Machine-Learning-Assisted Multi-Element Optimization of Mechanical Properties in Spinel Refractory Materials

Using machine learning models, this study innovatively introduces multi-element compositions to optimize the performance of spinel refractories. A total of 1120 spinel samples were fabricated at 1600 °C for 2 h, and an experimental database containing 112 data points was constructed. High-throughput performance predictions and experimental verifications were conducted, identifying the sample with the highest hardness, (Al₂Fe_0.25Zn_0.25Mg_0.25Mn_0.25)O₄ (1770.6 ± 79.1 HV1, 3.35 times that of MgAl₂O₄), and the highest flexural strength, (Al₂Cr_0.5Zn_0.1Mg_0.2Mn_0.2)O₄ (161.2 ± 9.7 MPa, 1.4 times that of MgAl₂O₄). Further analysis of phase composition and microstructure shows that the mechanism of hardness enhancement is mainly the solid solution strengthening of multi-element doping, the energy dissipation of the large-grain layered structure, and the reinforcement of the zigzag grain boundary. In addition to solid solution strengthening and a compact low-pore structure, the mechanism of improving bending strength also includes second-phase strengthening and phase concentration gradient distribution. This method provides a promising way to optimize the performance of refractory materials.