Topology optimization utilizing deep learning technique with an emphasis on feature extraction

This paper utilizes a novel deep learning architecture-Optimizer-Net― specifically developed for structural optimization, to extract informative feature maps as part of an automated optimization process. Addressing key challenges such as high computational cost, complex parameter tuning, and convergence issues, Optimizer-Net transforms normalized energy data into contour or energy images that, together with optimized structures, serve as training input. The architecture consists of ۱۳ carefully designed layers, including convolutional, max-pooling, fully connected, transposed convolutional, and upsampling layers. These are further enhanced with batch normalization, leaky ReLU activation, dropout, and padding to ensure stable and efficient learning. Central to the model's success is its ability to extract and refine high-quality feature maps, which play a crucial role in capturing latent structural patterns within the energy contours. Performance evaluation using Mean Squared Error (MSE) demonstrates superior accuracy and optimization efficiency compared to conventional methods.

کلیدواژه ها:

Deep learning ، Mean squared error ، Structural optimization ، Feature map ، Optimizer Net