Predicting structures and energetics of metallic nanoparticles across wide size ranges remains challenging because the balance of interaction scales changes rapidly with system size. We introduce a multi-scale Mixture-of-Experts (MoE) architecture for Cu clusters and nanoparticles that explicitly separates short-, medium-, and long-range interactions using three specialised machine-learning interatomic potential experts combined through a learnable size-conditioned gating network. The resulting MoE aggregates per-atom energies into a single conservative potential, ensuring forces are obtained as energy gradients and enabling stable molecular dynamics. Across mixed cluster-nanoparticle test sets, the MoE improves accuracy relative to both an off-the-shelf foundation potential and a finetuned single-expert baseline. Stress tests show that force errors remain comparatively stable across cluster sizes and that the model retains robust energetics under morphology out-of-distribution shifts quantified using a structural outlier score based on similarity measures. The learned gating weights further provide an interpretable, size-dependent decomposition of interaction scales. Finally, validation against linear-scaling density functional theory using the ONETEP code, together with finite-temperature molecular dynamics tests, demonstrates consistent energetics, stable force behaviour, and well-behaved dynamical trajectories, supporting the use of the model for efficient structural optimisation and configurational sampling across nanoparticle sizes.
Yunyu Zhang, Keith T Butler, C. Richard A. Catlow· npj Computational Materials· 0 citations
Heterogeneous interfaces underpin technologies from microelectronics to energy conversion and storage, but their configurational complexity precludes exhaustive first-principles screening of interface registries. Although data-driven approaches can alleviate this burden, they remain limited by sparse interface datasets. Here, we introduce an energy-independent workflow that represents coherent interfaces as attributed graphs, quantifies their similarity to parent bulk environments using the fused Gromov-Wasserstein (FGW) distance, and couples this metric with Bayesian optimization over the in-plane registry space. We assess the approach for KI/NaCl, GaP/GaAs and GaN/$\mathrm{Al_{2}O_{3}}$ interfaces spanning ionic, covalent and mixed-bonding regimes, using hierarchical validation with MACE and density functional theory (DFT). Comparison with single-point energy landscapes shows that the FGW distance captures registry-dependent periodicity, while interfaces exhibit deviations between structural and energetic extrema, reflecting additional chemistry-specific contributions. Furthermore, FGW distances show an overall association with relaxed energies. Under limited screening budgets, FGW-guided registry selection consistently outperforms random search and is more robust across interface systems than selection guided by pretrained MACE energies. The workflow converts the qualitative notion of bulk-like continuity into a quantitative prescreening criterion, enabling efficient registry exploration and providing physically informed candidate structures for materials discovery workflows.