Interpretable machine learning-aided prediction of ferroelectric property in HfO2-based thin films

Hafnium oxide (HfO 2 )-based ferroelectrics are promising for non-volatile memories. While the trial-and-error approach is inefficient, machine learning (ML) offers a rapid alternative. However, ML prediction for HfO 2 -based ferroelectrics is hindered by data sparsity. In this work, an intelligence ferroelectricity predictive framework with high predictive precision was proposed. Considering the formation of unique HfO 2 metastable ferroelectric phase assisted by the synergistic effects of doping, size effect, electrode clamping effect and so on, the processing conditions and dopant physicochemical features were used as joint inputs to capture their coulpling information and mitigate data sparsity. Among 107 models evaluated under an automated ML framework, the random forest model achieved the highest accuracy and robustness in predicting remanent polarization ( P r ) of HfO 2 -based ferroelectric thin films, with R 2 values of 0.86 (training) and 0.91 (testing). In contrast, the framework with only physicochemical features performed poorly with the highest R ² values of 0.42 (training) and 0.50 (testing), revealed the critical synergistic effect between dopant characteristics and processing conditions on P r . 7 key features were identified from the initial 37 features with top three important features of mean molar volume, annealing temperature, and top electrode thickness. Interpretable ML-based SHAP analysis revealed nonlinear relationships between these features and P r , elucidating potential underlying physical mechanisms. Lastly, the well-matched experimental results with predictions further validated this predictive framework.