Integrating Positron Annihilation Lifetime Spectroscopy with Interpretable Machine Learning for Dielectric Property Analysis of Low-k Polyimide Films

Low-k polyimide films are important dielectric candidates for advanced microelectronic and packaging applications, where reducing capacitive delay and power dissipation requires a clearer understanding of how nanoscale free volume affects dielectric response. Positron annihilation lifetime spectroscopy (PALS) provides direct information on subnanometer free-volume cavities, but conventional analyses often rely on single-parameter correlations. Here, we integrate PALS-derived descriptors with interpretable machine learning to analyze dielectric constant variation in 17 low-k polyimide samples prepared and characterized under comparable experimental conditions. Three descriptor sets and five regression models were evaluated using leave-one-out cross-validation. The ElasticNet model based on the expanded descriptor set achieved the best performance, with an R 2 of 0.877, a mean absolute error of 0.192, and a root-mean-square error of 0.236, while a compact physics-informed descriptor set retained strong predictive capability. Feature importance, permutation importance, and SHAP analyses consistently highlighted I 3, the intensity-weighted free-volume descriptor T ve, and cavity size and free-volume-related descriptors as important variables in the model. This study provides an interpretable small-data framework for connecting PALS-measured free-volume characteristics with dielectric behavior in low-k polyimide films.