Machine learning-guided atomic-scale design of binary rare-earth niobates with ultra-low thermal conductivity

Achieving ultralow thermal conductivity in ceramics remains a central challenge for thermal management in extreme aerospace and energy applications. While high-entropy strategies suppress phonon transport through lattice disorder, their utility is constrained by synthetic complexity. Here, we introduce a machine learning-accelerated, atomic-scale design strategy that enables the discovery of a simple binary rare-earth niobate with ultra-low thermal conductivity. Guided by an optimized gradient boosting regression model, we identified and synthesized (Y_0.3Dy_0.7)₃NbO₇, exhibiting 0.98 W m⁻¹ K⁻¹, lower than most of the reported high-entropy niobates despite its low compositional complexity. Atomic-scale characterization reveals that size-mismatched cations induce local lattice distortion, thereby enhancing phonon scattering. Combined DFT and MD simulations confirm that Dy substitution disrupts Y vibrations, reduces phonon participation, and promotes diffuson-mediated transport. These findings not only establish a benchmark for low-complexity thermal barrier ceramics but also highlight the transformative potential of machine-learning-driven materials discovery for next-generation thermal management.