Abstract
With the advancement of scientific research, understanding the physical parameters governing acoustic wave propagation in sea ice has become increasingly important. Among these parameters, shear wave velocity plays a crucial role. However, as measurements progressed, it became apparent that there was a large discrepancy between measured values of shear waves and predictions based on empirical formulas or existing models. These inconsistencies stem primarily from the complex internal structure of natural sea ice, which significantly influences its physical behavior. Research reveals that shear wave velocity is not only influenced by bulk properties such as density, temperature, and stress state but is also sensitive to microstructural features, including air bubbles, inclusions, and ice crystal orientation. Compared to longitudinal wave velocity, the characterization of shear wave velocity is far more challenging due to these inherent complexities, underscoring the need for more precise measurement and modeling techniques. To address the challenges posed by the complex internal structure of natural sea ice and improve prediction accuracy, this study introduces a novel, integrated approach combining simulation, measurement, and inversion intelligent learning model. First, a laboratory-based method for generating sea ice layers under controlled formation conditions is developed. The produced sea ice layers align closely with measured values for Poisson’s ratio, multi-year sea ice density, and uniaxial compression modulus, particularly in the high-temperature range. Second, enhancements to shear wave velocity measurement equipment have been implemented. The improved device achieves measurement accuracy exceeding 1%, offers portability, and meets the demands of high-precision experiments conducted in harsh polar environments. Finally, according to the characteristics of small sample data. The ANN neural network was improved to a deep residual neural network with the addition of Squeeze-and-Excitation Attention (SE-ResNet) to predict longitudinal and transverse wave velocities. This prediction method improves the accuracy of shear and longitudinal wave velocity prediction by 24.87% and 39.59%, respectively, compared to the ANN neural network.