Prof. Marco F. Leon
,
Institute for Energy and Materials Research, Universidad San Francisco de Quito, Quito, Ecuador
Prof. Miryan Lorena Bejarano
,
Institute for Energy and Materials Research, Universidad San Francisco de Quito, Quito, Ecuador
Mr. Carlos Jarrin
,
Institute for Energy and Materials Research, Universidad San Francisco de Quito, Quito, Ecuador
Mr. Sebastian Isuasti
,
Institute for Energy and Materials Research, Universidad San Francisco de Quito, Quito, Ecuador
Mr. Westly Castro
,
Institute for Energy and Materials Research, Universidad San Francisco de Quito, Quito, Ecuador
Maria Gracia Velez
,
Institute for Energy and Materials Research, Universidad San Francisco de Quito, Quito, Ecuador
Jackson Alcivar
,
Institute for Energy and Materials Research, Universidad San Francisco de Quito, Quito, Ecuador
Ms. Krutskaya Irene Yepez
,
Alberta Next-Gen AM, University of Alberta, Edmonton, AB, Canada
Prof. Alfredo Valarezo
,
Institute for Energy and Materials Research, Universidad San Francisco de Quito, Quito, Ecuador
The microstructure of a material defines its physical and chemical properties, making its characterization crucial for material analysis. Traditionally, microstructure assessment relies on expert evaluation, which is inherently subjective and difficult to automate. Recent advances in artificial intelligence (AI) and deep learning have demonstrated promising capabilities in automating this process. This study presents a methodology for reconstructing 3D microstructures of nodular and gray cast iron from 2D metallographically characterized images using AI and advanced processing techniques. The first one presents a regular model shape of globular nodules, whereas the second presents a more exigent amorphous shape of graphite flakes.
The proposed methodology integrates experimental and computational techniques. Metallographic samples were prepared, and optical microscopy images were acquired for both cast irons. Image processing was carried out using deep learning methods: clustering model, Segment Anything Model (SAM), and Watershed model. Additionally, conventional computer vision techniques were applied to enhance data processing. Following segmentation, a 3D reconstruction of microstructures was performed using Python and MeshLab. The accuracy of these reconstructions was validated through 3D-finite element (FE) simulations of hardness indentation using ABAQUS.
The developed methodology achieved a 90% accuracy in microstructure identification, as evaluated using the intersection over union metric. Moreover, the 3D-FE simulations yielded a very strong correlation with experimentally obtained mechanical properties, demonstrating the effectiveness of the proposed method.
This study enhances metallographic analysis by reducing reliance on subjective human evaluation, improving reproducibility, and increasing efficiency. The results highlight the potential of deep learning-based microstructure characterization for accurate mechanical property prediction, paving the way for further advancements in AI-driven material analysis.
