A Trust-region Funnel Algorithm for Grey Box Optimisation

Gul Hameed, Tao Chen, Antonio del Rio Chanona, Lorenz T. Biegler, Michael Short

Introduction

In process systems engineering, researchers often encounter "grey-box" optimization problems. These problems combine "glass-box" models—explicit algebraic equations with accessible derivatives—with "black-box" components, such as proprietary simulators or computationally intensive CFD models that provide outputs without analytical gradients. Navigating this landscape requires balancing the precision of known physics with the uncertainty of external surrogates.

A recent study by Hameed et al. introduces a novel Trust-region (TR) funnel algorithm designed to simplify these complex problems. Implemented within the open-source Pyomo framework, this method replaces traditional, complex "filter" acceptance criteria with a more intuitive and self-correcting "funnel" mechanism.

The Grey-Box Challenge

Modern chemical plants and energy systems are frequently too complex for a single set of equations. Engineers typically use Reduced Models (RMs) to approximate black-box behavior. The standard approach, the Trust-Region Filter method, maintains a list of past successful points to decide whether to accept a new step. However, filter methods are algorithmically complex, often requiring extensive manual parameter tuning and risking stalls if surrogate model errors become too high.

The Funnel Mechanism

The core innovation of this work is the Funnel Mechanism. Instead of a list of points, the algorithm maintains a single scalar value called the "funnel width." This width serves as a monotonically non-increasing upper bound on the approximation error (infeasibility) of the black-box components.

Funnel condition satisfaction (a) and the logic for a theta-type step (b) which focuses on reducing infeasibility.

Key technical features include:

  • Decoupled Regions: The algorithm utilizes a "sampling region" to build local surrogates and a separate "trust region" to govern solver movement. This allows for larger steps in well-behaved glass-box variables even while black-box surrogates are being refined.
  • Switching Logic: The algorithm intelligently alternates between $f$-type steps (optimizing the objective function) and $\theta$-type steps (reducing infeasibility or error).
  • Open-Source Integration: Built into Pyomo, the algorithm leverages powerful NLP solvers like IPOPT, making it accessible for broader research use.

Step Types and Convergence

The funnel approach is largely problem-independent. Unlike filter methods that require heuristic tuning of multiple parameters, the funnel width initializes itself based on the starting point's infeasibility. As the algorithm progresses, it "squeezes" the error down naturally as it approaches the solution.

Logic for an f-type step (c) and the visual representation of the algorithm's convergence (d) toward an optimal point x*.

The paper provides a formal global convergence proof, ensuring that the algorithm will find a first-order critical point under standard assumptions. This mathematical rigor is supported by the visual convergence of the funnel width toward zero as the iterate approaches the optimal solution $x^*$.

Key Results and Performance

The authors evaluated the TR funnel algorithm against a diverse benchmark suite, including numerical tests (Colville, Himmelblau), engineering design problems (Wing Weight, Welded Beam), and process case studies (Williams-Otto reactor, biomass-based hydrogen production).

  • Efficiency: In numerical tests, the TR funnel required 4 to 135 times fewer external black-box evaluations than traditional filter methods when using lower-fidelity surrogates.
  • Robustness: In problems starting from highly infeasible points, the funnel rapidly restored feasibility before focusing on optimality.
  • Consistency: Across all tests, the funnel method converged to the same or better local optima as filter methods, typically with a smoother convergence profile and fewer rejected steps.

Conclusion and Future Work

While the funnel method significantly reduces the need for parameter tuning and improves efficiency, it still faces the "curse of dimensionality" as the number of black-box inputs increases. Future research will focus on incorporating Hessian-informed methods and more sophisticated sampling update mechanisms to further enhance convergence speed for high-dimensional problems.

Summary updated

Abstract

Abstract Gray‐box optimization, where parts of optimization problems are represented by algebraic models while others are treated as black‐box models lacking analytic derivatives, remains a challenge. Trust‐region (TR) methods provide a robust framework for gray‐box problems through local reduced models (RMs) for black‐box components, but they are complex and require extensive parameter tuning. Motivated by recent advances in funnel‐based convergence theory for nonlinear optimization, we propose a novel TR funnel algorithm for gray‐box optimization, replacing the filter acceptance criterion with a uni‐dimensional funnel, maintaining a monotonically decreasing upper bound on approximation error of local black‐box RMs. A global convergence proof to a first‐order critical point is established. The algorithm, implemented open‐source in Pyomo, supports multiple RM forms and globalization strategies (filter or funnel). Benchmark tests show the TR funnel algorithm achieves comparable and often improved performance relative to the classical TR filter method, thus providing a simpler, effective alternative for gray‐box optimization.
BibTeX
@inproceedings{hameed2026trust,
  author = {Gul Hameed and Tao Chen and Antonio del Rio Chanona and Lorenz T. Biegler and Michael Short},
  title = {{A Trust-region Funnel Algorithm for Grey Box Optimisation}},
  booktitle = {AIChE},
  year = {2026},
  doi = {10.1002/aic.70258},
}
Figure