Introduction
In modern engineering, optimization extends beyond finding the highest-performing design; it must also ensure that designs are robust against real-world uncertainties, such as manufacturing tolerances or environmental variability. This field, known as Reliability-Based Design Optimization (RBDO), often relies on computationally expensive simulations like Computational Fluid Dynamics (CFD). This paper introduces EFISUR (Expected Feasible Improvement with Stepwise Uncertainty Reduction), a novel Bayesian optimization framework designed to find reliable optima with a minimal number of costly simulations.
The Challenge: The "Double Loop" Problem
Traditional RBDO faces a significant hurdle: the "double loop" problem. To verify if a design meets a reliability threshold (e.g., a 95% probability of satisfying constraints), one must evaluate that design under numerous uncertainty scenarios. When each evaluation takes hours or days, performing this check for every candidate design becomes computationally prohibitive. While standard Bayesian Optimization (BO) often treats uncertainty as uncontrollable noise, EFISUR leverages the fact that researchers can often choose which specific uncertainty parameters ($u$) to simulate alongside design variables ($x$).
The EFISUR Framework
EFISUR addresses the computational bottleneck through a two-step acquisition strategy that operates in the joint space of design variables and uncertain parameters.
1. Joint Space Modeling
Instead of modeling the design space alone, EFISUR constructs Gaussian Process (GP) surrogates over the combined space of $x$ and $u$. This allows the model to learn how specific uncertainties affect performance across different designs.
2. Two-Step Acquisition
The algorithm proceeds in two distinct phases for each iteration:
- Target Selection ($x_{targ}$): The algorithm identifies the most promising design point by maximizing the Expected Feasible Improvement (EFI). This criterion balances the search for better objective values with the probability of satisfying "chance constraints."
- Uncertainty Sampling ($u_{t+1}$): Once a target design is selected, the algorithm determines which specific uncertainty scenario to simulate. It uses a Stepwise Uncertainty Reduction (SUR) criterion ($S$) to select the $u$ that most reduces the variance of the feasible improvement at the target point.
By providing closed-form analytical expressions for these criteria, the authors ensure that the internal optimization of the acquisition function remains efficient.
Experimental Validation
The effectiveness of EFISUR was demonstrated on two primary benchmarks:
- Analytical Test Case: A 4D problem (2D design, 2D uncertainty) with non-linear constraints, used to verify convergence against a known optimum.
- NASA Rotor 37: A high-fidelity aerodynamic optimization of a compressor blade. This industrial-scale challenge involved 20 design variables and 7 uncertain parameters (27D total) with five constraints.
Results and Impact
EFISUR demonstrated significant advantages over existing methods like random uncertainty sampling (EFIrand) or quantile-based approaches (cEIdevNum):
- High-Dimensional Efficiency: In the 27-dimensional NASA Rotor case, EFISUR identified a superior feasible design in only 237 total simulations, a remarkably low budget for such a complex space.
- Reliability and Convergence: The method converged faster and with lower variance than competitors. Crucially, the designs proposed by EFISUR were more likely to satisfy actual "a posteriori" reliability constraints, effectively mapping the failure boundaries.
- Strategic Probing: The algorithm successfully "breaks symmetry" by ignoring uncertainty scenarios that are clearly safe or clearly failing, focusing its budget on the "edge of knowledge" where information gain is highest.
Limitations and Future Directions
While EFISUR represents a major step forward, several areas for improvement remain. The current model assumes independence between the objective and constraints; incorporating their physical correlations could further enhance efficiency. Additionally, as the number of data points grows, the $O(t^3)$ scaling of Gaussian Process operations becomes a bottleneck. Future work may also explore constraint prioritization to further reduce simulation costs by only evaluating the most critical constraints for a given pair.
Conclusion
EFISUR provides a mathematically rigorous and computationally efficient path for RBDO in high-dimensional engineering problems. By treating uncertainty as a controllable input and strategically sampling the joint space, it enables the discovery of reliable, high-performance designs that were previously out of reach for expensive simulation budgets.
Summary updated