Probabilistic resistance predictions of laterally restrained cellular steel beams by natural gradient boosting

Abstract

Accurate and reliable cellular steel beam resistance predictions are essential for economical and safe designs of steel-framed buildings with such beams. This paper proposes a new machine-learning (ML) model based on the natural gradient boosting (NGBoost) algorithm to predict probabilistic load-bearing capacities of laterally restrained cellular beams subjected to uniformly distributed loads, considering all possible failure modes and their interactions. The NGBoost model was developed based on a database with 14,094 numerical simulation results and interpreted using the SHapley Additive exPlanations (SHAP) method commonly used for ML model explanation and interpretation. The resistance reduction factors required for the NGBoost model to meet the reliability requirements of the European and US design frameworks were determined via reliability analyses using the methods given in the respective standards and the improved Hasofer–Lind–Rackwitz–Fiessler (iHL-RF) method. Comparisons of the developed NGBoost model with other ML models and existing design provisions indicate that the former is as accurate as other ML models (while offering probabilistic predictions) and significantly outperforms the existing design provisions. A web application was developed and deployed online to predict the ultimate uniform loads of laterally restrained cellular beams with the developed NGBoost model. The proposed NGBoost model can facilitate preliminary cellular steel beam designs and investigating parameters affecting their resistance.