Patrick Petter
Get in touch
Surrogate Optimization · In-context learning for optimization

Fewer expensive simulations, same engineering answer.

Transformer surrogates in the prior-fitted-network style for expensive, constrained engineering optimization — packaged as a reusable tool with a license-clean benchmark suite and an interactive comparison of search strategies.

ConceptOptimization
The question

Fewer expensive simulations, same engineering answer.

Engineering design problems are expensive to evaluate and hedged in by constraints. Every simulation saved matters, but the strongest research code for surrogate-based optimization is rarely packaged in a way practitioners can use.

Approach

How it works

  1. Prior-fitted network surrogates. In-context-learning surrogates rather than classical Gaussian-process or ensemble surrogates.

  2. Engineering-first benchmarks. Constrained simulation benchmarks rather than hyperparameter optimization.

  3. Honest positioning. The contribution is a cleanly packaged tool, benchmark suite and visual comparison of gradient descent, surrogate and evolutionary search — explicitly not a new algorithm.

Stack & sources

Under the hood

PythonPyTorchprior-fitted networks

References

  • Müller et al. (2023). PFNs4BO. ICML 2023.
  • Yu et al. (2025). PFN4CBO / BOEngineeringBenchmark. arXiv:2404.04495.
Discuss this research

Working on a similar problem?

Collaboration, an application of Surrogate Optimization in your context, or a speaking slot — I reply personally.

Start a conversation
Related research