Artificial Neural Networks (ANNs) are promising computational techniques capable of mapping and capturing many features and sub-features embedded in a large set of data that yield a certain output. A network that has successfully captured the governing relationships between the input and the output can be used as a prediction/simulation tool for cases when the output solution is not available. This learning-based technique proved to be an efficient methodology when dealing with a sufficient amount of data representing a complex structure or phenomenon, especially, when there is a highly nonlinear or complex unrecognized governing relations describing the available data sets. Equally efficient practice is to use ANN methodology to handle poorly-defined patterns that have no explicit set of rules.; In order to effectively demonstrate the potential use of ANNs in geotechnical and earthquake engineering fields, ANN-based learning technique was utilized in this research study as a new computational (numerical) approach to model two challenging geo-engineering applications. In the first application, a back-propagation ANN algorithm was used to develop liquefaction potential assessment models using good-sized SPT- and CPT-liquefaction databases representing various earthquake sites from around the world. ANN-based models were further simplified to produce accurate and easy to use liquefaction assessment potential charts/equations that are capable of assessing liquefaction potential based on given soil-, site- and earthquake-related parameters. Additionally, to effectively illustrate the credibility of the developed ANN-based models and their corresponding charts/equations, assessments obtained using well-known previously developed methods were compared with the ANN-based predictions.; In the second part of this study, the results of various experimental monotonic stress-strain triaxial tests on sandy soil were used to investigate the viability of using recurrent (feed-back) ANN-based models as efficient numerical generators/simulators. Accordingly, the feed-back ANN technique was used to develop models that can effectively characterize/simulate the consolidated drained and undrained stress-strain behavior of Nevada sand. Compression and extension stress-path experimental data generated from various Triaxial-based tests on Nevada sand were used to train, test and validate the desired ANN-based simulation models. All experimental tests were performed on samples with different relative densities ranging between 40% to 60%. and under various initial consolidation pressure values. Comprehensive investigation was conducted to assess the agreement between actual and ANN-based simulated stress-strain responses. Moreover, key issues pertaining to various ANN-based model development strategies are presented in this study. Overall, it is noted that developed feed-back ANN-based models were effective in simulating the drained and undrained stress-strain response behavior of Nevada sand.
展开▼
