The detrimental effects of Cu contamination during steel recycling and production are primarily due to the segregation of Cu at Fe grain boundaries (GBs). A promising approach to mitigate these effects is the introduction of alloying elements that inhibit Cu segregation at Fe GBs, which can be investigated through atomistic simulations. Currently, popular simulation methods include density functional theory (DFT)-based simulations and empirical interatomic potentials (EIPs)-based simulations. DFT calculations can provide simulations with high accuracy, while the high computational cost limits their application to simpler GB structures. Moreover, simulations utilizing EIPs may lack reliable potentials, especially when predicting GB segregation tendencies across a broader range of alloy systems. To address these challenges, universal machine learning interatomic potentials (uMLIPs), which are trained on DFT data and applicable for most of elements, have emerged as a promising alternative. Although uMLIPs have shown potential in various materials simulation tasks, their reliability for out-of-distribution tasks, such as simulating GB segregation behavior, remains unproven.

This thesis project evaluates the performance of the best available uMLIPs, specifically MACE-MP-0, CHGNet, M3GNet, and SevenNet-0, in predicting single-solute GB segregation energies, GB energies for both body-centered cubic (BCC) Fe and face-centered cubic (FCC) Fe systems, and solution enthalpies for BCC Fe and cementite. The results were compared against existing studies conducted via DFT calculations to assess the accuracy and applicability of each uMLIP. Additionally, some EIPs were tested for comparison, serving as an extra reference. The findings reveal that MACE-MP-0 generally outperforms the other uMLIPs in both accuracy and stability of convergence. While all tested uMLIPs perform well in BCC Fe systems, CHGNet(v0.2.0) and SevenNet-0 show reduced accuracy in FCC Fe simulations. Although most of the simulations using uMLIPs converged well in BCC Fe GBs, many unconverged cases were reported in FCC Fe systems, particularly for uMLIPs other than MACE-MP-0 and CHGNet(v0.3.0). Furthermore, a consistent underprediction of segregation tendencies for highly segregating solute elements, such as Cu, is observed in the results of MACE-MP-0 and CHGNet. This suggests that while uMLIPs hold significant potential for atomistic simulations, fine-tuning pre-trained uMLIP models for out-of-distribution tasks, such as calculating GB segregation energy, is recommended to improve accuracy and convergence behavior. This work offers a valuable benchmark for using uMLIPs in future GB segregation studies.