Inside a surrogate-assisted optimization loop

2025-03-26T14:24:00+00:00

In engineering design, evaluating a candidate can mean running an expensive electromagnetic simulation. A search algorithm therefore faces two questions: which design might perform well, and which evaluation would teach it something useful?

A surrogate approximates the relationship between a design and its performance using the evaluations already available. The search uses that approximation to propose candidates, then checks them with the original simulator. Each checked candidate adds another observation to the model.

Watch the loop learn

Interactive illustration

Prediction → evaluation → update

Step through a synthetic one-dimensional design problem. The line is the surrogate prediction, the band shows its uncertainty, and each dot is an evaluated design.

Schematic teaching example. It illustrates the mechanism, not measured or simulated performance from the paper.

The illustration uses a small Gaussian process so the mechanism is easy to inspect. It starts with three samples and selects subsequent candidates using a lower confidence bound: predicted cost minus a multiple of predictive uncertainty. A low predicted cost makes a candidate attractive; uncertainty can also make an unexplored region worth evaluating.

Use Reveal synthetic landscape to compare the prediction with the teaching function. In a real engineering problem, that complete landscape is unavailable: each point has to be evaluated individually.

What the uncertainty band tells us

Near observations, the surrogate has more evidence. Far from them, the band is usually wider. This is a model’s uncertainty under its assumptions, not a guarantee that an engineering design is safe or meets a specification. Model mismatch, noisy evaluations, and an unfamiliar region can all affect calibration.

A simulation still has the final word

The surrogate guides the allocation of expensive evaluations. Feasibility and performance still need to be checked with the simulator and, where appropriate, measurement. In constrained problems, the search must also account for multiple performance requirements and geometry restrictions.

Where this appears in my research

Power amplifier design: Bayesian neural networks support E-GASPAD’s hybrid optimization loop.
Pixelated antenna design: ensemble learning and geometry-aware genetic operators support the search over binary layouts.
LEAM and LADS: modeling artifacts and optimization setup connect the search to an engineering workflow.