dimensional learning

In metal additive manufacturing applications, attaining enhanced material properties relies on mastering and tailoring precise microstructural characteristics, while maintaining repeatability from build to build. As a result, microstructure analysis and characterization play a central role in product qualification and process verification procedures. The common industrial practice for analyzing microstructure data primarily relies on the subjective evaluation of human experts. This practice is time consuming, expensive, and inherently affected by subjective assessments. Such limitations have motivated increasing research efforts devoted to the development of data analytics and machine learning solutions for automated processing and classification of microstructural measurements. This study presents a novel approach to automatically extract comprehensive informative features from complex microstructural data gathered through electron backscattered diffraction (EBSD). The proposed method relies on the automated extraction of both morphological grain properties and crystal orientation distribution features by means of a unified low-dimensional learning approach specifically conceived to deal with the multi-dimensional nature of ESBD data. The comparison against benchmark approaches highlights the benefits of the proposed solution from a practical usage perspective. Alongside a simulation analysis, this work presents two real case studies involving the classification of microstructural properties in electron beam powder bed fusion and in directed energy deposition.