Knowledge-enhanced generative adversarial networks for schematic design of framed tube structures

Abstract

In the schematic design phase of framed tube structures, component sizing is a vital task that requires expert experience and domain knowledge. Deep learning-based structural design methods enable machines to acquire expert experiences, but domain knowledge (e.g., empirical laws summarized by engineers from engineering practices) has not been embedded into such data-driven methods, resulting in common sense-conflicting designs. A knowledge-enhanced generative adversarial network is proposed by incorporating a novel differentiable evaluator for compliance checking of domain knowledge. A comparative study indicates that the proposed knowledge-enhanced method is 51% superior to the conventional data-driven method and 150 times faster than a competent engineer. The proposed method facilitates the schematic design of framed tube structures to be automatic and efficient, hence improving the productivity of structural engineers.