Date of Completion
Honors College Thesis
College of Engineering and Mathematical Sciences
Artificial Neural Network, Backpropagation, ANN, BPNN, Residual Friction Angle, Predict
A complex nonlinear relationship exists between indicative soil parameters (liquid limit, plastic limit, clay fraction, sand fraction, and normal stress loading) and drained secant residual friction angle (ør’). Academic literature provides various empirical models for the prediction of ør’, though their predictions generally suffer from insufficient representation of the nonlinear relationship between parameters. In this study, an artificial neural network was developed for the prediction of ør’. Artificial neural networks are computational learning algorithms that derive their structure and learning procedures from phenomena observed in biological nervous systems. Their complex, interconnected structures allow for successful mapping of nonlinear relationships between parameters. This motivated the development and application of a network known as a backpropagation neural network (BPNN) as an alternative predictive model for ør’.
The BPNN was trained to successfully map the relationship between indicative soil parameters and ør’ using a variety of soil datasets provided from academic literature and other sources. The BPNN’s performance was evaluated using a normalized root mean square error (RMSE) term. It was posited that the BPNN predictions could be improved by training individual networks on soil data subdivided by clay fraction ranges. Analysis showed that the division of data into subsets significantly reduced the BPNN’s predictive performance by limiting the amount of data available for individual network training.
Where other predictive models generally neglect sand fraction as a predictive parameter for ør’, this study attempted to evaluate its predictive value. Comparison between a BPNN that included sand fraction and one that did not proved inconclusive as the results from multiple RMSE analyses between the two models were not statistically different.
Correlation-based equations for the prediction of ør’ in Stark and Hussain (2013) are based on the subdivision of soil data by clay fraction ranges. A comparison of the BPNN predictive model to other empirical models was performed in order to evaluate the viability of a BPNN as an alternative to current predictive models. The BPNN outperformed a traditional, multivariate least-squares linear regression model as well as correlation-based equations from Stark and Hussain (2013). The normalized root mean square error for Stark and Hussain equation-based predictions of ør’ was 0.2270 in comparison to 0.1278 for the BPNN. Where Stark and Hussain’s equations performed well on the particular dataset used to create these empirical equations, it failed to accurately predict ør’ for other data sets provided in the literature. The BPNN model was more robust in its ability to predict ør’ over a variety of datasets in this study and suggests that the BPNN may provide a viable alternative to other predictive models for ør’.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License.
Detwiler, Luke, "Using Backpropagation Neural Networks for the Prediction of Residual Shear Strength of Cohesive Soils" (2014). UVM Honors College Senior Theses. 12.