AI-driven multiscale modeling of mechanical and corrosion properties in biodegradable Mg-Lu alloys

Magnesium-lutetium (Mg-Lu) alloys exhibit significant potential as biodegradable bone implant materials. However, their low alloying content necessitates the use of excessively large simulation supercells, resulting in high computational costs. Moreover, conventional empirical interatomic potentials face substantial challenges, thereby limiting the efficiency and accuracy of materials design. To address these issues, within the AI for Science paradigm, we developed a deep neural network-based machine learning potential (DNN-MLP) model to systematically investigate the crystallographic characteristics, mechanical properties, energetic properties, and dynamic corrosion behavior of Mg-Lu alloys. The results indicate that the constructed DNN-MLP model achieves a root-mean-square error (RMSE) of 2.57 × 10^-3 eV per atom for energy predictions and below 2.39 × 10^-3 eV Å^-1 for force predictions. The lattice constants predicted by the DNN-MLP model show excellent agreement with experimental data and density functional theory (DFT) calculations. Meanwhile, the DNN-MLP model accurately captures the mechanical properties of Mg-Lu alloys and outperforms traditional empirical potentials, such as the Gupta potential; furthermore, the strengthening mechanism of Mg-Lu alloys is elucidated based on stacking fault energy analysis. In addition, the DNN-MLP model successfully predicts the corrosion current density and corrosion potential of Mg-Lu alloys. Overall, the DNN-MLP model offers high accuracy and reliability, enabling efficient multiscale simulation and optimized design of biomedical bone implant materials.