TY - JOUR
T1 - Multi-objective design optimization for graphite-based nanomaterials reinforced cementitious composites
T2 - A data-driven method with machine learning and NSGA-Ⅱ
AU - Dong, Wei
AU - Huang, Yimiao
AU - Lehane, Barry
AU - Ma, Guowei
PY - 2022/5/9
Y1 - 2022/5/9
N2 - Graphite-based nanomaterials (GNs) are promising conductive fillers for producing highly effective electrically conductive cementitious composites (ECCC) and promoting non-destructive structural health monitoring (SHM) methods. Since acceptable mechanical strength and electrical resistivity are both required, the design of GN-reinforced cementitious composites (GNRCC) is a complicated multi-objective optimization problem (MOOP). The present study proposes a comprehensive data-driven method to address this multi-objective design optimization (MODO) issue for GNRCC using machine learning (ML) techniques and non-dominated sorting genetic algorithm (NSGA-Ⅱ). First, prediction models of uniaxial compressive strength (UCS) and electrical resistivity (ER) of GNRCC are established by Bayesian-tuned XGBoost with prepared experimental datasets. The results show that they have excellent performance in predicting both properties with high R2 (0.95 and 0.92, 0.99 and 0.98) and low mean absolute error (MAE) scores (1.24 and 3.44, 0.15 and 0.22). The influence of critical features on GNRCC's properties are quantified by ML theories. This helps determine the variables to be optimized and define their constraints for the MODO. Finally, the MODO program is developed on the basis of NSGA-Ⅱ. It optimizes GNRCC's properties of UCS and ER simultaneously with the proposed prediction models as objective functions. It successfully achieves a set of Pareto solutions, which can facilitate appropriate parameters selections for the GNRCC design.
AB - Graphite-based nanomaterials (GNs) are promising conductive fillers for producing highly effective electrically conductive cementitious composites (ECCC) and promoting non-destructive structural health monitoring (SHM) methods. Since acceptable mechanical strength and electrical resistivity are both required, the design of GN-reinforced cementitious composites (GNRCC) is a complicated multi-objective optimization problem (MOOP). The present study proposes a comprehensive data-driven method to address this multi-objective design optimization (MODO) issue for GNRCC using machine learning (ML) techniques and non-dominated sorting genetic algorithm (NSGA-Ⅱ). First, prediction models of uniaxial compressive strength (UCS) and electrical resistivity (ER) of GNRCC are established by Bayesian-tuned XGBoost with prepared experimental datasets. The results show that they have excellent performance in predicting both properties with high R2 (0.95 and 0.92, 0.99 and 0.98) and low mean absolute error (MAE) scores (1.24 and 3.44, 0.15 and 0.22). The influence of critical features on GNRCC's properties are quantified by ML theories. This helps determine the variables to be optimized and define their constraints for the MODO. Finally, the MODO program is developed on the basis of NSGA-Ⅱ. It optimizes GNRCC's properties of UCS and ER simultaneously with the proposed prediction models as objective functions. It successfully achieves a set of Pareto solutions, which can facilitate appropriate parameters selections for the GNRCC design.
KW - Compressive strength
KW - Electrical resistivity
KW - GNRCC
KW - Machine learning
KW - Multi-objective design optimization
KW - NSGA-Ⅱ
UR - http://www.scopus.com/inward/record.url?scp=85127214828&partnerID=8YFLogxK
U2 - 10.1016/j.conbuildmat.2022.127198
DO - 10.1016/j.conbuildmat.2022.127198
M3 - Article
AN - SCOPUS:85127214828
SN - 0950-0618
VL - 331
JO - Construction and Building Materials
JF - Construction and Building Materials
M1 - 127198
ER -