Solving Geoengineering Problems by Neural Networks

(BridgeArt.net Press Release) - Wood-Ridge, New Jersey, 28 December 2006 - In a continuous effort to solve complex, yet common, geoengineering problems, engineers and researchers are attempting to use new methods and approaches, such as using artificial neural networks. Consider cracking of highway surfaces or forming of pot holes, which cause headaches to many drivers. These two undesirable phenomena may be linked to insufficient stiffness or compaction of soils beneath the roadway surfaces and to insufficient fatigue life of asphalt pavement. Team led by Dr. Najjar of Kansas State University used neural network modelling to better understand the behavior of materials in question.

Adding chemical agent to stabilize problematic highway subgrade soil is a common engineering practice in the United States. Due to the fact that theoretical accomplishments in soil chemical stabilization lag far behind the engineering practice, laboratory testing, which is expensive and time-consuming, is almost always necessary to determine the effectiveness of the soil stabilizer in enhancing engineering properties of the soil. Over the years, large amount of valuable data from laboratory tests on stabilizing different soils with different chemical stabilizers was accumulated in the literature. Efforts to extract the relationships and associations from the existing test data in order to provide guidance for new soil chemical stabilization cases were carried out for many years, however, due to the technology (statistic regression) limitations, reliable models are still not available. In a study led by Dr. Najjar, Artificial Neural Network (ANN) approach to study soil chemical stabilization was introduced. An ANN model to predict the unconfined compression strength of the stabilized soil was built based on the experimental data from stabilizing three representative Kansas embankment soils with five chemical stabilizers. The results showed that the trained ANN model could precisely predict the UCS of stabilized soil. Furthermore, ANN model enables us to study the significance of each input factor, thus providing a powerful tool for optimizing the mixture and construction design.

The fatigue behavior of asphalt concrete is very complicated that a comprehensive fundamental theoretical model is not available. Therefore, a reliable empirical method for predicting fatigue life based on experimental data remains a desirable approach. However, the complexity of the fatigue process and the noise associated with the fatigue test results make even the traditional empirical methods, such as regression analysis, handicapped in producing a sufficiently accurate model. Artificial neural networks have the ability to derive considerable complex relationships and associations from experimental data while filtering out the effect of noisy data. In this study, the potential use of ANNs for fatigue life prediction was explored and the comparisons between ANN-based model predictions and predictions via multi-linear as well as other published models showed that ANN-based models provide much more accurate predictions.

Detailed results of two studies are provided in two papers, which will be presented in a poster session on the Transportation Research Board (TRB) 2007 annual meeting.


The study involved development of several programs in MS Excel that illustrate the practical use of the developed neural networks. To request a copy of the software, please use the contact below.


Press Contact:
Chune Huang
email: info (at) bridgeart (dot) net


References

  • Chune Huang, Yacoub M. Najjar, and Stefan A. Romanoschi: Development of a Soil Chemical Stabilization Model Using Artificial Neural Network Approach. Proceedings of 86th Annual Meeting of Transportation Research Board. January 2007.
  • Chune Huang, Yacoub M. Najjar, and Stefan A. Romanoschi: Predicting Asphalt Concrete Fatigue Life Using Artificial Neural Network Approach. Proceedings of 86th Annual Meeting of Transportation Research Board. January 2007.
Article Type: 
Channel: 
Keywords: