Challenge 1: Amentum

Amentum: Machine Learning Interatomic Potentials for Defect Kinetics in Radiation-Tolerant FeCr(Al) Structural Alloys

Amentum is a global engineering and technology company with deep expertise in nuclear energy, providing support across the full lifecycle of nuclear facilities, from design and construction through operation, life extension, and decommissioning.

A central challenge in nuclear structural materials qualification is predicting how alloys degrade under sustained neutron irradiation over operational timescales. Iron-chromium-aluminium (FeCrAl) alloys with body-centred cubic (BCC) structure are among the leading candidates for next-generation nuclear applications, particularly as accident-tolerant fuel cladding for light water reactors. Here, they offer promising high-temperature steam oxidation resistance compared to conventional Zr-based systems, and the broader Fe-Cr binary system is also important as a structural material candidate, with well-established radiation damage phenomenology that underpins FeCrAl behaviour.

Long-term performance under irradiation is governed by a hierarchy of defect processes, and further mechanistic understanding of these is needed to help narrow the range of candidates for further experimental testing. Displacement cascades produce surviving Frenkel pairs whose subsequent clustering and evolution determine microstructural change. In FeCrAl, Cr reduces interstitial mobility through the formation of stable Fe-Cr dumbbells, while Al has little influence on surviving defect numbers but restrains cluster formation. At longer timescales, dislocation loop formation and coarsening - with both ½⟨111⟩ and ⟨100⟩ loop types present - contribute to irradiation hardening and embrittlement. Grain boundaries act as sinks for mobile defects and as sites for radiation-induced segregation, directly affecting strength and corrosion resistance. The ultimate engineering concern is creep and corrosion resistance over operational lifetimes, which depend sensitively on the rates of all these underlying processes.

The challenge is to explore how MACE-based machine-learning interatomic potentials (MLIPs) can be developed, benchmarked, and applied to defect kinetics in BCC FeCr(Al) alloys, such as:

Vacancy and interstitial diffusivity: Can a MACE potential accurately reproduce migration barriers and diffusion coefficients across the Fe-Cr-Al composition space, and how does it compare against existing empirical potentials and available experimental and DFT data?
Grain boundary energetics: Can MACE reliably predict grain boundary energies and defect–boundary interactions relevant to segregation and sink strength modelling?
Composition dependence: How do defect kinetics vary across the relevant Cr and Al concentration range, and can a MACE model trained on a manageable DFT dataset generalise reliably across this space?
Uncertainty quantification: Where are MACE predictions reliable, and how should uncertainty estimates inform decisions about where additional DFT reference data is needed?
Stretch goal: Dislocation formation and migration: Can MACE capture the formation energies and migration mechanisms of dislocation including the competition between ½⟨111⟩ and ⟨100⟩ loop types?

Approaches

Participants will be encouraged to use MACE as the primary modelling framework, with existing empirical Fe-Cr-Al potentials serving as baselines for validation. Approaches of interest include evaluating existing MACE foundation models; considering active learning or committee-based strategies for efficient training data construction; and deployment of trained potentials within LAMMPS for defect migration and microstructural evolution studies.

Amentum will provide scientific context, guidance on the alloy composition range of interest, and clarity on the target defect properties, with the open literature and publicly available DFT databases forming the primary source of reference data for the group.

Extra background

Atomistic modelling of defect kinetics has traditionally relied on classical empirical potentials such as the embedded atom method (EAM) fitted to experimental data and DFT calculations. While these potentials can reproduce point-defect properties in Fe-Cr and Fe-Al subsystems, existing models struggle to simultaneously capture the mixing enthalpy, dumbbell binding energies, and stacking fault energies across the full Fe–Cr–Al composition space. Machine learning interatomic potentials (MLIPs) offer a route beyond these limitations and can achieve transferability across properties relevant to collision cascades and plasticity simultaneously, for example accurately reproducing short-range repulsive interactions, stacking fault energies, dislocation core structures, and defect cluster formation energies within a single model. For FeCr alloys specifically, MLIPs have demonstrated near-DFT precision in capturing atomic-scale energetics whilst maintaining the computational efficiency needed for large-scale molecular dynamics. However, validated MLIPs for the full ternary FeCrAl system, with explicit treatment of defect kinetics across the relevant composition range, do not yet exist.

The MACE architecture, an equivariant, higher-order message-passing neural network framework, is of particular interest for this challenge. MACE has demonstrated state-of-the-art accuracy across a range of materials systems and is deployable within LAMMPS for production-scale molecular dynamics, making it a practical candidate for integration into existing nuclear materials simulation workflows. Foundation models such as MACE-MP-0 and successors provide a natural starting point, and the study group setting offers an opportunity to benchmark foundation models for Fe-Cr-Al for defect kinetics problem and consider future active-learning and fine-tuning approaches.