TY - JOUR
T1 - Logic-guided neural network for predicting steel-concrete interfacial behaviors
AU - Mahjoubi, Soroush
AU - Meng, Weina
AU - Bao, Yi
N1 - Publisher Copyright:
© 2022 Elsevier Ltd
PY - 2022/7/15
Y1 - 2022/7/15
N2 - The interfacial behaviors play significant roles in various composite materials and structures. This paper presents a logic-guided neural network to seamlessly integrate data-driven methods and scientific knowledge in predicting interfacial properties of steel-concrete composites. The investigated properties include the bond strength, interface slip, and bond-slip relationship. Three methods are proposed to conform to logic and scientific principles: (1) logic data are generated to supplement experimental data; (2) a logic loss function is presented to guide the learning process; and (3) unstructured data and incomplete data are utilized to enlarge the dataset. The performance of the presented method is compared with four representative machine learning methods, which are artificial neural network, tree boosting, random forest, and epsilon-support vector regression. The results indicated that the proposed method achieved the highest accuracy. The coefficients of determination of the bond stress and compressive strength are higher than 0.95, and the predicted bond-slip behaviors conform to prior knowledge. The proposed method is useful for prediction of properties for composite materials and structures.
AB - The interfacial behaviors play significant roles in various composite materials and structures. This paper presents a logic-guided neural network to seamlessly integrate data-driven methods and scientific knowledge in predicting interfacial properties of steel-concrete composites. The investigated properties include the bond strength, interface slip, and bond-slip relationship. Three methods are proposed to conform to logic and scientific principles: (1) logic data are generated to supplement experimental data; (2) a logic loss function is presented to guide the learning process; and (3) unstructured data and incomplete data are utilized to enlarge the dataset. The performance of the presented method is compared with four representative machine learning methods, which are artificial neural network, tree boosting, random forest, and epsilon-support vector regression. The results indicated that the proposed method achieved the highest accuracy. The coefficients of determination of the bond stress and compressive strength are higher than 0.95, and the predicted bond-slip behaviors conform to prior knowledge. The proposed method is useful for prediction of properties for composite materials and structures.
KW - Deep learning
KW - Incomplete data
KW - Interfacial properties
KW - Logic-guided neural network
KW - Machine learning
KW - Unstructured data
UR - http://www.scopus.com/inward/record.url?scp=85126021074&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85126021074&partnerID=8YFLogxK
U2 - 10.1016/j.eswa.2022.116820
DO - 10.1016/j.eswa.2022.116820
M3 - Article
AN - SCOPUS:85126021074
SN - 0957-4174
VL - 198
JO - Expert Systems with Applications
JF - Expert Systems with Applications
M1 - 116820
ER -