Date of Award
Doctor of Philosophy
School of Engineering
Associate Professor Sanjay Kumar Shukla
Dr Themelina Paraskeva
During the past five decades, numerous studies have been conducted to investigate the load-settlement behaviour of geosynthetic-reinforced soil. The main advantage of reinforced soil foundations are the increase in the bearing capacity and decrease in the settlement. Whereas, for pavement foundation design, the strength of the subgrade soil is often measured in terms of California bearing ratio (CBR). The researchers have suggested various methods to improve the quality of geosynthetic-reinforced foundations soils. In the recent past, the wraparound geosynthetic reinforcement technique has been proposed to strengthen the foundation soil effectively. However, there are several research gaps in the area; for example, there has been no analytical solution for estimating the ultimate bearing capacity of wraparound reinforced foundations, and there has been no evaluation of this technique under repeated loading conditions. Similarly, for planar geosynthetic-reinforced soil foundations, the prediction of load-settlement behaviour also requires more attention. The advent of artificial intelligence (AI) based modelling techniques has made many traditional approaches antiquated. Despite this, there is limited research on using AI techniques to derive mathematical expressions for predicting the load-settlement behaviour of reinforced soil foundations or the strength of reinforced subgrade soil.
This research is undertaken to examine the load-settlement behaviour of geosynthetic-reinforced foundation soils using experimental, analytical, and intelligent modelling methods. For this purpose, extensive laboratory measurements, analytical, numerical and AI-based modelling and analysis have been conducted to: (i) derive theoretical expression to estimate the ultimate bearing capacity of footing resting on soil bed strengthened by wraparound reinforcement technique; (ii) using detail experimental study, present the effectiveness of wraparound reinforcement for improving the load-settlement characteristics of sandy soil under repeated loading conditions; (iii) to build the executable artificial intelligence-based or computationally intelligent soft computing models and converting them into simple mathematical equations for estimating the (a) ultimate bearing capacity of reinforced soil foundations; (b) settlement at peak footing loads; (c) strength (California bearing ratio) of geosynthetic-reinforced subgrade soil; and (iv) to examine and predict the settlement of geosynthetic-reinforced soil foundations (GRSF) under service loading condition using novel hybrid approach, that is, finite element modelling (FEM) and AI modelling.
In the analytical phase, a theoretical expression has been developed for estimating the ultimate bearing capacity of strip footing resting on soil bed reinforced with the geosynthetic layer having the wraparound ends. The wraparound ends of the geosynthetic reinforcement are considered to provide the shearing resistance at the soil-geosynthetic interface as well as the passive resistance due to confinement of soil by the geosynthetic reinforcement. The values of ultimate load-bearing capacity determined by using the developed analytical expression have predicted values closer to the model studies reported in the literature, with a difference in the range of 0% to 25% with an average difference of 10%.
In the experimental phase, model footing load tests have been conducted on strip footing resting on a sandy soil bed reinforced with geosynthetic in wraparound and planar forms under monotonic and repeated loadings. The geosynthetic layers were laid according to the reinforcement ratio to minimise the scale effect. The effect of repeated load amplitude and the number of cycles, and the effect of reinforcement parameters, such as number of layers, reinforcement width, lap-length ratio and planar width of wraparound, were investigated, and their potential effect on the load-settlement behaviour has been studied. The wraparound reinforced model has shown about 45% lower average total settlement than the unreinforced model. In comparison, the double-layer reinforced model has about 41% at the cost of twice the material and 1.5 times the occupied land width ratio. Additionally, for lower settlement levels (s/B ≤ 5%), the wraparound geotextile with a smaller occupied land width ratio (bp/B = 3.5) has performed well in comparison to the wraparound with a slightly larger occupied land width ratio (bp/B = 4). However, the wraparound with occupied width ratio of 4 provides more stability to the foundation soil for higher settlement levels. The performance of the fully wrapped model (bp/B = 2.8) is more similar to that of the planar double-layer reinforced model (b/B = 4); however, it is noted that even the fully wrapped model outperforms the planar single-layer reinforced model with the same amount of geotextile and 50% less occupied land width
For data analytic methods, first historical data has been collected to build the various machine learning (ML) models, and then detailed comparison has been presented among the ML-based models and with other available theoretical methods. A comprehensive study was conducted for each model to choose its structure, optimisation, and tuning of hyperparameters and its interpretation in the form of mathematical expressions. The forecasting strength of the models was assessed through a cross-validation approach, rigorous statistical testing, multi-criteria approach, and external validation process. The traditional statistical indices such as coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percent deviation (MAPD); along with several other modern model performance indicators, were utilised to evaluate the accuracy of the developed models. For ultimate bearing capacity (UBC) estimation, the performance of the extreme learning machine (ELM) and TreeNet models has shown a good degree of prediction accuracy in comparison with traditional methods over the test dataset. However, the overall performance of the ELM model (R2 = 0.9586, MAPD=12.8%) was better than that of the TreeNet model (R2 = 0.9147, MAPD =17.2%). Similarly, for settlement estimation at peak footing loads, multivariate adaptive regression splines (MARS) modelling technique has outperformed (R2 = 0.974, RMSE = 1.19 mm, and MAPD = 7.19%) several other robust AI-based models, namely ELM, support vector regression (SVR), Gaussian process regression (GPR), and stochastic gradient boosting trees (SGBT). For CBR, the competency and reliability of the several intelligent models such as artificial neural network (ANN), least median of squares regression (LMSR), GPR, elastic net regularization regression (ENRR), lazy K-star (LKS), M5 model trees, alternating model trees (AMT), and random forest (RF). Among all the intelligent modelling techniques, ANN (R2 = 0.944, RMSE = 1.74, and MAE = 1.27) and LKS (R2 = 0.955, RMSE = 1.52, and MAE = 1.04) has achieved the highest ranking score of 35 and 40, respectively, in predicting the CBR of geosynthetic-reinforced soil. Moreover, for UBC and settlement at peak footing loads, new model footing load tests, and for strength of reinforced subgrade soil, new CBR tests were also conducted to verify the predictive veracity of the developed AI-based models.
For predicting the settlement behaviour of GRSF under various service loads, an integrated numerical-artificial intelligence approach was utilised. First, the large-scale footing load tests were simulated using the FEM technique. At the second stage, a detailed parametric study was conducted to find the effect of footing-, geosynthetic- and soil strength- parameters on the settlement of GRSF under various service loads. Afterward, a novel evolutionary artificial intelligence model, that is, grey-wolf optimised artificial neural network (ANN-GWO), was developed and translated to the simple mathematical equation for estimating the load-settlement behaviour of GRSF. The results of this study indicate that the proposed ANN-GWO model predict the settlement of GRSF with high accuracy for training (RMSE = 0.472 mm, MAE = 0.833, R2 = 0.982), and testing (RMSE = 0.612 mm, MAE = 0.363, R2 = 0.962,) dataset. Furthermore, the predictive veracity of the model was verified by detailed and rigorous statistical testing and against several independent scientific studies as reported in the literature.
This work is practically valuable for understanding and predicting the load-settlement behaviour of reinforced soil foundations and applies to traditional planar geosynthetic-reinforced and as well as recently developed wraparound geosynthetic-reinforced foundation soil technique. For wraparound reinforced soil foundations, the analytical expression will be helpful in the estimation of ultimate bearing capacity, and experimental study shows the beneficial effects of such foundations systems in terms of enhancement in bearing capacity and reduction in the settlement, and economic benefits in terms of saving land area and amount of geosynthetic, under repeated loading conditions. Moreover, the developed AI-based models and mathematical expressions will be helpful for the practitioners in predicting the strength and settlement of reinforced soil in an effective and intelligent way and will be beneficial in the broader understanding of embedding the intelligent modelling techniques with geosynthetic-reinforced soil (GRS) technology for the automation in construction projects.
Raja, M. N. A. (2021). Load-settlement investigation of geosynthetic-reinforced soil using experimental, analytical, and intelligent modelling techniques. https://ro.ecu.edu.au/theses/2455
Available for download on Wednesday, October 19, 2022