Vitor F. Grizzi, Benjamin Nebgen, Yang Zhang
The introduction of high-valence actinides, such as thorium and uranium, into molten alkali halide induces complex structural correlations and intermediate-range ordering, presenting a daunting challenge for computational materials science. Because these dynamic topological features span length scales and timescales that are inaccessible to traditional ab initio molecular dynamics, capturing their influence on macroscopic transport properties requires highly scalable, high-fidelity sampling. Here, we employ an uncertainty-driven active-learning framework to construct a robust machine-learned interatomic potential (MLIP) for the multicomponent ${\text{LiF-NaF-ThF}}_{4}$ ionic liquid. With the resulting MLIP, we performed 400 ps simulations of $100\phantom{\rule{0.16em}{0ex}}000$-atom systems at six temperatures spanning 973--1473 K with near--density functional theory accuracy. These large-scale simulations allowed us to compute key thermophysical, structural, and transport properties over a broad temperature range while substantially reducing the finite-size effects that limit the convergence of Green-Kubo transport coefficients in highly correlated liquids. Our simulations reveal that the solvation of ${\mathrm{Th}}^{4+}$ drives the formation of transient, medium-range (${\mathrm{ThF}}_{n}$) clusters. We demonstrate that this emergent structural inhomogeneity fundamentally alters the transport mechanisms of the melt: the bulky clusters impede ionic mobility and disrupt collective vibrational modes. This interplay between dynamic local structure and energy dissipation leads to an increased shear viscosity and a significantly reduced thermal conductivity relative to the pure eutectic salt. Collectively, these results elucidate the profound impact of dynamic structural heterogeneities on dissipative transport, underscoring MLIP-driven simulations as a necessary tool to achieve the spatiotemporal resolution required to study correlated alkali halide melts.