Transfer learning-driven prediction of mechanical properties in membrane resonators

We demonstrate the effectiveness of transfer learning for enhancing predictions of critical mechanical properties of membrane resonators. This approach benefits from robust pretrained features, which improve convergence and generalization even with large training datasets. By fine-tuning pretrained deep learning architectures—including VisionTransformer, ResNet152, and VGG19—we accurately predict eigenfrequencies, dilution coefficients, and buckling behaviors. Using a large dataset of 170 000 samples generated with COMSOL Multiphysics, our models capture complex physical phenomena, delivering performance comparable to traditional finite element methods but at significantly reduced computational cost. Specifically, eigenfrequency predictions across 15 eigenmodes consistently achieve low relative errors (1%–3%) with minimal variance. Dilution coefficient predictions remain accurate for lower eigenmodes, maintaining relative errors below 10% for the four lowest modes using static pruning thresholds. However, for higher modes, accuracy deteriorates—potentially due to the increased complexity and localized nature of higher-order mode shapes. In addition, our models exhibit robust performance in classifying membrane buckling behaviors in a binary setting, achieving over 95% accuracy.