Geotechnical Assessment of selected lateritic soils in southwest Nigeria for road construction and development of artificial neural network mathematical based model for prediction of the California bearing ratio

Investigation of the geotechnical characteristics of eighteen different lateritic soils within southwestern Nigeria was carried out to determine their suitability for road construction. To achieve this goal, the lateritic soils samples were subjected to different laboratory tests, including specific gravity, Atterberg limits, grain size analysis, California bearing ratio, and compaction, in accordance with the ASTM standard procedure. The results of the tests showed that the specific gravity varies between 2.55 and 2.81; the linear shrinkage varies between 6.68% and 10.98%; the liquid limit varies between 37.17% and 56.93%; the plastic limit ranges from 19.47% to 37.14%; the plasticity index ranges from 3.81% to 30.29%; the fine sand content ranges from 37.07% to 62..93%; the fines content ranges from 36.4% and 60.9%; the maximum dry density ranges from 1747 kg/m3 to 2056 kg/m3; the optimum moisture content ranges from 10.94% to 20.51%; the un-soaked California bearing ratio ranges from 14.7% to 45.6%; and the soaked California bearing ratio ranges from 10% to 31%. Based on these results, all the studied soil can be used as road sub-grade while none is suitable for road sub-base except that of Loc.5. However, none of the soil meets up with the requirement as road base course. The suitability of laterite for the construction of road depends largely on the California bearing ratio. However, laboratory tests for determining the California bearing ratio is tedious, time consuming and costly. As a result of this difficulty, there is a need to develop soft computing models to predict laterite California bearing ratio from index properties with cheap and simple tests. Thus, the experimental datasets of the eighteen studied lateritic soils were used to create and train artificial neural network (ANN) models to predict California bearing ratio from liquid limits, plasticity index, linear shrinkage, fine sand content and fines content. The proposed ANN models were compared with regression-based models proposed in this study and various regression based models from the literature via statistical analyses. Based on the model comparison, the proposed ANN models outperformed the rest of the models, as they presented the highest R2 and the lowest RMSE, MAPE and MAE values. Thus, the ANN models are validated. To enhance the practical applications of the proposed ANN models, they were transformed into simple mathematical equations, which gave the same predictions as the direct ANN models. Thus, they can be used for practical purposes.