Data-driven accelerated design of non-equimolar rare-earth silicates with phase stability and high CMAS resistance at 1550 °C

The vast compositional space of non-equimolar systems requires data-driven methods to predict phase stability and CMAS resistance of rare-earth silicates (RESs). A dataset of phase structure and CMAS resistance was established by combining RESs synthesized in this work via sol–gel methods with open data. Rietveld-refined XRD data revealed linear correlations among $I_{max}$$I_{max}$, lattice constants, distortion, and ${\bar{r}}_{Me}$${\bar{r}}_{Me}$. A total of 37 potential descriptors were screened using machine learning. A voting ensemble classifier identified rare-earth disilicates compositions forming a stable $\beta$$\beta$-phase, without enforcing a strict distinction between multi- and $\gamma$$\gamma$-phase, due to varying testing temperatures across research groups. A corrosion grading function (CRG) was introduced to mathematically classify the CMAS resistance levels of RESs, reducing the dimensionality of experimental datasets. The XGBoost model was used to predict about 3.5 million non-equimolar compositions by systematic enumeration five elements from a cost-effective RE pool (Yb, Tm, Er, Y, Ho, Tb, Gd), identifying optimized formulations. Experimental validation at 1550 $°\mathrm{C}$$°\mathrm{C}$ for 24 h on the composition ${\left({\mathrm{Gd}}_{0.05}{\mathrm{Er}}_{0.1}{\mathrm{Yb}}_{0.375}{\mathrm{Y}}_{0.025}{\mathrm{Tm}}_{0.45}\right)}_2{\mathrm{SiO}}_5$${\left({\mathrm{Gd}}_{0.05}{\mathrm{Er}}_{0.1}{\mathrm{Yb}}_{0.375}{\mathrm{Y}}_{0.025}{\mathrm{Tm}}_{0.45}\right)}_2{\mathrm{SiO}}_5$ showed an average corrosion layer thickness of approximately $33.5\mathrm{\mu }\mathrm{m}$$33.5\mathrm{\mu }\mathrm{m}$, confirming the effectiveness of the data-driven approach in accelerating the design of non-equimolar RESs for environmental barrier coatings.