Discovery of High-Voltage Magnesium-Ion Cathodes using Machine Learning and First-Principles Calculations

Abstract

Developing high-performance cathode materials for magnesium-ion batteries (MIBs) remains challenging because Mg$^{2+}$ ions move slowly, and conventional materials exhibit low voltage outputs. In this study, machine learning and first-principles calculations were combined to investigate topological quantum materials (TQMs) as a new class of cathode candidates. A modified crystal graph convolutional neural network (mCGCNN) was used to screen 917 Mg-containing TQMs, identifying a small subset of materials with predicted voltages above 3 V and high volumetric capacities. Among these, Mg$_2$VO$_4$ and Mg$_6$MnO$_8$ were selected for detailed density functional theory (DFT) analysis. Formation energy and convex-hull calculations indicate that Mg$_x$VO$_4$ exhibits a fully stable magnesiation pathway, whereas Mg$_x$MnO$_8$ demonstrates minor metastability at intermediate compositions. The calculated voltage profiles yield average voltages of 3.66 V for Mg$_2$VO$_4$ and 4.06 V for Mg$_6$MnO$_8$, in good agreement with machine learning predictions. Electronic structure analysis, supported by Wannier interpolation, confirms that both materials are semiconducting, with valence bands dominated by O $2p$ states and conduction bands by transition-metal $d$ states, indicating a charge-transfer redox mechanism. Compared to conventional Mg cathodes, these TQMs exhibit higher voltages and competitive capacities, underscoring their potential for next-generation multivalent batteries. This study demonstrates that integrating machine learning with first-principles calculations offers an efficient approach for discovering and understanding novel cathode materials.