Prediction of flowability and strength in controlled low-strength material through regression and oversampling algorithm with deep neural network

Abstract

In urban areas, backfilling voids with complex and narrow shapes necessitates alternative backfill methods and materials, such as controlled low-strength materials (CLSMs), to minimize ground subsidence caused by improper compaction of backfill soils. This study aims to propose a predictive methodology for the mechanical properties of CLSMs using regression analysis and a deep neural network (DNN). CLSM mixtures are prepared with various mixing ratios of calcium sulfoaluminate (CSA) expansive admixture, water, Portland cement, fly ash, sand, silt, and alkali-free accelerator. The flow consistency and compressive strength at 12 hrs and 7 days post-mixing are estimated. The relationships between CLSM mixing ratios and the estimated mechanical properties are established through multiple regression analysis and DNN. The DNN's performance is evaluated, with coefficients of determination being 0.0874, 0.8432, and 0.6826 for flowability, and compressive strength at 12 hrs and 7 days, respectively. To address the low performance, oversampling algorithms like the synthetic minority oversampling technique (SMOTE) and the conditional tabular generative adversarial network (CTGAN) are utilized. Analysis of the oversampled data using SMOTE indicates improved performance, with the coefficients of determination rising to 0.6818, 0.9856, and 0.983 for flowability, and compressive strength at 12 hrs and 7 days, respectively. This study illustrates that the identified correlations may be effectively used to predict flowability and compressive strength based on the mixing ratio.

Keywords

• Controlled low-strength material
• Compressive strength
• Deep neural network
• Flowability
• Oversampling