Comparison of artificial neural network (ANN) and linear regression modeling with residual errors to predict the unconfined compressive strength and compression index for Erbil City soils, Kurdistan-Iraq

Abstract

The focus of this research is to develop models for predicting the unconfined compressive strength and compression index (Cc) of clay soils according to Atterberg limits, natural water content, dry density, void ratio, and fine content. The soil unconfined compressive strength (UCS) ranged from 24 to 340 kPa and has been quantified with great precision using tested data and data collected from published research studies. The soil compression index ranged from 0.014 to 0.831 which is also correlated by the easily measurable soil properties of Atterberg limits, natural moisture content, density, void ratio, and fine contents (passes no. 200 sieves). A wide experimental test results (a total of 253 tested soils) were combined with more than 350 data collected from different academic research studies, and total data were statistically analyzed and modeled using a linear regression model (LR). According to the correlation determination (R2), mean absolute error (MAE), and the root mean square error (RMSE), the unconfined compressive strength and compression index of soil can be well predicted in terms of liquid limit (LL), plasticity index (PI), moisture content (MC), dry density (γd), void ratio, and percentage passing no. 200 sieve using linear simulation techniques. The performance of the LR models was compared using artificial neural networks (ANN) models with different hidden layers and input layers. Based on the statistical assessments such as R2, RMSE, and MAE the LR models were close to the ANN model to predict the UCS and Cc of the soils. The sensitivity study concludes that the dry density and void ratio are the most influential parameter in determining the unconfined compressive strength and compression index with the training dataset.