Data-driven screening of stable lead-free piezoelectric materials via machine learning and graph-based models

The identification of lead-free, stable piezoelectric materials remains a key challenge for sustainable technologies. Although first-principles calculations based on density functional theory (DFT) provide reliable property predictions, their high computational cost prevents their direct use in large-scale screening. Here we present a systematic machine learning (ML) benchmark for the prediction of the maximum piezoelectric modulus e m of inorganic crystalline materials, trained on > 3,000 DFT-computed compounds from the Materials Project. We evaluate thirteen predictive architectures spanning gradient-boosted tree ensembles, deep neural networks, and seven graph neural network (GNN) models. The three top-performing models are then used to screen promising piezoelectric materials among > 10,000 unlabeled Materials Project compounds. After filtering for thermodynamic near-stability ( E hull < 0.1 eV/atom), insulating behavior ( E g > 0.1 eV), and non-centrosymmetric crystal structures, eight promising lead-free candidates are identified. Independent DFT calculations of the full piezoelectric tensor for twenty selected compounds confirm the screening results. Among the identified materials, SrTaNO 2 , YWN 3 , and SrTa 2 Bi 2 O 9 emerge as the most promising targets for experimental synthesis as they exhibit large piezoelectric moduli in the range of 3–103 C/m 2 .