Answers are grounded in this paper and cite their sources.
Machine learning (ML) models are increasingly adopted in structural engineering to capture the complex behavior of structural components, connections, and systems. When large, high-quality datasets are available, such models can outperform traditional analytical and empirical approaches. Recently, empirical predictive models based on linear regression were developed to estimate the full-range moment–rotation response of semirigid extended endplate connections using a comprehensive dataset. In this study, the same dataset is employed to develop artificial neural network (ANN) models with different architectures and extreme gradient boosting models. The performance of these models is systematically evaluated and compared against one another, as well as against code-based provisions and conventional linear regression models. The new ML models show superior performance, with more than 75% of the data predicted with an error less than 20%, even for the most sensitive response parameters, elastic stiffness and rotational ductility. The consistency of learned relationships with established joint mechanics is examined using Shapley additive explanations and Accumulated Local Effects analysis. Most importantly, it is shown that multi-input multi-output ANNs can achieve predictions as high as multi-input single-output ANNs with almost half the data set size. This can inform the efficient development of data-driven models for similar problems. The developed models’ mathematical sources are made available along with a computer interactive tool.
No takes yet. Share an insight, caveat, or question.
Ding et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: