Unconfined compressive strength of bio-cemented sand: state-of-the-art review and MEP-MC-based model development

As a clean and sustainable method, the microbially induced calcite precipitation (MICP) approach has been widely used for reinforcing weak soils. This study presents a state-of-the-art review on the unconfined compressive strength (UCS) of bio-cemented sand treated by MICP, followed by the high-performance prediction using a machine learning algorithm combined with the Monte-Carlo (MC) method. First, various influencing parameters affecting the UCS of bio-cemented sand are identified, such as initial relative density, angularity of particle shape, bacterial concentration, precipitated calcium carbonate content, temperature and degree of saturation. Besides, the particle size distribution, urea and calcium concentration, and initial pH level also influence the UCS of the bio-cemented sand, but the effects remain contradictory or unclear. Following the state-ofthe-art review, a large database covering 351 bio-cemented sand samples is developed, with the UCS as the output and seven influencing parameters (median grain size, coefficient of uniformity, initial void ratio, optical density of bacterial suspension, urea concentration, calcium concentration and precipitated calcium carbonate content) as inputs for the correlation. The multi expression programming (MEP) method combined with the MC method is proposed to develop the prediction models. All data groups randomly generated from the database are with 80% of the samples as the training sets and 20% as the testing sets. Finally, the optimal prediction model is selected with the lowest mean absolute error, further based on the analyses of monotonicity, sensitivity and robustness regarding more general applications.

Automated summary

The study reviewed the influence of MICP on the UCS of bio-cemented sand and made high-performance predictions using machine learning algorithm and MC method. A large database was developed with MEP method proposed for developing prediction models with the lowest error.