Calibration of polyvinylidene fluoride (PVDF) stress gauges under high-impact dynamic compression by machine learning

Calibration of stress gauges is of great importance for understanding the behaviors of materials under high dynamic impacts. However, commonly used calibration models have little transferability due to ignoring the influences of the gauge parameters. In this work, we propose a systematic approach that can generate effective and transferable calibration models including multiple independent variables by machine learning. Specifically, we conduct high-impact dynamic compression experiments using polyvinylidene fluoride (PVDF) stress gauges with two different thicknesses and varying remnant polarizations at shock levels from 0.3 to 10 GPa. To best characterize the comprehensive calibration relationship, we select a set of five features (combined by strain, remnant polarization, and film thickness) by feature engineering and use Lasso with the bagging ensemble as an algorithm to train the machine learning model. For comparison, we also propose semiempirical models that calibrate PVDF gauges effectively, but without including thickness and remnant polarization. Our results show that the machine learning model is more precise and more reasonable in physics. The predicted dependences of the calibration curves on remnant polarization and film thickness by the machine learning model are qualitatively consistent with the physics scenario. This work reveals the potential of machine learning methods to improve gauge calibration for better performance and transferability. The method used in this work is applicable to the calibration of any stress gauges with multiple variables.