A hierarchical neural hybrid method for failure probability estimation

Ke Li, Kejun Tang, Jinglai Li, Tianfan Wu, Qifeng LiaoIEEE Access 2019 · OpenReview

TL;DR

This paper introduces a novel Hierarchical Neural Hybrid (HNH) method to efficiently estimate failure probabilities in complex, high-dimensional systems governed by partial differential equations (PDEs). By using a hierarchy of neural network surrogates with varying fidelity, HNH significantly reduces the computational cost compared to traditional Monte Carlo methods and existing hybrid approaches, achieving high accuracy with far fewer PDE solves.

What Problem Does This Paper Address?

Estimating the probability of failure in complex engineering systems, such as those described by partial differential equations (PDEs), is crucial for risk management. However, these systems often involve numerous uncertain parameters, making direct simulation computationally prohibitive. Traditional Monte Carlo (MC) methods require solving the expensive PDE model many times, which becomes infeasible for rare failure events. Surrogate-based methods, which build cheaper approximations of the PDE model, are often hampered by the "curse of dimensionality," meaning their performance degrades rapidly with an increasing number of uncertain parameters. This paper aims to bridge this gap by developing a method that is both accurate for rare events and efficient in high-dimensional settings.

Methods

The core of the proposed method is the Hierarchical Neural Hybrid (HNH) approach. It combines the strengths of:

  • Neural Networks: Used as universal function approximators to build surrogate models for the complex PDE. The paper notes that deeper neural networks generally offer higher fidelity but are more computationally expensive.
  • Hybrid Method: This approach, building on ideas from response surface methods and Monte Carlo sampling, uses a surrogate model to estimate failure probability. It defines a "suspicious domain" around the limit state function's zero crossing, where more refined estimation is needed.
  • Multifidelity Surrogates: Instead of relying on a single, expensive, high-fidelity neural network surrogate, HNH constructs a hierarchy of neural networks with varying depths (and thus varying fidelities and costs). Coarser fidelity surrogates (shallower networks) are used for initial estimations, and more expensive, fine-fidelity surrogates (deeper networks) are selectively employed only in the critical "suspicious domain." This adaptive approach significantly reduces overall computational cost.

The HNH method involves an iterative process:

  1. Hierarchical Training: Multiple neural network surrogates are trained with different numbers of layers (fidelity).
  2. Modification Procedure: Data is sorted based on the output of the coarsest fidelity surrogate. Predictions are then "modified" using progressively higher fidelity surrogates in the suspicious regions.
  3. Iterative Hybrid Estimation: The modified predictions are used within an iterative hybrid method to refine the failure probability estimate, avoiding the need to pre-define a difficult threshold parameter.

Datasets

The paper does not introduce new benchmark datasets. Instead, it evaluates its method on three established problems:

  • A high-dimensional multivariate benchmark problem from structural safety.
  • A diffusion equation problem.
  • A Helmholtz equation problem.

For training the neural network surrogates, the paper mentions using 10³ training samples for the numerical experiments. The evaluation of the reference solutions often involves a large number of samples (e.g., 5 × 10⁶ for Monte Carlo).

Key Insights and Ideas

  • Adaptive Fidelity is Key: The "curse of dimensionality" is tackled not by making a single surrogate infinitely accurate, but by intelligently using a hierarchy of surrogates. Cheaper, less accurate models handle the bulk of the problem, while expensive, accurate models are reserved for the critical regions.
  • Neural Networks as Universal Surrogates: The paper leverages the proven ability of neural networks to approximate complex functions, making them suitable for a wide range of PDE-based systems.
  • Reduced PDE Solves: The primary benefit is a drastic reduction in the number of times the original, expensive PDE model needs to be solved. This is the bottleneck for traditional MC methods.
  • Theoretical Guarantees: The paper provides theoretical analysis to show that the HNH method can achieve an error bound (ε) for the failure probability estimation.

Key Results

The HNH method demonstrates significant improvements in both accuracy and efficiency:

  • Efficiency: For estimating rare failure probabilities (e.g., 10⁻⁵), traditional Monte Carlo methods might require over a million PDE solves. The HNH method achieves comparable accuracy with only a few thousand solves.
  • Speedup: In numerical experiments, HNH consistently outperforms a standard neural hybrid (NH) method in terms of runtime, especially as the number of samples increases. For instance, in the Helmholtz problem with 10⁶ samples, HNH was approximately four times faster than NH.
  • Accuracy: The method accurately captures failure probabilities, often with significantly lower relative errors than MC, even with a much smaller number of samples. For example, in the diffusion equation test, HNH achieved a relative error of 4.17% with 2 × 10³ samples, compared to MC with 10⁶ samples.

Limitations and Future Work

  • Tuning Parameters: While the HNH method aims to reduce the need for manual parameter tuning (like the threshold γ in traditional hybrid methods), some parameters related to the hierarchy structure (number of layers, number of fidelity levels) and the modification process still require careful selection.
  • Computational Cost of Training: Training deep neural networks can still be computationally intensive, although it's a one-time cost for constructing the surrogates.
  • Future Work: The authors suggest that the universal nature of neural networks means the HNH method can be extended to other surrogate modeling problems beyond failure probability estimation. Further exploration of optimal hierarchy structures and adaptive refinement strategies could also be beneficial.

Summary updated

Abstract

Failure probability evaluation for complex physical and engineering systems governed by partial differential equations (PDEs) are computationally intensive, especially when high-dimensional random parameters are involved. Since standard numerical schemes for solving these complex PDEs are expensive, traditional Monte Carlo methods which require repeatedly solving PDEs are infeasible. Alternative approaches which are typically the surrogate based methods suffer from the so-called ``curse of dimensionality'', which limits their application to problems with high-dimensional parameters. For this purpose, we develop a novel hierarchical neural hybrid (HNH) method to efficiently compute failure probabilities of these challenging high-dimensional problems. Especially, multifidelity surrogates are constructed based on neural networks with different levels of layers, such that expensive highfidelity surrogates are adapted only when the parameters are in the suspicious domain. The efficiency of our new HNH method is theoretically analyzed and is demonstrated with numerical experiments. From numerical results, we show that to achieve an accuracy in estimating the rare failure probability (e.g., $10^{-5}$), the traditional Monte Carlo method needs to solve PDEs more than a million times, while our HNH only requires solving them a few thousand times.
BibTeX
@inproceedings{li2019hierarchical,
  author = {Ke Li and Kejun Tang and Jinglai Li and Tianfan Wu and Qifeng Liao},
  title = {{A hierarchical neural hybrid method for failure probability estimation}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=2jTR4KWOFXd},
}
Figure