TY - JOUR
T1 - ML-based prediction of shear strength for steel sheathed cold-formed steel shear wall
AU - Liu, Shen
AU - Wei, Jiaxin
AU - Lin, Bozhong
AU - Yao, Yao
N1 - Publisher Copyright:
Copyright © 2026. Published by Elsevier Ltd.
PY - 2026/12
Y1 - 2026/12
N2 - Cold-formed steel (CFS) shear walls with steel sheathing have gained widespread adoption in multi-story buildings owing to favorable seismic performance. However, existing design approaches such as the effective strip method (ESM), tabulated methods, and numerical simulations suffer from inherent trade-offs among prediction accuracy, computational cost, and applicability. To address these limitations, this paper develops a machine learning (ML) framework for reliably predicting the shear strength for such shear walls. A hybrid database combining 230 experimental and 100 validated finite element (FE) results was developed, with input features standardized via Z -score normalization and stratified sampling applied to preserve data representativeness. Six ML models, including Ridge Regression (RR), Lasso Regression (LR), Support Vector Regression (SVR), Decision Tree (DT), Random Forest (RF), and eXtreme Gradient Boosting (XGB), were trained and evaluated against current methods in AISI S400. Results indicate that ensemble models, especially XGB and RF, achieve superior predictive accuracy and robustness compared to linear models. Notably, the proposed XGB model consistently outperforms existing codified methods across U.S., Canadian, and other regional datasets, while quantifying the contribution of the blocking to shear strength. Reliability analysis further yields a resistance factor of 0.82 for the XGB model, with a divisor of 1.33 to maintain consistency with the AISI design framework.
AB - Cold-formed steel (CFS) shear walls with steel sheathing have gained widespread adoption in multi-story buildings owing to favorable seismic performance. However, existing design approaches such as the effective strip method (ESM), tabulated methods, and numerical simulations suffer from inherent trade-offs among prediction accuracy, computational cost, and applicability. To address these limitations, this paper develops a machine learning (ML) framework for reliably predicting the shear strength for such shear walls. A hybrid database combining 230 experimental and 100 validated finite element (FE) results was developed, with input features standardized via Z -score normalization and stratified sampling applied to preserve data representativeness. Six ML models, including Ridge Regression (RR), Lasso Regression (LR), Support Vector Regression (SVR), Decision Tree (DT), Random Forest (RF), and eXtreme Gradient Boosting (XGB), were trained and evaluated against current methods in AISI S400. Results indicate that ensemble models, especially XGB and RF, achieve superior predictive accuracy and robustness compared to linear models. Notably, the proposed XGB model consistently outperforms existing codified methods across U.S., Canadian, and other regional datasets, while quantifying the contribution of the blocking to shear strength. Reliability analysis further yields a resistance factor of 0.82 for the XGB model, with a divisor of 1.33 to maintain consistency with the AISI design framework.
KW - Cold-formed steel shear wall
KW - Effective strip method
KW - Machine learning
KW - Shear strength
KW - Steel sheathing
UR - https://www.scopus.com/pages/publications/105048166456
U2 - 10.1016/j.jcsr.2026.110648
DO - 10.1016/j.jcsr.2026.110648
M3 - 文章
AN - SCOPUS:105048166456
SN - 0143-974X
VL - 247
JO - Journal of Constructional Steel Research
JF - Journal of Constructional Steel Research
M1 - 110648
ER -