.png){kind=icon}
.png){kind=icon}
Return
Reliability-informed inverse design of dual tunnels with deep evidential regression
.png){kind=icon}
PRE
AI
H
T
T
DOI：10.1016/j.ress.2025.112134.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This study presents an integrated, data-driven framework for evaluating the stability and performing inverse design of dual tunnel systems in cohesive–frictional soils subjected to uniform surcharge loading. The approach couples finite element limit analysis, Deep Evidential Regression, and reliability-informed optimization to deliver interpretable, probabilistic, and cost-efficient tunnel designs. A dataset of 9306 FELA simulations covering six tunnel types and broad geotechnical–geometric ranges was used to train a Deep Evidential Regression model optimized via a tree-structured Parzen estimator. The model achieved high predictive accuracy (R² = 0.989, RMSE = 0.552) and well-calibrated uncertainty (mean and standard deviation of standardized residuals: −0.036 and 1.144, respectively). Global sensitivity analysis and model interpretability through Shapley Additive Explanations and Partial Dependence Plots revealed γD/c’, S/D, H/D, and φ’ as key factors, reflecting arching, confinement, and tunnel–tunnel interaction mechanisms, and delineated optimal stability ranges for symmetric layouts. Inverse optimization over 48,000 configurations per tunnel type under dual thresholds (σₛ/c’ ≥ 2, 5; P ≥ 95 %) produced reliability-consistent solutions balancing performance and economy. Cross-validation using independently generated datasets confirmed the model’s predictive reliability, accurately reproducing central stability trends and calibrated uncertainty, thereby reinforcing its physical consistency and applicability to reliability-informed tunnel design.
Journal
R
IF:
11
Papers:
9.0K
Citations:
4.2W
Organization
Cited Papers