Constitutive modelling of cemented paste backfill: A data-mining approach. (10th February 2019)

• Record Type: Journal Article
• Authors: Qi, Chongchong; Chen, Qiusong; Fourie, Andy; Tang, Xiaolin; Zhang, Qinli; Dong, Xiangjian; Feng, Yan
• Is Part Of: Construction & building materials. Volume 197(2019)
• Page Start: 262
• Page End: 270
• Publication Date: 2019-02-10
• Subjects: Cemented paste backfill -- Constitutive modelling -- Unconfined compression tests -- Random forest -- Firefly algorithm; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2018.11.142
• Languages: English
• ISSNs: 0950-0618

Abstract

Graphical abstract: Highlights: A novel approach is proposed for constitutive modelling of cemented paste backfill. Random forest was used to model the constitutive relations of CPB. FA was employed to optimize the hyper-parameters of RF. Unconfined compression tests were performed for dataset preparation. The optimum RF model could model the constitutive relations with high accuracy. Abstract: The environmental risks posed by mine tailings suggest the necessity of recycling tailings as cemented paste backfill (CPB) and constitutive modelling is an important step to understand its mechanical stability. In the present work, a novel data-mining approach is proposed for the stress–strain relationship modelling of CPB considering the coupled effect of cement/tailings ratio, solids content and curing time. The proposed approach is based on the random forest (RF) and firefly algorithm (FA), which can operate on large quantities of data for nonlinear and complex relationships modelling. RF was used to model the CPB constitutive relations while FA was used to tune the RF hyper-parameters. Unconfined compression tests were performed for the dataset preparation. The reliability and robustness of the proposed approach, the RF_FA, has been verified against experimental data. Results showed that the hyper-parameters of RF could be efficiently tuned by FA and the optimum hyper-parameters were obtained at the fifth generation. Moreover, the RF_FA possessed excellent prediction capability for the stress–strain relationship modelling (the correlation coefficient values between predicted and experimental stress values were 0.991 on the training set and 0.989 on the testing set). External verifications were further carried out to illustrate the performance of the RF_FA using several statistical criteria recommended in the literature. Consequently, it can be suggested that the RF_FA paves a new way in the constitutive modelling of CPB, which is of great significance for its engineering application.