Material-efficient 2D skeletal structure design using convolutional neural networks

Structural optimization plays a critical role in achieving efficient material utilization and maximizing structural performance. However, conventional approaches often suffer from high computational costs due to their iterative nature. Incorporating code-compliant design criteria enhances structural integrity and practical applicability but further escalates computational demands. This study presents a convolutional neural network-based framework augmented with a member section classification algorithm, achieving over a sixfold improvement in computational efficiency for generating code-compliant skeletal structures. The proposed workflow generates robust, ready-to-assemble frame designs based on user-defined parameters such as boundary conditions, initial domain geometry, and loading scenarios. Unlike most existing approaches that focus solely on material layout, this study introduces a method capable of jointly determining both structural topology and member dimensions, which is essential for creating assembly-ready designs. Generative convolutional neural models achieve over 91 % accuracy in generating optimal frame images. However, by integrating a strategic geometric feature identification algorithm, geometric features are identified with near-perfect accuracy, effectively overcoming the inherent limitations of generator-induced losses. The effectiveness of the approach is demonstrated through multiple case studies, validating the accuracy of predicted member sections under combined axial, bending, and shear loading, in accordance with established design codes. By incorporating a robust geometric feature extraction mechanism, the framework reliably produces designs that pass critical integrity checks and can be applied across multiple engineering domains.