Session: 12-06-01: Scientific Machine Learning (SciML) for Characterization, Modeling, and Design of Structures and Materials

Paper Number: 140771

140771 - 3d Microstructure and Finite Element Modeling of the Mechanical Behavior of Cast Irons Using Deep Learning for Metallographic Analysis 

The microstructure of a material represents its internal structure, containing crucial information from which its physical and chemical properties are defined. Microstructure analysis is traditionally carried out exclusively by experts due to the difficulty in determining its characteristics, such as phases, pores, cracks, precipitates and other complex structures. This makes the process subjective and difficult to automate. However, recent advances in artificial intelligence (AI) and deep learning methods (DLM) have offered promising solutions. The purpose of this work is to develop an image analysis process using computer vision and deep learning to identify features in a reference material system, Nodular Cast Iron (NCI) and Gray Cast Iron (GCI) microstructures. 

This study significantly contributes to the field of metallographic analysis by reducing reliance on subjective human evaluation. The analyses become more impartial and at the same time, realistic, leading to more efficient use of time and resources. To achieve this, we proceeded to metallographically prepare and collect optical microscope images for both NCI and GCI. Subsequently, these images were subjected to pixel segmentation using a fully convolutional neural network accompanied by a max-voting scheme that allowed us to define the final microstructure images. 

Along the investigation, several methods for deep learning were employed as pre-trained model segmentation labels, such as Segment Anything Meta (SAM), U-Net Model and unsupervised learning skills like clustering. Additionally, the analysis employed various traditional computer vision techniques to preprocess the input data for the neural networks. Masks were used that had a size of 1024x1280 pixels, which is the size returned by the optical microscope. However, cropping was also necessary to use smaller masks, thus increasing the number of images for the most appropriate training of the neural network. Once the system was trained, the reconstruction of the cropped images was performed, and the most accurate method was selected. Our system achieved 90% microstructure identification accuracy when comparing the masks used for training against the ones obtained from the deep learning system.  

The automated analysis for the images is then applied to microstructures generated at different depths obtained by mild polishing. The depth is controlled/measured by a fiduciary mark (hardness indent). This procedure allows the reconstruction in three-dimensions (3D) of the microstructures. 

In the 3D-virtual microstructures, mechanical simulations using the Finite Element Method (FEM) are performed using software ABAQUS and ANSYS for comparison. DREAM3D software was used previously to obtain a mesh file, identify the two distinct phases of cast iron: α-iron and graphite, and assign its respective mechanical properties. The simulations consist of reproducing spherical indentations at various loads to analyze the elastic-plastic deformations and compare its results to actual experimental results. The consistency between the modeling and experimental results underscores the success of the method used. 

Presenting Author: Marco León USFQ

Presenting Author Biography: Marco Francisco León Dunia is a Mechanical Engineer graduated from the Escuela Politécnica Nacional (Ecuador). He also holds a MSc in Mechanical Engineering Sciences with a specialisation in Corrosion. He has studies at Institut für Oberflächentechnik, RWTH Aachen University (Germany). He has worked for the Institute of Surface Engineering (IOT - Germany) and several companies in the oil industry. He is a research professor in the Department of Mechanical Engineering. His main areas of research focus on the study and characterisation of materials for industrial and biomedical applications, modelling and development of new coatings, failure analysis, and the application, development and refinement of methods for monitoring and combating corrosion.

Authors:

Carlos Jarrin USFQ
Krutskaya Yépez USFQ
Sebastián Insuasti USFQ
Francisco Martinez USFQ
Marco León USFQ
Lorena Bejarano USFQ
Alfredo Valarezo USFQ