Active Learning Driven Materials Discovery for Low Thermal Conductivity Rare-Earth Pyrochlore for Thermal Barrier Coatings

TL;DR

This work addresses the challenge of discovering low-thermal-conductivity pyrochlore oxides for thermal barrier coatings in a data-scarce regime. It applies an active learning loop with a Random Forest surrogate and Expected Improvement to navigate the multi-component rare-earth pyrochlore space, using crystallographic descriptors and lattice entropy as design features. The study identifies four candidate compositions across two iterations, with two single-phase pyrochlores showing measured TCs near 2.0 W/mK and the second iteration revealing dual-phase products, underscoring the need to incorporate phase formation into the AL framework. Overall, the approach demonstrates accelerated materials discovery toward tailored TBCs, while highlighting practical constraints related to phase stability and data efficiency that guide future improvements.

Abstract

High-Entropy/multicomponent rare-earth oxides (HECs and MCCs) show promise as alternative materials for thermal barrier coatings (TBC) with the ability to tailor properties based on the combination of rare-earth elements present. By enabling the substitution of scarce or supply-risk rare-earths with more readily available alternatives while maintaining comparable material performance, HECs and MCCs offer a valuable path towards alternative TBC material design. However, navigating this search space of compositionally complex materials is both time and resource intensive. In this study, an active learning (AL) framework was employed to identify HEC/MCC materials with a pyrochlore structure, with acceptable thermal conductivity (TC) for TBC applications. The AL framework was applied through a Bayesian optimisation (BO) strategy, coupled with a random forest surrogate model. TC was selected as the optimisation criterion as that is the most basic requirement of TBC materials. Over two iterations of the AL cycle, four compositions were generated and synthesized in the lab for experimental evaluation. The first iteration yielded two single-phase pyrochlores, $(La_{0.29}Nd_{0.36}Gd_{0.36})_2Zr_2O_7$ and $(La_{0.333}Nd_{0.26}Gd_{0.15}Ho_{0.15}Yb_{0.111})_2Zr_2O_7$, with measured thermal conductivities of 2.03 and 1.90 $W/mK$, respectively. The surrogate model predicted a TC of 2.009 $W/mK$ for both compositions, demonstrating it's accuracy for completely new compositions. The second iteration compositions showed dual-phase when synthesized, highlighting the need to take into account phase formation in the AL framework.

Figures (13)

• Figure 1: Active learning workflow
• Figure 2: Thermal conductivity distribution of training data
• Figure 3: Thermal conductivity vs $R_{A}$ of lanthanum gadolinium zirconate series
• Figure 4: XRD patterns of active learning generated compositions compared to LZO and GZO. The (622) peaks are shown magnified on the right.
• Figure 5: Predicted thermal conductivity vs actual thermal conductivity of test data using surrogate model trained on the original dataset