Prediction of fracture behavior of Al2O3-Cr2O3 ceramics in different Cr2O3 ratios under flexure load using machine learning methods

Machine learning (ML) algorithms have recently shown considerable success in the field of materials science, particularly for modeling novel materials, analyzing enormous volumes of data, and making material property predictions using such data. In this study, the best-suited machine learning algorithm was investigated to predict the fracture behavior of Al2O3-Cr2O3 ceramics in different Cr2O3 volume ratios (0.5, 1, 5) under bending. To improve the performance of the model, the hyperparameters were optimized and MLP (Multi-layer Perceptron), RF (Random Forest), and XGBoost (Extreme Gradient Boosting) algorithms were used. The leave-one-out technique was employed as a cross-validation method to make sure the consistency of the results. The heat map technique was used to choose the appropriate input and output parameters from the dataset. The effect of the input factors was also studied using the experimental method of surface response. The RF model provided the closest predictions to the experimental values for most of the samples. The most important input variable was found to be the Cr2O3 ratio for the relative density, diameter for the fracture strength, and thickness for the total crack length according to the feature importance results from the RF algorithm. The optimal solution found by the GA (Genetic Algorithm) is 0.7 for Cr2O3% concentration, with a corresponding diameter of 28.5 mm and a thickness of 2.2 mm with a fracture strength of 325.8 MPa.