Answers are grounded in this paper and cite their sources.
A physics-constrained, FEA-coupled deep neural optimization framework is proposed for the discrete topology and size optimization of steel structural systems. The framework generates candidate designs with a deep neural network and evaluates them through SAP2000 finite element analyses accessed via its Open Application Programming Interface (OAPI). AISC-based strength and serviceability constraints are incorporated into an adaptive penalized objective function, so no manually labeled dataset is required; the training signal is derived from the code-compliant response of each candidate design computed by the external solver. Geometric and topological characteristics of structural members are encoded through centroid coordinates together with binary indicators representing the presence or absence of bracing members, forming the input space of the network. The trained network predicts cross-sectional areas while satisfying the strength and serviceability requirements defined by AISC-ASD and LRFD provisions. Because structural responses are obtained from an external finite element solver rather than embedded differential-equation residuals, the proposed method should be distinguished from classical continuous-domain physics-informed neural networks. The framework is assessed on ten-story and twenty-four-story steel braced frames, a five-bay six-story frame, and a pedestrian bridge considered for topology optimization. Compared with the metaheuristics MVPO, TLBO, SLOCA, and ABC, the proposed method yields competitive or lighter designs while reducing the number of required structural analyses by approximately 50–70%, indicating its suitability for large-scale structural optimization.
No takes yet. Share an insight, caveat, or question.
Ghasri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: