Scaling reliable interatomic potentials to complex nuclear alloys via pretrained atomic models

Abstract Next-generation fission and fusion reactors impose unprecedented demands on structural materials, requiring simultaneous resistance to high temperatures, high-dose irradiation, and aggressive corrosion. Designing materials that harness the intrinsic properties of multiple elements and their synergistic interactions has emerged as a key strategy to achieve such integrated performance. To guide this design paradigm, a mechanistic understanding of chemically and structurally complex systems is essential. However, such understanding is currently constrained by the lack of high-fidelity interatomic potentials (IAPs) that enable predictive, large-scale atomistic simulations. Here, we employ, for the first time, a multi-task, physics-informed pretraining strategy with the large atomic model (LAM) to systematically evaluate the construction and predictive capability of IAPs for complex nuclear alloy systems. Using Ta-Nb-W-Mo-V as a representative case, the resulting DPA2-5E model—trained solely on the quinary dataset—significantly outperforms conventional machine learning IAPs, demonstrates superior transferability to lower-order subsystems, and accurately reproduces cascade damage and stress-strain behavior. Furthermore, this approach extends to nuclear-relevant structures and corrosive/oxide environments, enabling high-fidelity IAPs and large-scale simulations at reactor extremes.