Enhanced active learning Gaussian process metamodel for estimating the one-sided tail probability of nonlinear...

Area of Science:

• Structural Engineering
• Computational Mechanics
• Risk Assessment

Background:

• Mega-structures face safety challenges from rare, high-impact failures.
• Accurate estimation of low-probability structural response is difficult.
• Failures in underground structures, like subway tunnels, can result from insufficient anchorage length.

Purpose of the Study:

• To develop a novel framework for accurately estimating rare event probabilities in structural engineering.
• To improve the estimation of one-sided tail probabilities for structural response distributions.
• To provide a practical tool for uncertainty analysis in critical infrastructure safety.

Main Methods:

• Introduction of the Tail-Sensitive Global Learning (TS-GL) algorithm.
• TS-GL features a tail-focused search mechanism and a new weight function.
• Investigation of activation functions for computational efficiency.

Main Results:

• TS-GL significantly improves the estimation of one-sided tail probabilities compared to existing methods.
• The algorithm was validated on the bond-slip relationship in steel-concrete connections.
• TS-GL demonstrated superior accuracy and efficiency over active learning-based Gaussian process (AL-GP) metamodels for rare event quantification.

Conclusions:

• The TS-GL algorithm offers a practical and effective solution for uncertainty analysis in critical infrastructure.
• This novel framework enhances the safety operation and maintenance of mega-structures.
• Improved estimation of rare event probabilities is crucial for preventing structural failures.