Classification of cast iron alloys through convolutional neural networks applied on optical microscopy images

Abstract

Classification of cast iron alloys based on graphite morphology plays a crucial role in materials science and engineering. Traditionally, this classification has relied on visual analysis, a method that is not only time-consuming but also suffers from subjectivity, leading to inconsistencies. This study introduces a novel approach utilizing convolutional neural networks - MobileNet for image classification and U-Net for semantic segmentation - to automate the classification process of cast iron alloys. A significant challenge in this domain is the limited availability of diverse and comprehensive datasets necessary for training effective machine learning models. This is addressed by generating a synthetic dataset, creating a rich collection of 2400 pure and 1500 mixed images based on the ISO 945-1:2019 standard. This ensures a robust training process, enhancing the model's ability to generalize across various morphologies of graphite particles. The findings showcase a remarkable accuracy in classifying cast iron alloys (achieving an overall accuracy of 98.9±0.4% and exceeding 97% for all six classes - for classification of pure images and ranging between 84% and 93% for semantic segmentation of mixed images) and also demonstrate the model's ability to consistently identify and graphite morphology with a level of precision and speed unattainable through manual methods.