Quantifying the reserve capacity of deformed steel members: predictions from deep learning models utilizing point clouds

Sustainable construction and engineering risk management require the development of methods that enable reliable diagnosis of structural damage. This paper addresses the problem using deep learning applied to point clouds, demonstrating that geometric deformation alone provides sufficient information for accurate reserve capacity prediction. The deformed geometry of steel columns is mapped to their corresponding reserve capacities. Reserve capacity is defined as the ratio of the flexural strength of a column in its current deformed configuration to the maximum flexural strength attained during its loading history. A new dataset of point clouds and corresponding reserve capacities, obtained through validated high-fidelity simulations is used for model development and testing. As load history is unavailable in real-world applications, different point-based deep learning models relying solely on point clouds are explored. Results demonstrate the potential to train a single deep learning model capable of achieving robust reserve capacity estimations, yielding a mean prediction error of 2.6 % for samples with reserve capacity above 0.4. Incorporating additional structure-related knowledge as predictive features, such as axial load ratio and cross-section profile, improves the prediction error to 2.1 %. This study provides compelling evidence that the reserve capacity of steel columns can be reliably quantified using only geometric information. • Reserve capacity of steel members is quantified using deep learning on point clouds. • Deformed 3D geometry of steel columns is directly mapped to reserve capacities. • Extensive dataset of high-fidelity simulations is used for model development/testing. • Reserve strength of steel members is reliably quantified using only visual information. • Knowledge of the loading history does not meaningfully affect model predictions.