Recycling steel at scale is hindered by tramp elements such as Cu and Sn, which degrade material properties. Atomistic simulations using foundational machine-learned interatomic potentials (MLIPs) trained on large databases, such as Materials Project, Alexandria, and OMAT, offer a promising approach to study the effects of these impurities. However, fine-tuning these models to specific systems can lead to catastrophic forgetting–the loss of general chemical knowledge acquired during pretraining. Here, we evaluate forgetting in three foundational MLIPs: CHGNet, SevenNet-O, and MACE, by fine-tuning on a data set of bcc-based structures, with Fe atoms only. When evaluated on a subset of the Materials Project data set with a learning rate of 0.0001, the fine-tuned MLIPs of CHGNet and SevenNet-O exhibited only a minor increase in RMSE of 0.047 and 0.022 eV/atom, respectively, indicating markedly minor forgetting. In contrast, fine-tuned MACE exhibited catastrophic forgetting, despite a range of additional strategies such as layer freezing and data set replay. We attribute the catastrophic forgetting to architectural sensitivity. These results highlight the importance of fine-tuning hyperparameters, model architecture, and data set design, with fine-tuned models of CHGnet and SevenNet-O showing some potential for efficient and transferable modeling of recycled steels.

Bearing steels are complex materials composed of an iron matrix and a well defined and precise amount of several alloying elements. In order to improve sustainability and circularity, there is a tendency to increase the utilisation of scrap material for their production. The variability of the composition of scrap material has a direct impact on the properties of the final steels: There is less control on their composition due to the possible presence of larger amounts of tramp and alloying elements. One way to study the effect of tramp elements is by using universal machine learning interatomic potentials. These types of potential render the investigation of multi-element systems possible. They permit the study of interactions between iron atoms in the matrix and multiple concurrent tramp and alloying elements, a feature that is currently not available in classical potentials. In this work, we present a benchmark of four state-of-the-art universal machine learning interatomic potentials (Crystal Hamiltonian Graph Neural Network (Deng et al 2023 Nat. Mach. Intell. 5 1031–41) (v0.2.0 and v0.3.0), Materials 3-body Graph Network (Chen and Ping Ong 2022 Nat. Comput. Sci. 2 718–28), Multiple Atomic Cluster Expansion (Batatia et al 2022 Advances in Neural Information Processing Systems vol 35 pp 11423–36)) and SevenNet (Park et al 2024 J. Chem. Theory Comput. 20 4857–68), and study their applicability to the simulation of systems relevant to steels. For pure Fe, all potentials accurately predict the equilibrium lattice parameter, but the accuracy varies for other properties. For most solute–solute and solute–vacancy interactions all interatomic potentials tend to capture the general trends though there is a disparity in the predicted magnitudes. While currently 'off-the-shelf' universal machine learning interatomic potentials fail to predict some key properties, some of them show significant potential to serve as starting point for further training and refinement.

Dark etching regions (DERs) are a widely studied phenomenon in the context of sub-surface microstructure decay in rolling bearings. These regions result from dislocation motion (plasticity) and carbon migration due to rolling contact fatigue (RCF). We conducted a systematic study using Mössbauer spectroscopy to identify phase evolution during tempering and DER formation. Our findings resolve a long-standing debate in the literature by demonstrating that fine temper carbides dissolve during DER formation. A slight decrease in non-stoichiometric carbides was observed in the DERs, indicating the dissolution of fine carbides. The released carbon is believed to back-diffuse into ferrite and/or martensite.