Integration of local and global surrogates for failure probability estimation

This paper introduces the Global-Local Hybrid Surrogate (GLHS) algorithm, which combines a global surrogate model with adaptive local surrogates in a buffer zone around the limit state surface to efficiently and accurately estimate rare failure probabilities in high-dimensional complex systems.

The Gist

Imagine you are trying to find a tiny, hidden leak in a massive, complex plumbing system (like a skyscraper's water pipes). Your goal is to calculate the exact probability that the building will flood.

In the world of engineering and science, this "leak" is called a failure, and the "plumbing system" is a complex mathematical model of a real-world object (like a bridge, a chemical reactor, or a spacecraft entering an atmosphere).

The Problem: The "Needle in a Haystack" Dilemma

Traditionally, to find this leak, engineers use a method called Monte Carlo Simulation. Imagine throwing a million darts at a giant map of the building to see where they land.

• The Catch: If the leak is very rare (a "rare event"), you might throw a million darts and still miss the leak entirely. To be sure, you'd need to throw billions of darts.
• The Cost: Each "dart throw" requires running a super-complex computer simulation that takes hours or days. Throwing billions of darts would cost more money and time than the building itself is worth.

The Old Solution: The "Rough Sketch"

To save time, engineers started using Surrogate Models. Instead of running the expensive simulation every time, they build a cheap, fast "rough sketch" (a mathematical approximation) of the system based on a few test runs.

• The Flaw: A rough sketch is great for seeing the general shape of the building, but it's terrible at finding the tiny, specific spot where the leak is. If the sketch is slightly wrong near the leak, your probability calculation will be wildly inaccurate.

The New Solution: GLHS (Global-Local Hybrid Surrogate)

This paper introduces a clever new algorithm called GLHS. Think of it as a two-step detective strategy that combines a "bird's-eye view" with a "microscope."

Step 1: The Bird's-Eye View (The Global Surrogate)

First, the algorithm takes a few quick, cheap measurements to build a Global Surrogate.

• Analogy: Imagine you are looking at a map of a forest from a helicopter. You can see the general shape of the trees and the terrain. You can tell roughly where the "danger zone" (the area where a fire might start) is, but you can't see the individual dry leaves on the ground.
• This global map is fast to make but not precise enough to find the exact fire risk.

Step 2: The Magic "Buffer Zone"

The algorithm looks at its rough map and says, "I'm pretty sure the fire risk is somewhere in this specific valley." It draws a circle around that valley, calling it a Buffer Zone.

• The Buffer Zone: This is a safety margin. It's not just the exact line where failure happens; it's a fuzzy area around that line where we need to be extra careful.

Step 3: The Microscope (The Local Surrogate)

Now, instead of checking the whole forest again, the algorithm zooms in only on that Buffer Zone.

• The Trick: It uses a special technique called Christoffel Adaptive Sampling. Imagine a smart robot that knows exactly where to look. Instead of randomly throwing darts in the valley, the robot knows, "The ground is tricky here; I need to check these specific spots to get the best picture."
• It builds a Local Surrogate—a high-definition, zoomed-in 3D model of just that small valley. This local model is incredibly accurate.

Step 4: The Hybrid Result

Finally, the algorithm stitches the two together:

1. It uses the Global Map for 99% of the forest (where nothing bad is happening).
2. It swaps in the High-Def Local Model for the tiny Buffer Zone (where the danger is).

Why This is a Game-Changer

• Efficiency: You don't need to run expensive simulations for the whole system. You only run them for the tiny, critical area where failure might happen.
• Accuracy: Because the "Local Model" is built with smart, targeted sampling, it catches the tiny details the "Global Map" missed.
• Speed: In the paper's tests, this method reduced the error in predicting failure probability from nearly 50% down to less than 1%, using only a tiny fraction of the computer power required by traditional methods.

Real-World Example from the Paper

The authors tested this on a simulation of a spacecraft entering Titan's atmosphere (a moon of Saturn).

• The Challenge: They needed to predict the chance of the heat shield failing due to chemical reactions.
• The Result: The standard "rough sketch" was off by 50%. The new GLHS method, by zooming in on the critical chemical reaction zone, got the answer almost perfectly right, saving massive amounts of computing time.

Summary

GLHS is like hiring a general contractor to look at your whole house (Global) and then calling in a specialized plumber with a microscope to inspect only the one pipe you think is leaking (Local). You get the accuracy of a full inspection without paying for the time and effort of checking every single pipe in the house.

Technical Summary

1. Problem Statement

The paper addresses the challenge of estimating the probability of rare failure events ($P_f$) in complex, high-dimensional engineering systems.

• The Core Issue: Traditional Monte Carlo Simulation (MCS) is accurate but computationally prohibitive for rare events because the required sample size scales as $O(1/P_f)$.
• Limitations of Existing Surrogates: While surrogate modeling (e.g., Polynomial Chaos Expansions - PCE) reduces computational cost, global surrogates often lack the necessary accuracy near the limit state surface (the boundary between safe and failure modes, defined by $g_T(x) = 0$). Small local errors in this critical region can lead to significant biases in the estimated failure probability.
• Limitations of Previous Hybrid Methods: Previous hybrid approaches (e.g., Li and Xiu, 2010/2011) integrate global and local models but often rely on pre-defined, static buffer zones or iterative strategies that may still require excessive model evaluations or suffer from sensitivity to threshold parameters.

2. Methodology: Global-Local Hybrid Surrogate (GLHS)

The authors propose the Global-Local Hybrid Surrogate (GLHS) algorithm, an iterative framework that combines a global surrogate with adaptively constructed local surrogates within a dynamically learned "buffer zone."

Key Components:

1. Global Surrogate ($g_S$):

   • Constructed initially using a small set of high-fidelity samples ($m_0$) via Least Squares PCE.
   • Captures the overall trend of the system but may be inaccurate near the limit state.
   • Used to identify an initial buffer zone ($S_{\eta}$), defined as the region where $|g_S(x)| \leq \eta$.

2. Domain Learning (DL) & Buffer Zone Expansion:

   • The algorithm identifies the buffer zone using the current surrogate.
   • To ensure a rich representation of this zone, it constructs a hyperrectangle (or hyperplanes in 1D) enclosing the identified samples.
   • New samples are generated within this hyperrectangle to "learn" the domain of the buffer zone, expanding the grid to size $m_d \gg 1$.

3. Christoffel Adaptive Sampling (CS):

   • Instead of uniform sampling within the buffer zone, GLHS employs Christoffel Adaptive Sampling.
   • CS selects sample points based on the Christoffel function, which quantifies how well a point is represented by the polynomial basis. It inherently favors regions where the regression problem is ill-conditioned or where more data is needed for convergence.
   • This ensures optimal sample placement for constructing the local surrogate with minimal evaluations.

4. Local Surrogate Construction ($g_L^{(l)}$):

   • A local PCE is constructed using the CS-selected samples and the corresponding high-fidelity evaluations ($g_T$).
   • The polynomial order is determined via cross-validation to balance complexity and accuracy.
   • The local surrogate is designed specifically to refine the approximation within the buffer zone.

5. Hybrid Integration & Iteration:

   • A combined surrogate ($\tilde{g}^{(l)}$) is formed by replacing the global surrogate values with the local surrogate values inside the buffer zone.
   • The threshold $\eta_l$ is updated based on the Mean Squared Error (MSE) of the new hybrid model.
   • The process iterates (Steps 2–4) until convergence (e.g., the buffer zone becomes empty or the failure probability stabilizes).

3. Key Contributions

• Adaptive Buffer Zone Strategy: Unlike static methods, GLHS dynamically learns the buffer zone and expands it using a hyperrectangle approach to ensure the true limit state is captured, addressing the sensitivity to the initial threshold $\eta_0$.
• Integration of Christoffel Sampling: The novel application of Christoffel Adaptive Sampling within the buffer zone allows for the construction of highly accurate local surrogates with a minimal number of expensive model evaluations.
• Hybrid Architecture: The method effectively decouples the problem: the global surrogate handles the bulk of the domain, while local surrogates focus computational resources exclusively on the critical failure transition region.
• Handling Rare Events: The framework is specifically optimized for high-dimensional problems and rare failure probabilities ($P_f < 1\%$), where standard global surrogates fail.

4. Numerical Results

The authors validated GLHS on three test cases:

• 1D Test Case (Analytical Function):

  • Demonstrated that a global surrogate alone had a 6.8% error in failure probability.
  • After one iteration of GLHS (adding only 6 samples), the error dropped to 1.6%.
  • With a slightly more conservative initial buffer zone, the error reached 0.0%.

• 2D Test Case (Analytical Function):

  • Global surrogate error: 3.5%.
  • GLHS (1 iteration, 17 additional samples) reduced error to 0.4%.
  • Comparison: GLHS significantly outperformed the non-iterative Li and Xiu method when both were restricted to the same computational budget (number of model evaluations). GLHS achieved lower variance and higher accuracy.

• 4D Test Case (Chemical Reaction behind a Shock - PLATO):

  • Applied to a complex plasma simulation involving chemical reactions during atmospheric entry.
  • Scenario A ($P_f = 1\%$): Global surrogate error was ~3%; GLHS reduced it to 0.01%.
  • Scenario B ($P_f = 0.1\%$): Global surrogate error spiked to 49.5% (due to the rarity of events). GLHS reduced the error to 0.2% using only 56 additional high-fidelity evaluations.
  • Sensitivity Analysis: Showed that increasing the conservativeness parameter ($c$) expands the buffer zone, slightly increasing variance but ensuring the failure region is fully captured.

5. Significance and Conclusion

• Computational Efficiency: GLHS achieves high accuracy with a fraction of the computational cost required by standard MCS or non-iterative hybrid methods. It concentrates expensive model evaluations exactly where they are needed (near the limit state).
• Robustness for Rare Events: The method is particularly effective for rare failure probabilities ($<1\%$), a regime where global surrogates typically fail due to insufficient resolution of the failure boundary.
• Scalability: The approach successfully handled a 4-dimensional, computationally expensive physics simulation, demonstrating its potential for real-world engineering applications.
• Future Work: The authors suggest extending the method to handle multiple disconnected failure domains (via clustering) and adapting the sampling measures for non-uniform input distributions (e.g., Gaussian) using generalized polynomial bases.

In summary, the GLHS method represents a significant advancement in reliability analysis by intelligently coupling global trend approximation with localized, adaptively sampled high-fidelity modeling, offering a robust solution for rare event estimation in complex systems.