Civil Engineering Faculty Publications
Document Type
Article
Publication Date
10-1-2026
Abstract
Machine learning (ML) has increasingly been applied in civil and structural engineering for predicting material and structural performance. However, most existing studies either benchmark several algorithms within a single dataset or review such findings, leaving the cross-dataset stability of one fixed model configuration largely unquantified. This study proposes a generalized ML evaluation framework to investigate the robustness and stability of a fixed-configuration model across multiple civil engineering material datasets. Ten datasets representing diverse structural and material problems were compiled, including 3D printed concrete (3DPC), conventional and advanced concrete including normal and high strength concrete (NHSC), recycled aggregate concrete (RAC), ultra-high-performance concrete (UHPC), eco-friendly concrete (EFC), fly ash-based geopolymer concrete (FAGPC), reinforced concrete deep beam shear (RCDB), and self-compacted concrete (SCC). A unified CatBoost-based pipeline was implemented with identical preprocessing, training strategy, and evaluation metrics (R², MSE, RMSE, and MAE) to ensure fair cross-domain comparison. The results show that the CatBoost model demonstrates high predictive capability and stability in several material domains, particularly for 3DPC datasets, where testing R² values reached 0.971, 0.984, and 0.968 for compressive, flexural, and splitting tensile strength, respectively, with minimal performance degradation between training and testing phases (ΔR² ≤ 0.023). Conventional and alternative concrete datasets, including NHSC and RAC, also achieved strong generalization performance with testing R² values of 0.919 and 0.917, respectively. Stability analysis further confirmed that datasets with well-defined relationships between material composition and mechanical response demonstrate stronger model consistency, whereas datasets involving nonlinear structural behavior or rheological properties show larger performance degradation. In particular, stability decreased substantially for the UHPC and SCC datasets, indicating that a single fixed configuration cannot be assumed to generalize uniformly to highly complex or rheology-governed material domains.
Recommended Citation
Khan, Arslan Qayyum, Ali Raza, and Amorn Pimanmas. "Cross-dataset stability assessment of a fixed-configuration CatBoost model across heterogeneous civil engineering material datasets." Next Materials 13 (2026): 103079. https://doi.org/10.1016/j.nxmate.2026.103079
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Publication Title
Next Materials
DOI
10.1016/j.nxmate.2026.103079
Comments
2026 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).