Gears are one of the most representative mechanical components in power transmission systems, transmitting power, changing speeds, and amplifying torque. Gear lightweighting directly impacts improvements in fuel efficiency, energy savings, and manufac...
Gears are one of the most representative mechanical components in power transmission systems, transmitting power, changing speeds, and amplifying torque. Gear lightweighting directly impacts improvements in fuel efficiency, energy savings, and manufacturing cost reduction. However, achieving lightweight designs in gears is not simply a matter of reducing material; it requires careful consideration of the complex stress distribution changes and structural reliability associated with stiffness reduction. With the advancement of cutting-edge manufacturing technologies such as 3D printing, design freedom has significantly expanded, increasing the need to handle non-standard, irregular blank shapes that traditional design methods struggle to address. In lightweight gear design, it is essential not only to generate shapes but also to accurately predict the maximum stress and its location for any given design. Without guaranteed reliability, lightweight designs can lead to mechanical failures or fatigue fractures under real-world operating conditions, making stress prediction and reliability-based optimization indispensable elements in gear lightweighting research.
To address these challenges, this study proposes an integrated lightweight optimization framework that combines high-fidelity finite element analysis (FEA), machine learning–based surrogate modeling, and deep learning–based generative design techniques, covering both gear types with and without well-defined design variables.
First, to achieve a level of lightweighting beyond existing international standards, three types of test gears solid gears, thin-rimmed gears, and thin-rimmed gears with additional holes were designed, manufactured, and experimentally evaluated for tooth root stress and blank stress. The experiments revealed that thin-rimmed and holed gears showed approximately 2.5 times higher maximum root stress and 13 times higher blank stress compared to solid gears.A high-fidelity FEA model reflecting the same rotation, lubrication, boundary, and loading conditions as the experiments was developed and validated against the measured data, ensuring the reliability of the simulation.
For gears with defined design variables, a machine learning–based Gaussian Process Regression (GPR) surrogate model was developed to replace time-consuming FEA and enable rapid prediction of maximum stress and weight. Using this model, a reliability-based design optimization (RBDO) process that accounts for manufacturing tolerances was performed. Through iterative GA single-objective optimization, optimal design parameters achieving a target reliability of 97.5% were derived, resulting in a final design achieving approximately 53.59% weight reduction compared to solid gears while maintaining structural integrity within allowable stress limits.
For gears without clearly defined design variables, this study proposes a deep learning pipeline based on a multi-output 3D CNN model, trained on diverse random gear blank shapes generated through topology optimization and latent space analysis using autoencoders. By utilizing sparse tensor inputs, the model achieves substantial reductions in memory and computational demands while effectively capturing critical hotspot information. This enables simultaneous high-accuracy predictions of both maximum stress and its location.
The proposed methodology enables two complementary applications depending on the nature of the design problem. For cases where design variables are clearly defined, the trained GPR model allows rapid prediction and efficient exploration of the design space. For non-parametric cases involving complex or irregular geometries, the 3D CNN based generative approach can predict stress magnitude and location directly from geometry. By bridging variable based optimization with data driven learning, this work contributes to expanding the applicability of AI-based reliability evaluation and optimization for the design of complex-shaped gears.