Reliability and Efficiency of Metamodel for Numerical Back Analysis of Tunnel Excavation

Abstract

During tunnel construction, the ground properties, initially evaluated, are continuously assessed and verified through back analysis. This procedure generally requires many numerical analyses, so a metamodel based on artificial neural networks has been used to reduce the number of analyses. More datasets can be used to create more reliable metamodels. However, there are no established rules regarding the optimum number of datasets for a reliable metamodel. Metamodels predicting the vertical displacement of the tunnel crown using five ground parameters (unit weight (γ), uniaxial compressive strength (UCS), material constant mi, geological strength index (GSI), and coefficient of lateral pressure (K)), with 3, 4, 6, 8, and 10 values per property, were created to confirm the reliability of the metamodel based on the number of datasets in this study. Metamodels using 6 and 8 values for each property showed 5% and 1% mean absolute percent errors, respectively. These numbers of each of the properties would be appropriate for developing the metamodel. Among the five parameters, only the results of the global sensitivity analyses of GSI and K are higher than 0.9. According to these results, it is verified that assessments based only on these parameters are sufficient in the back analysis.

Keywords

• metamodel
• artificial neural networks
• back analysis
• reliability
• tunnel excavation