Swelling Potential of Kansas Soils: Modeling and Validation Using Artificial Neural Network Reliability Approach

Abstract

Determination of the swell potential of a troublesome soil can only be made possible through the use of systematic methods for identifying, testing, and evaluating that potential. For proper evaluation of the severity of swelling for such soils, an accurate soil swell potential assessment method is warranted. To address this key issue, a two-phase research study was performed to develop combined artificial neural network and reliability-based soil swell prediction models. In Phase 1, a reasonable-sized database representing 514 swell soil tests retrieved from over 51 different projects in Kansas was used to develop both neural network–based (NNB) and statistical-based (SB) swell potential prediction models. Direct comparison of results obtained showed that NNB models provide significant improvements in prediction accuracy over their SB counterparts. In the second phase, predictions obtained using the developed NNB models along with the available experimental database were used to produce reliability (probability) factor matrices, which are used to assign a specific confidence level to predictions obtained via NNB models in order to classify the soil under consideration as a swelling or nonswelling type.