Intelligent structural design of composite concrete‐encased steel columns based on hybrid machine learning and multiobjective optimization
- 5 August 2025
- journal article
- Published by Wiley in Structural Concrete
- Vol. 27 (1) , 49-72
Abstract
Accurate prediction of axial load capacity in composite concrete‐encased steel columns (CCESC) remains challenging due to complex nonlinear interactions between materials and geometric parameters, with traditional methods often showing up to 10%–30% prediction relative errors. To address these limitations, this study proposes a hybrid machine‐learning framework that integrates a Backpropagation Neural Network (BPNN) with a Dynamic Multi Population‐Particle Swarm (DMP‐PSO) algorithm for high‐precision axial load prediction. A comprehensive database was established, encompassing key variables such as concrete strength, steel yield strength, cross‐sectional dimensions, and reinforcement configurations. The DMP‐PSO algorithm dynamically optimizes the hyperparameters of the BPNN, achieving superior prediction accuracy. The predictive performance significantly outperforms that of BPNNs optimized using PSO, Bayesian optimization (BO), and grid search methods, as well as multi‐layer perceptron (MLP) optimized through DMP‐PSO. SHAP (SHapley Additive exPlanations) analysis identified concrete compressive strength and steel flange thickness as dominant factors influencing predictions, while parametric studies underscored the nonlinear interactions between material and geometric variables. Furthermore, the developed multi‐objective optimization (MOO) model employing NSGA‐II (Nondominated Sorting Genetic Algorithm II) enables automated design optimization under material, dimensional, and structural constraints. By systematically analyzing Pareto frontiers, this approach achieves rapid identification of cost‐minimized configurations while satisfying specified load‐bearing requirements, effectively replacing traditional manual trial‐and‐error processes. The integrated framework not only reduces material waste through ultra‐precise capacity predictions but also establishes a systematic methodology for balancing structural safety and economic efficiency in composite system design.Keywords
Funding Information
- National Key Research and Development Program of China (2024YFC3809803)
This publication has 65 references indexed in Scilit:
- Enhancing multi-objective prediction of settlement around foundation pit using explainable machine learningJournal of Civil Structural Health Monitoring, 2025
- Transfer learning framework for the wind pressure prediction of high-rise building surfaces using wind tunnel experiments and machine learningBuilding and Environment, 2025
- A novel fuzzy three-valued logic computational framework in machine learning for medicine datasetComputers in Biology and Medicine, 2025
- Experimental study on the axial compressive performance of prefabricated steel reinforcement cage‐cast‐in‐situ concrete column with an innovative sleeveStructural Concrete, 2024
- Experimental behaviour and design model of FRP-UHPC-steel tubular columns under monotonic axial compressionComposite Structures, 2024
- Experimental and numerical study of stirrup‐confined concrete‐filled L‐shaped steel tube stub column under axial compressionStructural Concrete, 2024
- Seismic fragility and post-earthquake reparability of concrete encased-and-filled steel tubular bridges columns with debonded high-strength reinforcementsBulletin of Earthquake Engineering, 2023
- STRUCTURAL BEHAVIOUR OF JOINT BETWEEN STEEL BEAM AND CONCRETE ENCASED STEEL COMPOSITE COLUMNce/papers, 2019
- A hybrid particle swarm optimization–back-propagation algorithm for feedforward neural network trainingApplied Mathematics and Computation, 2007
- The normalized normal constraint method for generating the Pareto frontierStructural and Multidisciplinary Optimization, 2003