Predictive modelling and latent space exploration of steel profile overstrength factors using multi‐head autoencoder‐regressors

This paper investigates the suitability and interpretability of a data‐driven deep learning algorithm for multi cross sectional overstrength factor prediction. For this purpose, we first compile datasets consisting of experiments from literature on the overstrength factor of circular, rectangular and square hollow sections as well as I‐ and H‐sections. We then propose a novel multi‐head encoder architecture consisting of three input heads (one head per section type represented by respective features), a shared embedding layer as well as a subsequent regression tail for predicting the overstrength factor. By construction, this multi‐head architecture simultaneously allows for (i) the exploration of the nonlinear embedding of different cross‐sectional profiles towards the overstrength factor within the shared layer, and (ii) a forward prediction of the overstrength factor given profile features. Our framework enables for the first …