A GA-Optimized CNN-LSTM Model for Predicting Simulated Anchor Radius in Glassblown Micro Hemispherical Resonators
Abstract Accurately predicting the anchor radius of micro hemispherical resonators (MHRs) in high-temperature glassblowing simulations is essential for identifying designs that meet assembly feasibility and performance targets prior to fabrication. However, the complex nonlinear deformation during the forming process makes this prediction challenging, and traditional finite element simulation is too time-consuming for efficient design exploration. This study presents a hybrid CNN (Convolutional Neural Network)-LSTM (Long Short-Term Memory) deep learning model optimized by a genetic algorithm to predict the simulated anchor radius from three design parameters based on finite element simulation data: resonator height, designed anchor radius, and edge thickness. Using a dataset of 314 finite element simulations, the proposed model achieves a test root mean square error of 0.0065 and an R2 of 0.9996. Ablation studies confirm that both the CNN and LSTM components contribute meaningfully to prediction accuracy, while comparative experiments show that the model performs well for this small-sample regression task. After approximately 4 hours of hyperparameter optimization and training, the model achieved prediction within milliseconds. Compared to the 260 hours required for generating training datasets using traditional finite element methods, this method significantly reduces the time required. This computational efficiency enables rapid design space exploration within the simulation environment and demonstrates the potential of hybrid deep learning approaches for efficient Micro-Electro-Mechanical System (MEMS) design under complex simulated manufacturing conditions.
Authors
- Gongliu Yang (ORCID: https://orcid.org/0009-0000-0677-9308)
- Yuyi Yao (ORCID: https://orcid.org/0009-0004-7947-9938)
Institutions
- Zhejiang University (CN)
Publication Details
- Journal
- Engineering Research Express
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1088/2631-8695/aec359
- Primary Topic
- Advanced MEMS and NEMS Technologies
- Type
- article
- Field-Weighted Citation Impact
- 0.00