Machine-learning-guided inverse design of lead-free relaxors enabled by multimodal literature mining.

Song, Kwanwoo; Kim, Youngmin; Kim, Jaehyun; Min, Byeong-Jae; Song, Hyun-Cheol; Kang, Nayeon; Ryu, Jungho; Jang, Ho Won · Nat Commun · 2026

basic_science · Level V

Where this comes from

Abstract

Lead-free relaxors combining high permittivity with thermal robustness are essential for integrated electronics and power systems operating at elevated temperatures. However, their development remains fundamentally hindered by inherent property trade-offs and the immense chemical search space. Here, we report an inverse-design framework integrating multimodal literature mining with physics-guided descriptor-based deep ensemble learning for systematic relaxor screening. By screening a combinatorial space exceeding 150 million candidates, we identify the (Sr<sub>0.48</sub>Na<sub>0.26</sub>Bi<sub>0.26</sub>)(Ti<sub>1-x</sub>Sn<sub>x</sub>)O<sub>3</sub> (0 ≤ x ≤ 0.02) compositions as a low-complexity composition window that satisfies a sparsely populated tri-target dielectric-property regime. Experimental validation confirms that compositions with x = 0.01 and 0.02 meet the predefined design targets for dielectric stability while maintaining a room temperature permittivity above 3300. This stability is further reflected in the simultaneous satisfaction of the upper-temperature-side X5R/X6R/X7R criteria. Atomic-scale characterization reveals that octahedral framework expansion induces polar heterogeneity, underlying the observed dielectric performance. Our results establish an experimentally validated down-selection framework for navigating complex relaxor design spaces and accelerating the data-driven discovery of temperature-stable lead-free dielectrics.