This study seeks to gain a thorough understanding of the unique features exhibited by aluminum inclusions in microscopic images. The ultimate goal is to leverage this knowledge to create a reliable machine learning-based classification system for accurately classifying these inclusions. Through a detailed methodology encompassing domain knowledge acquisition, preprocessing, feature selection, and modelling, the research investigates the performance of various classifiers, such as Support Vector Machine (SVM), Random Forest, and k-Nearest Neighbors (KNN), as well as ensemble techniques, specifically stacking. Initial results indicate suboptimal performance of traditional classifiers without domain-driven feature extraction, with SVM achieving an accuracy of 43%, Random Forest 63%, and KNN 29%. However, significant improvement in accuracy, precision, recall, and F1-score metrics was observed after employing preprocessing steps and leveraging domain knowledge. Specifically, SVM achieved 80% accuracy, KNN 75%, and Random Forest 78% accuracy after preprocessing. Further refinement through hyperparameter tuning and cross-validation led to SVM achieving 90% accuracy, KNN 82%, and Random Forest 90%. Finally, the stacking classifier, combining predictions from diverse base models with a logistic regression meta-learner, demonstrated superior performance with an accuracy of 93.4%. These findings underscore the critical role of domain knowledge in enhancing machine learning-based classification systems for microscopic image analysis, offering valuable insights applicable beyond aluminum manufacturing.