Mechanical properties of cemented tailings and waste-rock backfill (CTWB) materials: Laboratory tests and deep learning modeling

• Published in: Construction & building materials 2023-03, Vol.369, p.130610, Article 130610
• Main Authors: Yin, Shenghua, Yan, Zepeng, Chen, Xun, Yan, Rongfu, Chen, Dapeng, Chen, Junwei
• Format: Article
• Language: English
• Subjects: Cemented tailings and waste-rock backfill; Deep learning modeling; Long short-term memory network; Strength; Waste-rock content
• ISSN: 0950-0618; 1879-0526
• DOI: 10.1016/j.conbuildmat.2023.130610

•A model based on GA and LSTM was proposed to predict the CTWB strength.•Comparisons are made with GA-SVM, LSTM, and PSO-LSTM to verify the prediction model's robustness.•UCS test showed that the strength increases first and then decreases with the waste-rock content.•Grey relational analysis was used to determine the influence weight of each factor on the strength. Cemented tailings and waste-rock backfill (CTWB) is an effective way to solve the problem of mine solid waste. Waste-rock content is an important factor affecting UCS, but fully understanding the mechanism of waste-rock action on UCS would require extensive indoor testing, which would be tedious and expensive. Therefore, the mechanical properties of CTWB were investigated using a combination of laboratory testing and deep learning. Long short-term memory (LSTM) network prediction model based on genetic algorithm (GA) optimization was developed with cement content (290 ∼ 330 kg/m3), solid content (75 ∼ 81%), waste-rock content (40 ∼ 70%), and curing age as input variables and unconfined compressive strength (UCS) as output variables. Among them, the population size of GA was set to 40, and the crossover rate and mutation rate were set to 0.75 and 0.01, respectively. The study evaluated the models using RMSE, R2, and MAE and plotted normalized Taylor diagrams for different prediction models. The results show that the GA algorithm is effective for the parameter rectification of the LSTM model. The representative GA-LSTM model has an R2 of 0.9956 in the training set and an R2 of 0.9763 in the test set, indicating a good prediction effect. Deep learning models exhibit higher accuracy than machine learning models. Compared with GA-SVM, LSTM, and PSO-LSTM models, R2 increases from 0.9483, 0.8474, and 0.9611 to 0.9968, respectively, indicating that the model has good robustness and generalization ability. Moreover, the correlation analysis shows that the UCS reaches the maximum value when the waste rock content is 50%, and the hydration products can effectively fill the pores and form a more stable internal structure. The research results will provide technical support for safe, clean, and efficient recovery.